diff --git a/.github/FUNDING.yml b/.github/FUNDING.yml new file mode 100644 index 0000000..e69de29 diff --git a/CHANGELOG.md b/CHANGELOG.md new file mode 100644 index 0000000..244731f --- /dev/null +++ b/CHANGELOG.md @@ -0,0 +1,154 @@ +# Changelog +All notable changes to this project will be documented in this file. + +The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/), +and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html). + +## [1.8.0] - 2021-06-20 +### Added +- Morph factor option +- Migrated options from SAEHD to AMP: + - Loss function + - Random downsample + - Random noise + - Random blur + - Random jpeg + - Background Power + - CT mode: fs-aug + - Random color + +## [1.7.3] - 2021-06-16 +### Fixed +- AMP mask type + +## [1.7.2] - 2021-06-15 +### Added +- New sample degradation options (only affects input, similar to random warp): + - Random noise (gaussian/laplace/poisson) + - Random blur (gaussian/motion) + - Random jpeg compression + - Random downsampling +- New "warped" preview(s): Shows the input samples with any/all distortions. + +## [1.7.1] - 2021-06-15 +### Added +- New autobackup options: + - Session name + - ISO Timestamps (instead of numbered) + - Max number of backups to keep (use "0" for unlimited) + +## [1.7.0] - 2021-06-15 +### Updated +- Merged in latest changes from upstream, including new AMP model + +## [1.6.2] - 2021-05-08 +### Fixed +- Fixed bug with GAN smoothing/noisy labels with certain versions of Tensorflow + +## [1.6.1] - 2021-05-04 +### Fixed +- Fixed bug when `fs-aug` used on model with same resolution as dataset + +## [1.6.0] - 2021-05-04 +### Added +- New loss function "MS-SSIM+L1", based on ["Loss Functions for Image Restoration with Neural Networks"](https://research.nvidia.com/publication/loss-functions-image-restoration-neural-networks) + +## [1.5.1] - 2021-04-23 +### Fixed +- Fixes bug with MS-SSIM when using a version of tensorflow < 1.14 + +## [1.5.0] - 2021-03-29 +### Changed +- Web UI previews now show preview pane as PNG (loss-less), instead of JPG (lossy), so we can see the same output + as on desktop, without any changes from JPG compression. This has the side-effect of the preview images loading slower + over web, as they are now larger, a future update may be considered which would give the option to view as JPG + instead. + +## [1.4.2] - 2021-03-26 +### Fixed +- Fixes bug in background power with MS-SSIM, that misattributed loss from dst to src + +## [1.4.1] - 2021-03-25 +### Fixed +- When both Background Power and MS-SSIM were enabled, the src and dst losses were being overwritten with the + "background power" losses. Fixed so "background power" losses are properly added with the total losses. + - *Note: since all the other losses were being skipped when ms-ssim and background loss were being enabled, this had + the side-effect of lowering the memory requirements (and raising the max batch size). With this fix, you may + experience an OOM error on models ran with both these features enabled. I may revisit this in another feature, + allowing you to manually disable certain loss calculations, for similar performance benefits.* + +## [1.4.0] - 2021-03-24 +### Added +- [MS-SSIM loss training option](doc/features/ms-ssim) +- GAN version option (v2 - late 2020 or v3 - current GAN) +- [GAN label smoothing and label noise options](doc/features/gan-options) +### Fixed +- Background Power now uses the entire image, not just the area outside of the mask for comparison. +This should help with rough areas directly next to the mask + +## [1.3.0] - 2021-03-20 +### Added +- [Background Power training option](doc/features/background-power/README.md) + +## [1.2.1] - 2021-03-20 +### Fixed +- Fixes bug with `fs-aug` color mode. + +## [1.2.0] - 2021-03-17 +### Added +- [Random color training option](doc/features/random-color/README.md) + +## [1.1.5] - 2021-03-16 +### Fixed +- Fixed unclosed websocket in Web UI client when exiting + +## [1.1.4] - 2021-03-16 +### Fixed +- Fixed bug when exiting from Web UI + +## [1.1.3] - 2021-03-16 +### Changed +- Updated changelog with unreleased features, links to working branches + +## [1.1.2] - 2021-03-12 +### Fixed +- [Fixed missing predicted src mask in 'SAEHD masked' preview](doc/fixes/predicted_src_mask/README.md) + +## [1.1.1] - 2021-03-12 +### Added +- CHANGELOG file for tracking updates, new features, and bug fixes +- Documentation for Web UI +- Link to CHANGELOG at top of README + +## [1.1.0] - 2021-03-11 +### Added +- [Web UI for training preview](doc/features/webui/README.md) + +## [1.0.0] - 2021-03-09 +### Initialized +- Reset stale master branch to [seranus/DeepFaceLab](https://github.com/seranus/DeepFaceLab), + 21 commits ahead of [iperov/DeepFaceLab](https://github.com/iperov/DeepFaceLab) ([compare](https://github.com/iperov/DeepFaceLab/compare/4818183...seranus:3f5ae05)) + +[1.8.0]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.7.3...v1.8.0 +[1.7.3]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.7.2...v1.7.3 +[1.7.2]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.7.1...v1.7.2 +[1.7.1]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.7.0...v1.7.1 +[1.7.0]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.6.2...v1.7.0 +[1.6.2]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.6.1...v1.6.2 +[1.6.1]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.6.0...v1.6.1 +[1.6.0]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.5.1...v1.6.0 +[1.5.1]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.5.0...v1.5.1 +[1.5.0]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.4.2...v1.5.0 +[1.4.2]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.4.1...v1.4.2 +[1.4.1]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.4.0...v1.4.1 +[1.4.0]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.3.0...v1.4.0 +[1.3.0]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.2.1...v1.3.0 +[1.2.1]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.2.0...v1.2.1 +[1.2.0]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.1.5...v1.2.0 +[1.1.5]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.1.4...v1.1.5 +[1.1.4]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.1.3...v1.1.4 +[1.1.3]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.1.2...v1.1.3 +[1.1.2]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.1.1...v1.1.2 +[1.1.1]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.1.0...v1.1.1 +[1.1.0]: https://github.com/faceshiftlabs/DeepFaceLab/compare/v1.0.0...v1.1.0 +[1.0.0]: https://github.com/faceshiftlabs/DeepFaceLab/releases/tag/v1.0.0 diff --git a/README.md b/README.md index 39aba89..c860133 100644 --- a/README.md +++ b/README.md @@ -1,4 +1,11 @@ - +[![Patreon](https://c5.patreon.com/external/logo/become_a_patron_button@2x.png)](https://www.patreon.com/bePatron?u=22997465) + +# CHANGELOG +### [View most recent changes](CHANGELOG.md) + +![](doc/dfl_cover.png) + +
- -
@@ -19,9 +26,9 @@ https://arxiv.org/abs/2005.05535

-![](doc/logo_cuda.png) ![](doc/logo_tensorflow.png) -![](doc/logo_python.png) +![](doc/logo_cuda.png) +![](doc/logo_directx.png)

@@ -29,8 +36,8 @@ More than 95% of deepfake videos are created with DeepFaceLab. DeepFaceLab is used by such popular youtube channels as -|![](doc/tiktok_icon.png) [deeptomcruise](https://www.tiktok.com/@deeptomcruise)| -|---| +|![](doc/tiktok_icon.png) [deeptomcruise](https://www.tiktok.com/@deeptomcruise)|![](doc/tiktok_icon.png) [1facerussia](https://www.tiktok.com/@1facerussia)|![](doc/tiktok_icon.png) [arnoldschwarzneggar](https://www.tiktok.com/@arnoldschwarzneggar) +|---|---|---| |![](doc/youtube_icon.png) [Ctrl Shift Face](https://www.youtube.com/channel/UCKpH0CKltc73e4wh0_pgL3g)|![](doc/youtube_icon.png) [VFXChris Ume](https://www.youtube.com/channel/UCGf4OlX_aTt8DlrgiH3jN3g/videos)|![](doc/youtube_icon.png) [Sham00k](https://www.youtube.com/channel/UCZXbWcv7fSZFTAZV4beckyw/videos)| |---|---|---| @@ -194,7 +201,7 @@ Unfortunately, there is no "make everything ok" button in DeepFaceLab. You shoul
-Windows (magnet link) +Windows (magnet link) Last release. Use torrent client to download.
@@ -333,10 +340,6 @@ QQ 951138799 bitcoin:bc1qkhh7h0gwwhxgg6h6gpllfgstkd645fefrd5s6z
-Alipay 捐款 -
### Collect facesets diff --git a/XSegEditor/QIconDB.py b/XSegEditor/QIconDB.py index a2427c3..1fd9e3e 100644 --- a/XSegEditor/QIconDB.py +++ b/XSegEditor/QIconDB.py @@ -17,6 +17,7 @@ class QIconDB(): QIconDB.poly_type_exclude = QIcon ( str(icon_path / 'poly_type_exclude.png') ) QIconDB.left = QIcon ( str(icon_path / 'left.png') ) QIconDB.right = QIcon ( str(icon_path / 'right.png') ) + QIconDB.trashcan = QIcon ( str(icon_path / 'trashcan.png') ) QIconDB.pt_edit_mode = QIcon ( str(icon_path / 'pt_edit_mode.png') ) QIconDB.view_lock_center = QIcon ( str(icon_path / 'view_lock_center.png') ) QIconDB.view_baked = QIcon ( str(icon_path / 'view_baked.png') ) diff --git a/XSegEditor/QStringDB.py b/XSegEditor/QStringDB.py index 632419e..b9100d2 100644 --- a/XSegEditor/QStringDB.py +++ b/XSegEditor/QStringDB.py @@ -85,6 +85,11 @@ class QStringDB(): 'zh' : '保存并转到下一张图片\n按住SHIFT : 加快\n按住CTRL : 跳过未标记的\n', }[lang] + QStringDB.btn_delete_image_tip = { 'en' : 'Move to _trash and Next image\n', + 'ru' : 'Переместить в _trash и следующее изображение\n', + 'zh' : '移至_trash，转到下一张图片 ', + }[lang] + QStringDB.loading_tip = {'en' : 'Loading', 'ru' : 'Загрузка', 'zh' : '正在载入', diff --git a/XSegEditor/XSegEditor.py b/XSegEditor/XSegEditor.py index c9cb6aa..affc9f6 100644 --- a/XSegEditor/XSegEditor.py +++ b/XSegEditor/XSegEditor.py @@ -1164,6 +1164,7 @@ class MainWindow(QXMainWindow): super().__init__() self.input_dirpath = input_dirpath + self.trash_dirpath = input_dirpath.parent / (input_dirpath.name + '_trash') self.cfg_root_path = cfg_root_path self.cfg_path = cfg_root_path / 'MainWindow_cfg.dat' @@ -1341,7 +1342,18 @@ class MainWindow(QXMainWindow): self.update_cached_images() self.update_preview_bar() - + + def trash_current_image(self): + self.process_next_image() + + img_path = self.image_paths_done.pop(-1) + img_path = Path(img_path) + self.trash_dirpath.mkdir(parents=True, exist_ok=True) + img_path.rename( self.trash_dirpath / img_path.name ) + + self.update_cached_images() + self.update_preview_bar() + def initialize_ui(self): self.canvas = QCanvas() @@ -1356,20 +1368,36 @@ class MainWindow(QXMainWindow): btn_next_image = QXIconButton(QIconDB.right, QStringDB.btn_next_image_tip, shortcut='D', click_func=self.process_next_image) btn_next_image.setIconSize(QUIConfig.preview_bar_icon_q_size) - + btn_delete_image = QXIconButton(QIconDB.trashcan, QStringDB.btn_delete_image_tip, shortcut='X', click_func=self.trash_current_image) + btn_delete_image.setIconSize(QUIConfig.preview_bar_icon_q_size) + + pad_image = QWidget() + pad_image.setFixedSize(QUIConfig.preview_bar_icon_q_size) + preview_image_bar_frame_l = QHBoxLayout() preview_image_bar_frame_l.setContentsMargins(0,0,0,0) + preview_image_bar_frame_l.addWidget ( pad_image, alignment=Qt.AlignCenter) preview_image_bar_frame_l.addWidget ( btn_prev_image, alignment=Qt.AlignCenter) preview_image_bar_frame_l.addWidget ( image_bar) preview_image_bar_frame_l.addWidget ( btn_next_image, alignment=Qt.AlignCenter) + #preview_image_bar_frame_l.addWidget ( btn_delete_image, alignment=Qt.AlignCenter) preview_image_bar_frame = QFrame() preview_image_bar_frame.setSizePolicy ( QSizePolicy.Fixed, QSizePolicy.Fixed ) preview_image_bar_frame.setLayout(preview_image_bar_frame_l) - preview_image_bar_l = QHBoxLayout() - preview_image_bar_l.addWidget (preview_image_bar_frame) + preview_image_bar_frame2_l = QHBoxLayout() + preview_image_bar_frame2_l.setContentsMargins(0,0,0,0) + preview_image_bar_frame2_l.addWidget ( btn_delete_image, alignment=Qt.AlignCenter) + preview_image_bar_frame2 = QFrame() + preview_image_bar_frame2.setSizePolicy ( QSizePolicy.Fixed, QSizePolicy.Fixed ) + preview_image_bar_frame2.setLayout(preview_image_bar_frame2_l) + + preview_image_bar_l = QHBoxLayout() + preview_image_bar_l.addWidget (preview_image_bar_frame, alignment=Qt.AlignCenter) + preview_image_bar_l.addWidget (preview_image_bar_frame2) + preview_image_bar = QFrame() preview_image_bar.setFrameShape(QFrame.StyledPanel) preview_image_bar.setSizePolicy ( QSizePolicy.Expanding, QSizePolicy.Fixed ) diff --git a/XSegEditor/gfx/icons/trashcan.png b/XSegEditor/gfx/icons/trashcan.png new file mode 100644 index 0000000..a31285b Binary files /dev/null and b/XSegEditor/gfx/icons/trashcan.png differ diff --git a/core/imagelib/SegIEPolys.py b/core/imagelib/SegIEPolys.py index e658711..1a4c3d2 100644 --- a/core/imagelib/SegIEPolys.py +++ b/core/imagelib/SegIEPolys.py @@ -77,6 +77,8 @@ class SegIEPoly(): self.pts = np.array(pts) self.n_max = self.n = len(pts) + def mult_points(self, val): + self.pts *= val @@ -136,7 +138,11 @@ class SegIEPolys(): def dump(self): return {'polys' : [ poly.dump() for poly in self.polys ] } - + + def mult_points(self, val): + for poly in self.polys: + poly.mult_points(val) + @staticmethod def load(data=None): ie_polys = SegIEPolys() diff --git a/core/imagelib/__init__.py b/core/imagelib/__init__.py index e74d427..1ed95dc 100644 --- a/core/imagelib/__init__.py +++ b/core/imagelib/__init__.py @@ -14,14 +14,19 @@ from .reduce_colors import reduce_colors from .color_transfer import color_transfer, color_transfer_mix, color_transfer_sot, color_transfer_mkl, color_transfer_idt, color_hist_match, reinhard_color_transfer, linear_color_transfer, color_augmentation -from .common import normalize_channels, cut_odd_image, overlay_alpha_image +from .common import random_crop, normalize_channels, cut_odd_image, overlay_alpha_image from .SegIEPolys import * from .blursharpen import LinearMotionBlur, blursharpen from .filters import apply_random_rgb_levels, \ + apply_random_overlay_triangle, \ apply_random_hsv_shift, \ + apply_random_sharpen, \ apply_random_motion_blur, \ apply_random_gaussian_blur, \ - apply_random_bilinear_resize + apply_random_nearest_resize, \ + apply_random_bilinear_resize, \ + apply_random_jpeg_compress, \ + apply_random_relight diff --git a/core/imagelib/color_transfer.py b/core/imagelib/color_transfer.py index aaadd88..c9cafc5 100644 --- a/core/imagelib/color_transfer.py +++ b/core/imagelib/color_transfer.py @@ -373,6 +373,7 @@ def color_transfer(ct_mode, img_src, img_trg): # imported from faceswap def color_augmentation(img, seed=None): """ Color adjust RGB image """ + img = img.astype(np.float32) face = img face = np.clip(face*255.0, 0, 255).astype(np.uint8) face = random_clahe(face, seed) @@ -381,6 +382,25 @@ def color_augmentation(img, seed=None): return (face / 255.0).astype(np.float32) +def random_lab_rotation(image, seed=None): + """ + Randomly rotates image color around the L axis in LAB colorspace, + keeping perceptual lightness constant. + """ + image = cv2.cvtColor(image.astype(np.float32), cv2.COLOR_BGR2LAB) + M = np.eye(3) + M[1:, 1:] = special_ortho_group.rvs(2, 1, seed) + image = image.dot(M) + l, a, b = cv2.split(image) + l = np.clip(l, 0, 100) + a = np.clip(a, -127, 127) + b = np.clip(b, -127, 127) + image = cv2.merge([l, a, b]) + image = cv2.cvtColor(image.astype(np.float32), cv2.COLOR_LAB2BGR) + np.clip(image, 0, 1, out=image) + return image + + def random_lab(image, seed=None): """ Perform random color/lightness adjustment in L*a*b* colorspace """ random.seed(seed) diff --git a/core/imagelib/common.py b/core/imagelib/common.py index 6566819..4219d7d 100644 --- a/core/imagelib/common.py +++ b/core/imagelib/common.py @@ -1,5 +1,16 @@ import numpy as np +def random_crop(img, w, h): + height, width = img.shape[:2] + + h_rnd = height - h + w_rnd = width - w + + y = np.random.randint(0, h_rnd) if h_rnd > 0 else 0 + x = np.random.randint(0, w_rnd) if w_rnd > 0 else 0 + + return img[y:y+height, x:x+width] + def normalize_channels(img, target_channels): img_shape_len = len(img.shape) if img_shape_len == 2: diff --git a/core/imagelib/filters.py b/core/imagelib/filters.py index ba51e07..6b69576 100644 --- a/core/imagelib/filters.py +++ b/core/imagelib/filters.py @@ -1,47 +1,65 @@ import numpy as np -from .blursharpen import LinearMotionBlur +from .blursharpen import LinearMotionBlur, blursharpen import cv2 def apply_random_rgb_levels(img, mask=None, rnd_state=None): if rnd_state is None: rnd_state = np.random np_rnd = rnd_state.rand - + inBlack = np.array([np_rnd()*0.25 , np_rnd()*0.25 , np_rnd()*0.25], dtype=np.float32) inWhite = np.array([1.0-np_rnd()*0.25, 1.0-np_rnd()*0.25, 1.0-np_rnd()*0.25], dtype=np.float32) inGamma = np.array([0.5+np_rnd(), 0.5+np_rnd(), 0.5+np_rnd()], dtype=np.float32) - + outBlack = np.array([np_rnd()*0.25 , np_rnd()*0.25 , np_rnd()*0.25], dtype=np.float32) outWhite = np.array([1.0-np_rnd()*0.25, 1.0-np_rnd()*0.25, 1.0-np_rnd()*0.25], dtype=np.float32) result = np.clip( (img - inBlack) / (inWhite - inBlack), 0, 1 ) result = ( result ** (1/inGamma) ) * (outWhite - outBlack) + outBlack result = np.clip(result, 0, 1) - + if mask is not None: result = img*(1-mask) + result*mask - + return result - + def apply_random_hsv_shift(img, mask=None, rnd_state=None): if rnd_state is None: rnd_state = np.random - + h, s, v = cv2.split(cv2.cvtColor(img, cv2.COLOR_BGR2HSV)) h = ( h + rnd_state.randint(360) ) % 360 s = np.clip ( s + rnd_state.random()-0.5, 0, 1 ) - v = np.clip ( v + rnd_state.random()-0.5, 0, 1 ) - + v = np.clip ( v + rnd_state.random()-0.5, 0, 1 ) + result = np.clip( cv2.cvtColor(cv2.merge([h, s, v]), cv2.COLOR_HSV2BGR) , 0, 1 ) if mask is not None: result = img*(1-mask) + result*mask - + return result - + +def apply_random_sharpen( img, chance, kernel_max_size, mask=None, rnd_state=None ): + if rnd_state is None: + rnd_state = np.random + + sharp_rnd_kernel = rnd_state.randint(kernel_max_size)+1 + + result = img + if rnd_state.randint(100) < np.clip(chance, 0, 100): + if rnd_state.randint(2) == 0: + result = blursharpen(result, 1, sharp_rnd_kernel, rnd_state.randint(10) ) + else: + result = blursharpen(result, 2, sharp_rnd_kernel, rnd_state.randint(50) ) + + if mask is not None: + result = img*(1-mask) + result*mask + + return result + def apply_random_motion_blur( img, chance, mb_max_size, mask=None, rnd_state=None ): if rnd_state is None: rnd_state = np.random - + mblur_rnd_kernel = rnd_state.randint(mb_max_size)+1 mblur_rnd_deg = rnd_state.randint(360) @@ -50,38 +68,178 @@ def apply_random_motion_blur( img, chance, mb_max_size, mask=None, rnd_state=Non result = LinearMotionBlur (result, mblur_rnd_kernel, mblur_rnd_deg ) if mask is not None: result = img*(1-mask) + result*mask - + return result - + def apply_random_gaussian_blur( img, chance, kernel_max_size, mask=None, rnd_state=None ): if rnd_state is None: rnd_state = np.random - + result = img if rnd_state.randint(100) < np.clip(chance, 0, 100): gblur_rnd_kernel = rnd_state.randint(kernel_max_size)*2+1 result = cv2.GaussianBlur(result, (gblur_rnd_kernel,)*2 , 0) if mask is not None: result = img*(1-mask) + result*mask - + return result - - -def apply_random_bilinear_resize( img, chance, max_size_per, mask=None, rnd_state=None ): + +def apply_random_resize( img, chance, max_size_per, interpolation=cv2.INTER_LINEAR, mask=None, rnd_state=None ): if rnd_state is None: rnd_state = np.random result = img if rnd_state.randint(100) < np.clip(chance, 0, 100): h,w,c = result.shape - + trg = rnd_state.rand() - rw = w - int( trg * int(w*(max_size_per/100.0)) ) - rh = h - int( trg * int(h*(max_size_per/100.0)) ) - - result = cv2.resize (result, (rw,rh), interpolation=cv2.INTER_LINEAR ) - result = cv2.resize (result, (w,h), interpolation=cv2.INTER_LINEAR ) + rw = w - int( trg * int(w*(max_size_per/100.0)) ) + rh = h - int( trg * int(h*(max_size_per/100.0)) ) + + result = cv2.resize (result, (rw,rh), interpolation=interpolation ) + result = cv2.resize (result, (w,h), interpolation=interpolation ) if mask is not None: result = img*(1-mask) + result*mask - + + return result + +def apply_random_nearest_resize( img, chance, max_size_per, mask=None, rnd_state=None ): + return apply_random_resize( img, chance, max_size_per, interpolation=cv2.INTER_NEAREST, mask=mask, rnd_state=rnd_state ) + +def apply_random_bilinear_resize( img, chance, max_size_per, mask=None, rnd_state=None ): + return apply_random_resize( img, chance, max_size_per, interpolation=cv2.INTER_LINEAR, mask=mask, rnd_state=rnd_state ) + +def apply_random_jpeg_compress( img, chance, mask=None, rnd_state=None ): + if rnd_state is None: + rnd_state = np.random + + result = img + if rnd_state.randint(100) < np.clip(chance, 0, 100): + h,w,c = result.shape + + quality = rnd_state.randint(10,101) + + ret, result = cv2.imencode('.jpg', np.clip(img*255, 0,255).astype(np.uint8), [int(cv2.IMWRITE_JPEG_QUALITY), quality] ) + if ret == True: + result = cv2.imdecode(result, flags=cv2.IMREAD_UNCHANGED) + result = result.astype(np.float32) / 255.0 + if mask is not None: + result = img*(1-mask) + result*mask + + return result + +def apply_random_overlay_triangle( img, max_alpha, mask=None, rnd_state=None ): + if rnd_state is None: + rnd_state = np.random + + h,w,c = img.shape + pt1 = [rnd_state.randint(w), rnd_state.randint(h) ] + pt2 = [rnd_state.randint(w), rnd_state.randint(h) ] + pt3 = [rnd_state.randint(w), rnd_state.randint(h) ] + + alpha = rnd_state.uniform()*max_alpha + + tri_mask = cv2.fillPoly( np.zeros_like(img), [ np.array([pt1,pt2,pt3], np.int32) ], (alpha,)*c ) + + if rnd_state.randint(2) == 0: + result = np.clip(img+tri_mask, 0, 1) + else: + result = np.clip(img-tri_mask, 0, 1) + + if mask is not None: + result = img*(1-mask) + result*mask + + return result + +def _min_resize(x, m): + if x.shape[0] < x.shape[1]: + s0 = m + s1 = int(float(m) / float(x.shape[0]) * float(x.shape[1])) + else: + s0 = int(float(m) / float(x.shape[1]) * float(x.shape[0])) + s1 = m + new_max = min(s1, s0) + raw_max = min(x.shape[0], x.shape[1]) + return cv2.resize(x, (s1, s0), interpolation=cv2.INTER_LANCZOS4) + +def _d_resize(x, d, fac=1.0): + new_min = min(int(d[1] * fac), int(d[0] * fac)) + raw_min = min(x.shape[0], x.shape[1]) + if new_min < raw_min: + interpolation = cv2.INTER_AREA + else: + interpolation = cv2.INTER_LANCZOS4 + y = cv2.resize(x, (int(d[1] * fac), int(d[0] * fac)), interpolation=interpolation) + return y + +def _get_image_gradient(dist): + cols = cv2.filter2D(dist, cv2.CV_32F, np.array([[-1, 0, +1], [-2, 0, +2], [-1, 0, +1]])) + rows = cv2.filter2D(dist, cv2.CV_32F, np.array([[-1, -2, -1], [0, 0, 0], [+1, +2, +1]])) + return cols, rows + +def _generate_lighting_effects(content): + h512 = content + h256 = cv2.pyrDown(h512) + h128 = cv2.pyrDown(h256) + h64 = cv2.pyrDown(h128) + h32 = cv2.pyrDown(h64) + h16 = cv2.pyrDown(h32) + c512, r512 = _get_image_gradient(h512) + c256, r256 = _get_image_gradient(h256) + c128, r128 = _get_image_gradient(h128) + c64, r64 = _get_image_gradient(h64) + c32, r32 = _get_image_gradient(h32) + c16, r16 = _get_image_gradient(h16) + c = c16 + c = _d_resize(cv2.pyrUp(c), c32.shape) * 4.0 + c32 + c = _d_resize(cv2.pyrUp(c), c64.shape) * 4.0 + c64 + c = _d_resize(cv2.pyrUp(c), c128.shape) * 4.0 + c128 + c = _d_resize(cv2.pyrUp(c), c256.shape) * 4.0 + c256 + c = _d_resize(cv2.pyrUp(c), c512.shape) * 4.0 + c512 + r = r16 + r = _d_resize(cv2.pyrUp(r), r32.shape) * 4.0 + r32 + r = _d_resize(cv2.pyrUp(r), r64.shape) * 4.0 + r64 + r = _d_resize(cv2.pyrUp(r), r128.shape) * 4.0 + r128 + r = _d_resize(cv2.pyrUp(r), r256.shape) * 4.0 + r256 + r = _d_resize(cv2.pyrUp(r), r512.shape) * 4.0 + r512 + coarse_effect_cols = c + coarse_effect_rows = r + EPS = 1e-10 + + max_effect = np.max((coarse_effect_cols**2 + coarse_effect_rows**2)**0.5, axis=0, keepdims=True, ).max(1, keepdims=True) + coarse_effect_cols = (coarse_effect_cols + EPS) / (max_effect + EPS) + coarse_effect_rows = (coarse_effect_rows + EPS) / (max_effect + EPS) + + return np.stack([ np.zeros_like(coarse_effect_rows), coarse_effect_rows, coarse_effect_cols], axis=-1) + +def apply_random_relight(img, mask=None, rnd_state=None): + if rnd_state is None: + rnd_state = np.random + + def_img = img + + if rnd_state.randint(2) == 0: + light_pos_y = 1.0 if rnd_state.randint(2) == 0 else -1.0 + light_pos_x = rnd_state.uniform()*2-1.0 + else: + light_pos_y = rnd_state.uniform()*2-1.0 + light_pos_x = 1.0 if rnd_state.randint(2) == 0 else -1.0 + + light_source_height = 0.3*rnd_state.uniform()*0.7 + light_intensity = 1.0+rnd_state.uniform() + ambient_intensity = 0.5 + + light_source_location = np.array([[[light_source_height, light_pos_y, light_pos_x ]]], dtype=np.float32) + light_source_direction = light_source_location / np.sqrt(np.sum(np.square(light_source_location))) + + lighting_effect = _generate_lighting_effects(img) + lighting_effect = np.sum(lighting_effect * light_source_direction, axis=-1).clip(0, 1) + lighting_effect = np.mean(lighting_effect, axis=-1, keepdims=True) + + result = def_img * (ambient_intensity + lighting_effect * light_intensity) #light_source_color + result = np.clip(result, 0, 1) + + if mask is not None: + result = def_img*(1-mask) + result*mask + return result \ No newline at end of file diff --git a/core/imagelib/sd/__init__.py b/core/imagelib/sd/__init__.py index 2eafd4c..1cddc19 100644 --- a/core/imagelib/sd/__init__.py +++ b/core/imagelib/sd/__init__.py @@ -1,2 +1,2 @@ -from .draw import * +from .draw import circle_faded, random_circle_faded, bezier, random_bezier_split_faded, random_faded from .calc import * \ No newline at end of file diff --git a/core/imagelib/sd/draw.py b/core/imagelib/sd/draw.py index 77e9a46..711ad33 100644 --- a/core/imagelib/sd/draw.py +++ b/core/imagelib/sd/draw.py @@ -1,23 +1,36 @@ """ Signed distance drawing functions using numpy. """ +import math import numpy as np from numpy import linalg as npla -def circle_faded( hw, center, fade_dists ): + +def vector2_dot(a,b): + return a[...,0]*b[...,0]+a[...,1]*b[...,1] + +def vector2_dot2(a): + return a[...,0]*a[...,0]+a[...,1]*a[...,1] + +def vector2_cross(a,b): + return a[...,0]*b[...,1]-a[...,1]*b[...,0] + + +def circle_faded( wh, center, fade_dists ): """ returns drawn circle in [h,w,1] output range [0..1.0] float32 - hw = [h,w] resolution - center = [y,x] center of circle + wh = [w,h] resolution + center = [x,y] center of circle fade_dists = [fade_start, fade_end] fade values """ - h,w = hw + w,h = wh pts = np.empty( (h,w,2), dtype=np.float32 ) - pts[...,1] = np.arange(h)[None,:] pts[...,0] = np.arange(w)[:,None] + pts[...,1] = np.arange(h)[None,:] + pts = pts.reshape ( (h*w, -1) ) pts_dists = np.abs ( npla.norm(pts-center, axis=-1) ) @@ -30,15 +43,158 @@ def circle_faded( hw, center, fade_dists ): pts_dists = np.clip( 1-pts_dists, 0, 1) return pts_dists.reshape ( (h,w,1) ).astype(np.float32) + + +def bezier( wh, A, B, C ): + """ + returns drawn bezier in [h,w,1] output range float32, + every pixel contains signed distance to bezier line + + wh [w,h] resolution + A,B,C points [x,y] + """ -def random_circle_faded ( hw, rnd_state=None ): + width,height = wh + + A = np.float32(A) + B = np.float32(B) + C = np.float32(C) + + + pos = np.empty( (height,width,2), dtype=np.float32 ) + pos[...,0] = np.arange(width)[:,None] + pos[...,1] = np.arange(height)[None,:] + + + a = B-A + b = A - 2.0*B + C + c = a * 2.0 + d = A - pos + + b_dot = vector2_dot(b,b) + if b_dot == 0.0: + return np.zeros( (height,width), dtype=np.float32 ) + + kk = 1.0 / b_dot + + kx = kk * vector2_dot(a,b) + ky = kk * (2.0*vector2_dot(a,a)+vector2_dot(d,b))/3.0; + kz = kk * vector2_dot(d,a); + + res = 0.0; + sgn = 0.0; + + p = ky - kx*kx; + + p3 = p*p*p; + q = kx*(2.0*kx*kx - 3.0*ky) + kz; + h = q*q + 4.0*p3; + + hp_sel = h >= 0.0 + + hp_p = h[hp_sel] + hp_p = np.sqrt(hp_p) + + hp_x = ( np.stack( (hp_p,-hp_p), -1) -q[hp_sel,None] ) / 2.0 + hp_uv = np.sign(hp_x) * np.power( np.abs(hp_x), [1.0/3.0, 1.0/3.0] ) + hp_t = np.clip( hp_uv[...,0] + hp_uv[...,1] - kx, 0.0, 1.0 ) + + hp_t = hp_t[...,None] + hp_q = d[hp_sel]+(c+b*hp_t)*hp_t + hp_res = vector2_dot2(hp_q) + hp_sgn = vector2_cross(c+2.0*b*hp_t,hp_q) + + hl_sel = h < 0.0 + + hl_q = q[hl_sel] + hl_p = p[hl_sel] + hl_z = np.sqrt(-hl_p) + hl_v = np.arccos( hl_q / (hl_p*hl_z*2.0)) / 3.0 + + hl_m = np.cos(hl_v) + hl_n = np.sin(hl_v)*1.732050808; + + hl_t = np.clip( np.stack( (hl_m+hl_m,-hl_n-hl_m,hl_n-hl_m), -1)*hl_z[...,None]-kx, 0.0, 1.0 ); + + hl_d = d[hl_sel] + + hl_qx = hl_d+(c+b*hl_t[...,0:1])*hl_t[...,0:1] + + hl_dx = vector2_dot2(hl_qx) + hl_sx = vector2_cross(c+2.0*b*hl_t[...,0:1], hl_qx) + + hl_qy = hl_d+(c+b*hl_t[...,1:2])*hl_t[...,1:2] + hl_dy = vector2_dot2(hl_qy) + hl_sy = vector2_cross(c+2.0*b*hl_t[...,1:2],hl_qy); + + hl_dx_l_dy = hl_dx=hl_dy + + hl_res = np.empty_like(hl_dx) + hl_res[hl_dx_l_dy] = hl_dx[hl_dx_l_dy] + hl_res[hl_dx_ge_dy] = hl_dy[hl_dx_ge_dy] + + hl_sgn = np.empty_like(hl_sx) + hl_sgn[hl_dx_l_dy] = hl_sx[hl_dx_l_dy] + hl_sgn[hl_dx_ge_dy] = hl_sy[hl_dx_ge_dy] + + res = np.empty( (height, width), np.float32 ) + res[hp_sel] = hp_res + res[hl_sel] = hl_res + + sgn = np.empty( (height, width), np.float32 ) + sgn[hp_sel] = hp_sgn + sgn[hl_sel] = hl_sgn + + sgn = np.sign(sgn) + res = np.sqrt(res)*sgn + + return res[...,None] + +def random_faded(wh): + """ + apply one of them: + random_circle_faded + random_bezier_split_faded + """ + rnd = np.random.randint(2) + if rnd == 0: + return random_circle_faded(wh) + elif rnd == 1: + return random_bezier_split_faded(wh) + +def random_circle_faded ( wh, rnd_state=None ): if rnd_state is None: rnd_state = np.random - h,w = hw - hw_max = max(h,w) - fade_start = rnd_state.randint(hw_max) - fade_end = fade_start + rnd_state.randint(hw_max- fade_start) + w,h = wh + wh_max = max(w,h) + fade_start = rnd_state.randint(wh_max) + fade_end = fade_start + rnd_state.randint(wh_max- fade_start) - return circle_faded (hw, [ rnd_state.randint(h), rnd_state.randint(w) ], - [fade_start, fade_end] ) \ No newline at end of file + return circle_faded (wh, [ rnd_state.randint(h), rnd_state.randint(w) ], + [fade_start, fade_end] ) + +def random_bezier_split_faded( wh ): + width, height = wh + + degA = np.random.randint(360) + degB = np.random.randint(360) + degC = np.random.randint(360) + + deg_2_rad = math.pi / 180.0 + + center = np.float32([width / 2.0, height / 2.0]) + + radius = max(width, height) + + A = center + radius*np.float32([ math.sin( degA * deg_2_rad), math.cos( degA * deg_2_rad) ] ) + B = center + np.random.randint(radius)*np.float32([ math.sin( degB * deg_2_rad), math.cos( degB * deg_2_rad) ] ) + C = center + radius*np.float32([ math.sin( degC * deg_2_rad), math.cos( degC * deg_2_rad) ] ) + + x = bezier( (width,height), A, B, C ) + + x = x / (1+np.random.randint(radius)) + 0.5 + + x = np.clip(x, 0, 1) + return x diff --git a/core/imagelib/warp.py b/core/imagelib/warp.py index 559c2e8..fd4b3a9 100644 --- a/core/imagelib/warp.py +++ b/core/imagelib/warp.py @@ -2,7 +2,7 @@ import numpy as np import cv2 from core import randomex -def gen_warp_params (w, flip, rotation_range=[-2,2], scale_range=[-0.5, 0.5], tx_range=[-0.05, 0.05], ty_range=[-0.05, 0.05], rnd_state=None ): +def gen_warp_params (w, flip=False, rotation_range=[-2,2], scale_range=[-0.5, 0.5], tx_range=[-0.05, 0.05], ty_range=[-0.05, 0.05], rnd_state=None ): if rnd_state is None: rnd_state = np.random diff --git a/core/leras/device.py b/core/leras/device.py index 4d157f0..31d2f88 100644 --- a/core/leras/device.py +++ b/core/leras/device.py @@ -1,12 +1,19 @@ import sys import ctypes import os +import multiprocessing +import json +import time +from pathlib import Path +from core.interact import interact as io + class Device(object): - def __init__(self, index, name, total_mem, free_mem, cc=0): + def __init__(self, index, tf_dev_type, name, total_mem, free_mem): self.index = index + self.tf_dev_type = tf_dev_type self.name = name - self.cc = cc + self.total_mem = total_mem self.total_mem_gb = total_mem / 1024**3 self.free_mem = free_mem @@ -82,12 +89,135 @@ class Devices(object): result.append (device) return Devices(result) + @staticmethod + def _get_tf_devices_proc(q : multiprocessing.Queue): + + if sys.platform[0:3] == 'win': + compute_cache_path = Path(os.environ['APPDATA']) / 'NVIDIA' / ('ComputeCache_ALL') + os.environ['CUDA_CACHE_PATH'] = str(compute_cache_path) + if not compute_cache_path.exists(): + io.log_info("Caching GPU kernels...") + compute_cache_path.mkdir(parents=True, exist_ok=True) + + import tensorflow + + tf_version = tensorflow.version.VERSION + #if tf_version is None: + # tf_version = tensorflow.version.GIT_VERSION + if tf_version[0] == 'v': + tf_version = tf_version[1:] + if tf_version[0] == '2': + tf = tensorflow.compat.v1 + else: + tf = tensorflow + + import logging + # Disable tensorflow warnings + tf_logger = logging.getLogger('tensorflow') + tf_logger.setLevel(logging.ERROR) + + from tensorflow.python.client import device_lib + + devices = [] + + physical_devices = device_lib.list_local_devices() + physical_devices_f = {} + for dev in physical_devices: + dev_type = dev.device_type + dev_tf_name = dev.name + dev_tf_name = dev_tf_name[ dev_tf_name.index(dev_type) : ] + + dev_idx = int(dev_tf_name.split(':')[-1]) + + if dev_type in ['GPU','DML']: + dev_name = dev_tf_name + + dev_desc = dev.physical_device_desc + if len(dev_desc) != 0: + if dev_desc[0] == '{': + dev_desc_json = json.loads(dev_desc) + dev_desc_json_name = dev_desc_json.get('name',None) + if dev_desc_json_name is not None: + dev_name = dev_desc_json_name + else: + for param, value in ( v.split(':') for v in dev_desc.split(',') ): + param = param.strip() + value = value.strip() + if param == 'name': + dev_name = value + break + + physical_devices_f[dev_idx] = (dev_type, dev_name, dev.memory_limit) + + q.put(physical_devices_f) + time.sleep(0.1) + + @staticmethod def initialize_main_env(): - os.environ['NN_DEVICES_INITIALIZED'] = '1' - os.environ['NN_DEVICES_COUNT'] = '0' + if int(os.environ.get("NN_DEVICES_INITIALIZED", 0)) != 0: + return + + if 'CUDA_VISIBLE_DEVICES' in os.environ.keys(): + os.environ.pop('CUDA_VISIBLE_DEVICES') os.environ['CUDA_​CACHE_​MAXSIZE'] = '2147483647' + os.environ['TF_MIN_GPU_MULTIPROCESSOR_COUNT'] = '2' + os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' # tf log errors only + + q = multiprocessing.Queue() + p = multiprocessing.Process(target=Devices._get_tf_devices_proc, args=(q,), daemon=True) + p.start() + p.join() + + visible_devices = q.get() + + os.environ['NN_DEVICES_INITIALIZED'] = '1' + os.environ['NN_DEVICES_COUNT'] = str(len(visible_devices)) + + for i in visible_devices: + dev_type, name, total_mem = visible_devices[i] + + os.environ[f'NN_DEVICE_{i}_TF_DEV_TYPE'] = dev_type + os.environ[f'NN_DEVICE_{i}_NAME'] = name + os.environ[f'NN_DEVICE_{i}_TOTAL_MEM'] = str(total_mem) + os.environ[f'NN_DEVICE_{i}_FREE_MEM'] = str(total_mem) + + + + @staticmethod + def getDevices(): + if Devices.all_devices is None: + if int(os.environ.get("NN_DEVICES_INITIALIZED", 0)) != 1: + raise Exception("nn devices are not initialized. Run initialize_main_env() in main process.") + devices = [] + for i in range ( int(os.environ['NN_DEVICES_COUNT']) ): + devices.append ( Device(index=i, + tf_dev_type=os.environ[f'NN_DEVICE_{i}_TF_DEV_TYPE'], + name=os.environ[f'NN_DEVICE_{i}_NAME'], + total_mem=int(os.environ[f'NN_DEVICE_{i}_TOTAL_MEM']), + free_mem=int(os.environ[f'NN_DEVICE_{i}_FREE_MEM']), ) + ) + Devices.all_devices = Devices(devices) + + return Devices.all_devices + +""" + + + # {'name' : name.split(b'\0', 1)[0].decode(), + # 'total_mem' : totalMem.value + # } + + + + + + return + + + + min_cc = int(os.environ.get("TF_MIN_REQ_CAP", 35)) libnames = ('libcuda.so', 'libcuda.dylib', 'nvcuda.dll') for libname in libnames: @@ -139,70 +269,4 @@ class Devices(object): os.environ[f'NN_DEVICE_{i}_TOTAL_MEM'] = str(device['total_mem']) os.environ[f'NN_DEVICE_{i}_FREE_MEM'] = str(device['free_mem']) os.environ[f'NN_DEVICE_{i}_CC'] = str(device['cc']) - - @staticmethod - def getDevices(): - if Devices.all_devices is None: - if int(os.environ.get("NN_DEVICES_INITIALIZED", 0)) != 1: - raise Exception("nn devices are not initialized. Run initialize_main_env() in main process.") - devices = [] - for i in range ( int(os.environ['NN_DEVICES_COUNT']) ): - devices.append ( Device(index=i, - name=os.environ[f'NN_DEVICE_{i}_NAME'], - total_mem=int(os.environ[f'NN_DEVICE_{i}_TOTAL_MEM']), - free_mem=int(os.environ[f'NN_DEVICE_{i}_FREE_MEM']), - cc=int(os.environ[f'NN_DEVICE_{i}_CC']) )) - Devices.all_devices = Devices(devices) - - return Devices.all_devices - -""" -if Devices.all_devices is None: - min_cc = int(os.environ.get("TF_MIN_REQ_CAP", 35)) - - libnames = ('libcuda.so', 'libcuda.dylib', 'nvcuda.dll') - for libname in libnames: - try: - cuda = ctypes.CDLL(libname) - except: - continue - else: - break - else: - return Devices([]) - - nGpus = ctypes.c_int() - name = b' ' * 200 - cc_major = ctypes.c_int() - cc_minor = ctypes.c_int() - freeMem = ctypes.c_size_t() - totalMem = ctypes.c_size_t() - - result = ctypes.c_int() - device = ctypes.c_int() - context = ctypes.c_void_p() - error_str = ctypes.c_char_p() - - devices = [] - - if cuda.cuInit(0) == 0 and \ - cuda.cuDeviceGetCount(ctypes.byref(nGpus)) == 0: - for i in range(nGpus.value): - if cuda.cuDeviceGet(ctypes.byref(device), i) != 0 or \ - cuda.cuDeviceGetName(ctypes.c_char_p(name), len(name), device) != 0 or \ - cuda.cuDeviceComputeCapability(ctypes.byref(cc_major), ctypes.byref(cc_minor), device) != 0: - continue - - if cuda.cuCtxCreate_v2(ctypes.byref(context), 0, device) == 0: - if cuda.cuMemGetInfo_v2(ctypes.byref(freeMem), ctypes.byref(totalMem)) == 0: - cc = cc_major.value * 10 + cc_minor.value - if cc >= min_cc: - devices.append ( Device(index=i, - name=name.split(b'\0', 1)[0].decode(), - total_mem=totalMem.value, - free_mem=freeMem.value, - cc=cc) ) - cuda.cuCtxDetach(context) - Devices.all_devices = Devices(devices) - return Devices.all_devices """ \ No newline at end of file diff --git a/core/leras/layers/Conv2D.py b/core/leras/layers/Conv2D.py index ae37c50..7d4d444 100644 --- a/core/leras/layers/Conv2D.py +++ b/core/leras/layers/Conv2D.py @@ -23,28 +23,13 @@ class Conv2D(nn.LayerBase): if padding == "SAME": padding = ( (kernel_size - 1) * dilations + 1 ) // 2 elif padding == "VALID": - padding = 0 + padding = None else: raise ValueError ("Wrong padding type. Should be VALID SAME or INT or 4x INTs") - - if isinstance(padding, int): - if padding != 0: - if nn.data_format == "NHWC": - padding = [ [0,0], [padding,padding], [padding,padding], [0,0] ] - else: - padding = [ [0,0], [0,0], [padding,padding], [padding,padding] ] - else: - padding = None - - if nn.data_format == "NHWC": - strides = [1,strides,strides,1] else: - strides = [1,1,strides,strides] - - if nn.data_format == "NHWC": - dilations = [1,dilations,dilations,1] - else: - dilations = [1,1,dilations,dilations] + padding = int(padding) + + self.in_ch = in_ch self.out_ch = out_ch @@ -93,10 +78,27 @@ class Conv2D(nn.LayerBase): if self.use_wscale: weight = weight * self.wscale - if self.padding is not None: - x = tf.pad (x, self.padding, mode='CONSTANT') + padding = self.padding + if padding is not None: + if nn.data_format == "NHWC": + padding = [ [0,0], [padding,padding], [padding,padding], [0,0] ] + else: + padding = [ [0,0], [0,0], [padding,padding], [padding,padding] ] + x = tf.pad (x, padding, mode='CONSTANT') + + strides = self.strides + if nn.data_format == "NHWC": + strides = [1,strides,strides,1] + else: + strides = [1,1,strides,strides] - x = tf.nn.conv2d(x, weight, self.strides, 'VALID', dilations=self.dilations, data_format=nn.data_format) + dilations = self.dilations + if nn.data_format == "NHWC": + dilations = [1,dilations,dilations,1] + else: + dilations = [1,1,dilations,dilations] + + x = tf.nn.conv2d(x, weight, strides, 'VALID', dilations=dilations, data_format=nn.data_format) if self.use_bias: if nn.data_format == "NHWC": bias = tf.reshape (self.bias, (1,1,1,self.out_ch) ) diff --git a/core/leras/layers/MsSsim.py b/core/leras/layers/MsSsim.py new file mode 100644 index 0000000..d4987ed --- /dev/null +++ b/core/leras/layers/MsSsim.py @@ -0,0 +1,50 @@ +from core.leras import nn +tf = nn.tf + + +class MsSsim(nn.LayerBase): + default_power_factors = (0.0448, 0.2856, 0.3001, 0.2363, 0.1333) + default_l1_alpha = 0.84 + + def __init__(self, batch_size, in_ch, resolution, kernel_size=11, use_l1=False, **kwargs): + # restrict mssim factors to those greater/equal to kernel size + power_factors = [p for i, p in enumerate(self.default_power_factors) if resolution//(2**i) >= kernel_size] + # normalize power factors if reduced because of size + if sum(power_factors) < 1.0: + power_factors = [x/sum(power_factors) for x in power_factors] + self.power_factors = power_factors + self.num_scale = len(power_factors) + self.kernel_size = kernel_size + self.use_l1 = use_l1 + if use_l1: + self.gaussian_weights = nn.get_gaussian_weights(batch_size, in_ch, resolution, num_scale=self.num_scale) + + super().__init__(**kwargs) + + def __call__(self, y_true, y_pred, max_val): + # Transpose images from NCHW to NHWC + y_true_t = tf.transpose(tf.cast(y_true, tf.float32), [0, 2, 3, 1]) + y_pred_t = tf.transpose(tf.cast(y_pred, tf.float32), [0, 2, 3, 1]) + + # ssim_multiscale returns values in range [0, 1] (where 1 is completely identical) + # subtract from 1 to get loss + if tf.__version__ >= "1.14": + ms_ssim_loss = 1.0 - tf.image.ssim_multiscale(y_true_t, y_pred_t, max_val, power_factors=self.power_factors, filter_size=self.kernel_size) + else: + ms_ssim_loss = 1.0 - tf.image.ssim_multiscale(y_true_t, y_pred_t, max_val, power_factors=self.power_factors) + + # If use L1 is enabled, use mix of ms-ssim and L1 (weighted by gaussian filters) + # H. Zhao, O. Gallo, I. Frosio and J. Kautz, "Loss Functions for Image Restoration With Neural Networks," + # in IEEE Transactions on Computational Imaging, vol. 3, no. 1, pp. 47-57, March 2017, + # doi: 10.1109/TCI.2016.2644865. + # https://research.nvidia.com/publication/loss-functions-image-restoration-neural-networks + + if self.use_l1: + diff = tf.tile(tf.expand_dims(tf.abs(y_true - y_pred), axis=0), multiples=[self.num_scale, 1, 1, 1, 1]) + l1_loss = tf.reduce_mean(tf.reduce_sum(self.gaussian_weights[-1, :, :, :, :] * diff, axis=[0, 3, 4]), axis=[1]) + return self.default_l1_alpha * ms_ssim_loss + (1 - self.default_l1_alpha) * l1_loss + + return ms_ssim_loss + + +nn.MsSsim = MsSsim diff --git a/core/leras/layers/__init__.py b/core/leras/layers/__init__.py index 1c81963..d8f1c9d 100644 --- a/core/leras/layers/__init__.py +++ b/core/leras/layers/__init__.py @@ -13,4 +13,5 @@ from .FRNorm2D import * from .TLU import * from .ScaleAdd import * from .DenseNorm import * -from .AdaIN import * \ No newline at end of file +from .AdaIN import * +from .MsSsim import * diff --git a/core/leras/models/PatchDiscriminator.py b/core/leras/models/PatchDiscriminator.py index 343e000..dcaf941 100644 --- a/core/leras/models/PatchDiscriminator.py +++ b/core/leras/models/PatchDiscriminator.py @@ -111,7 +111,7 @@ class UNetPatchDiscriminator(nn.ModelBase): for i in range(layers_count-1): st = 1 + (1 if val & (1 << i) !=0 else 0 ) layers.append ( [3, st ]) - sum_st += st + sum_st += st rf = self.calc_receptive_field_size(layers) @@ -131,7 +131,7 @@ class UNetPatchDiscriminator(nn.ModelBase): return s[q][2] def on_build(self, patch_size, in_ch, base_ch = 16): - + class ResidualBlock(nn.ModelBase): def on_build(self, ch, kernel_size=3 ): self.conv1 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME') @@ -152,7 +152,7 @@ class UNetPatchDiscriminator(nn.ModelBase): self.upres1 = [] self.upres2 = [] layers = self.find_archi(patch_size) - + level_chs = { i-1:v for i,v in enumerate([ min( base_ch * (2**i), 512 ) for i in range(len(layers)+1)]) } self.in_conv = nn.Conv2D( in_ch, level_chs[-1], kernel_size=1, padding='VALID') @@ -162,12 +162,12 @@ class UNetPatchDiscriminator(nn.ModelBase): self.res1.append ( ResidualBlock(level_chs[i]) ) self.res2.append ( ResidualBlock(level_chs[i]) ) - + self.upconvs.insert (0, nn.Conv2DTranspose( level_chs[i]*(2 if i != len(layers)-1 else 1), level_chs[i-1], kernel_size=kernel_size, strides=strides, padding='SAME') ) self.upres1.insert (0, ResidualBlock(level_chs[i-1]*2) ) self.upres2.insert (0, ResidualBlock(level_chs[i-1]*2) ) - + self.out_conv = nn.Conv2D( level_chs[-1]*2, 1, kernel_size=1, padding='VALID') self.center_out = nn.Conv2D( level_chs[len(layers)-1], 1, kernel_size=1, padding='VALID') @@ -183,7 +183,7 @@ class UNetPatchDiscriminator(nn.ModelBase): x = tf.nn.leaky_relu( conv(x), 0.2 ) x = res1(x) x = res2(x) - + center_out, x = self.center_out(x), tf.nn.leaky_relu( self.center_conv(x), 0.2 ) for i, (upconv, enc, upres1, upres2 ) in enumerate(zip(self.upconvs, encs, self.upres1, self.upres2)): @@ -195,3 +195,117 @@ class UNetPatchDiscriminator(nn.ModelBase): return center_out, self.out_conv(x) nn.UNetPatchDiscriminator = UNetPatchDiscriminator + +class UNetPatchDiscriminatorV2(nn.ModelBase): + """ + Inspired by https://arxiv.org/abs/2002.12655 "A U-Net Based Discriminator for Generative Adversarial Networks" + """ + def calc_receptive_field_size(self, layers): + """ + result the same as https://fomoro.com/research/article/receptive-field-calculatorindex.html + """ + rf = 0 + ts = 1 + for i, (k, s) in enumerate(layers): + if i == 0: + rf = k + else: + rf += (k-1)*ts + ts *= s + return rf + + def find_archi(self, target_patch_size, max_layers=6): + """ + Find the best configuration of layers using only 3x3 convs for target patch size + """ + s = {} + for layers_count in range(1,max_layers+1): + val = 1 << (layers_count-1) + while True: + val -= 1 + + layers = [] + sum_st = 0 + for i in range(layers_count-1): + st = 1 + (1 if val & (1 << i) !=0 else 0 ) + layers.append ( [3, st ]) + sum_st += st + layers.append ( [3, 2]) + sum_st += 2 + + rf = self.calc_receptive_field_size(layers) + + s_rf = s.get(rf, None) + if s_rf is None: + s[rf] = (layers_count, sum_st, layers) + else: + if layers_count < s_rf[0] or \ + ( layers_count == s_rf[0] and sum_st > s_rf[1] ): + s[rf] = (layers_count, sum_st, layers) + + if val == 0: + break + + x = sorted(list(s.keys())) + q=x[np.abs(np.array(x)-target_patch_size).argmin()] + return s[q][2] + + def on_build(self, patch_size, in_ch): + class ResidualBlock(nn.ModelBase): + def on_build(self, ch, kernel_size=3 ): + self.conv1 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME') + self.conv2 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME') + + def forward(self, inp): + x = self.conv1(inp) + x = tf.nn.leaky_relu(x, 0.2) + x = self.conv2(x) + x = tf.nn.leaky_relu(inp + x, 0.2) + return x + + prev_ch = in_ch + self.convs = [] + self.res = [] + self.upconvs = [] + self.upres = [] + layers = self.find_archi(patch_size) + base_ch = 16 + + level_chs = { i-1:v for i,v in enumerate([ min( base_ch * (2**i), 512 ) for i in range(len(layers)+1)]) } + + self.in_conv = nn.Conv2D( in_ch, level_chs[-1], kernel_size=1, padding='VALID') + + for i, (kernel_size, strides) in enumerate(layers): + self.convs.append ( nn.Conv2D( level_chs[i-1], level_chs[i], kernel_size=kernel_size, strides=strides, padding='SAME') ) + + self.res.append ( ResidualBlock(level_chs[i]) ) + + self.upconvs.insert (0, nn.Conv2DTranspose( level_chs[i]*(2 if i != len(layers)-1 else 1), level_chs[i-1], kernel_size=kernel_size, strides=strides, padding='SAME') ) + + self.upres.insert (0, ResidualBlock(level_chs[i-1]*2) ) + + self.out_conv = nn.Conv2D( level_chs[-1]*2, 1, kernel_size=1, padding='VALID') + + self.center_out = nn.Conv2D( level_chs[len(layers)-1], 1, kernel_size=1, padding='VALID') + self.center_conv = nn.Conv2D( level_chs[len(layers)-1], level_chs[len(layers)-1], kernel_size=1, padding='VALID') + + + def forward(self, x): + x = tf.nn.leaky_relu( self.in_conv(x), 0.1 ) + + encs = [] + for conv, res in zip(self.convs, self.res): + encs.insert(0, x) + x = tf.nn.leaky_relu( conv(x), 0.1 ) + x = res(x) + + center_out, x = self.center_out(x), self.center_conv(x) + + for i, (upconv, enc, upres) in enumerate(zip(self.upconvs, encs, self.upres)): + x = tf.nn.leaky_relu( upconv(x), 0.1 ) + x = tf.concat( [enc, x], axis=nn.conv2d_ch_axis) + x = upres(x) + + return center_out, self.out_conv(x) + +nn.UNetPatchDiscriminatorV2 = UNetPatchDiscriminatorV2 diff --git a/core/leras/models/XSeg.py b/core/leras/models/XSeg.py index 0ba19a6..e6bde65 100644 --- a/core/leras/models/XSeg.py +++ b/core/leras/models/XSeg.py @@ -28,11 +28,12 @@ class XSeg(nn.ModelBase): x = self.frn(x) x = self.tlu(x) return x + + self.base_ch = base_ch self.conv01 = ConvBlock(in_ch, base_ch) self.conv02 = ConvBlock(base_ch, base_ch) - self.bp0 = nn.BlurPool (filt_size=3) - + self.bp0 = nn.BlurPool (filt_size=4) self.conv11 = ConvBlock(base_ch, base_ch*2) self.conv12 = ConvBlock(base_ch*2, base_ch*2) @@ -40,19 +41,30 @@ class XSeg(nn.ModelBase): self.conv21 = ConvBlock(base_ch*2, base_ch*4) self.conv22 = ConvBlock(base_ch*4, base_ch*4) - self.conv23 = ConvBlock(base_ch*4, base_ch*4) - self.bp2 = nn.BlurPool (filt_size=3) - + self.bp2 = nn.BlurPool (filt_size=2) self.conv31 = ConvBlock(base_ch*4, base_ch*8) self.conv32 = ConvBlock(base_ch*8, base_ch*8) self.conv33 = ConvBlock(base_ch*8, base_ch*8) - self.bp3 = nn.BlurPool (filt_size=3) + self.bp3 = nn.BlurPool (filt_size=2) self.conv41 = ConvBlock(base_ch*8, base_ch*8) self.conv42 = ConvBlock(base_ch*8, base_ch*8) self.conv43 = ConvBlock(base_ch*8, base_ch*8) - self.bp4 = nn.BlurPool (filt_size=3) + self.bp4 = nn.BlurPool (filt_size=2) + + self.conv51 = ConvBlock(base_ch*8, base_ch*8) + self.conv52 = ConvBlock(base_ch*8, base_ch*8) + self.conv53 = ConvBlock(base_ch*8, base_ch*8) + self.bp5 = nn.BlurPool (filt_size=2) + + self.dense1 = nn.Dense ( 4*4* base_ch*8, 512) + self.dense2 = nn.Dense ( 512, 4*4* base_ch*8) + + self.up5 = UpConvBlock (base_ch*8, base_ch*4) + self.uconv53 = ConvBlock(base_ch*12, base_ch*8) + self.uconv52 = ConvBlock(base_ch*8, base_ch*8) + self.uconv51 = ConvBlock(base_ch*8, base_ch*8) self.up4 = UpConvBlock (base_ch*8, base_ch*4) self.uconv43 = ConvBlock(base_ch*12, base_ch*8) @@ -65,8 +77,7 @@ class XSeg(nn.ModelBase): self.uconv31 = ConvBlock(base_ch*8, base_ch*8) self.up2 = UpConvBlock (base_ch*8, base_ch*4) - self.uconv23 = ConvBlock(base_ch*8, base_ch*4) - self.uconv22 = ConvBlock(base_ch*4, base_ch*4) + self.uconv22 = ConvBlock(base_ch*8, base_ch*4) self.uconv21 = ConvBlock(base_ch*4, base_ch*4) self.up1 = UpConvBlock (base_ch*4, base_ch*2) @@ -78,8 +89,7 @@ class XSeg(nn.ModelBase): self.uconv01 = ConvBlock(base_ch, base_ch) self.out_conv = nn.Conv2D (base_ch, out_ch, kernel_size=3, padding='SAME') - self.conv_center = ConvBlock(base_ch*8, base_ch*8) - + def forward(self, inp): x = inp @@ -92,8 +102,7 @@ class XSeg(nn.ModelBase): x = self.bp1(x) x = self.conv21(x) - x = self.conv22(x) - x = x2 = self.conv23(x) + x = x2 = self.conv22(x) x = self.bp2(x) x = self.conv31(x) @@ -106,8 +115,21 @@ class XSeg(nn.ModelBase): x = x4 = self.conv43(x) x = self.bp4(x) - x = self.conv_center(x) - + x = self.conv51(x) + x = self.conv52(x) + x = x5 = self.conv53(x) + x = self.bp5(x) + + x = nn.flatten(x) + x = self.dense1(x) + x = self.dense2(x) + x = nn.reshape_4D (x, 4, 4, self.base_ch*8 ) + + x = self.up5(x) + x = self.uconv53(tf.concat([x,x5],axis=nn.conv2d_ch_axis)) + x = self.uconv52(x) + x = self.uconv51(x) + x = self.up4(x) x = self.uconv43(tf.concat([x,x4],axis=nn.conv2d_ch_axis)) x = self.uconv42(x) @@ -119,8 +141,7 @@ class XSeg(nn.ModelBase): x = self.uconv31(x) x = self.up2(x) - x = self.uconv23(tf.concat([x,x2],axis=nn.conv2d_ch_axis)) - x = self.uconv22(x) + x = self.uconv22(tf.concat([x,x2],axis=nn.conv2d_ch_axis)) x = self.uconv21(x) x = self.up1(x) diff --git a/core/leras/nn.py b/core/leras/nn.py index ef5c2c9..7c28874 100644 --- a/core/leras/nn.py +++ b/core/leras/nn.py @@ -33,14 +33,14 @@ class nn(): tf = None tf_sess = None tf_sess_config = None - tf_default_device = None - + tf_default_device_name = None + data_format = None conv2d_ch_axis = None conv2d_spatial_axes = None floatx = None - + @staticmethod def initialize(device_config=None, floatx="float32", data_format="NHWC"): @@ -50,9 +50,6 @@ class nn(): nn.setCurrentDeviceConfig(device_config) # Manipulate environment variables before import tensorflow - - if 'CUDA_VISIBLE_DEVICES' in os.environ.keys(): - os.environ.pop('CUDA_VISIBLE_DEVICES') first_run = False if len(device_config.devices) != 0: @@ -68,22 +65,19 @@ class nn(): compute_cache_path = Path(os.environ['APPDATA']) / 'NVIDIA' / ('ComputeCache' + devices_str) if not compute_cache_path.exists(): first_run = True + compute_cache_path.mkdir(parents=True, exist_ok=True) os.environ['CUDA_CACHE_PATH'] = str(compute_cache_path) - - os.environ['TF_MIN_GPU_MULTIPROCESSOR_COUNT'] = '2' - os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' # tf log errors only - + if first_run: io.log_info("Caching GPU kernels...") import tensorflow - - tf_version = getattr(tensorflow,'VERSION', None) - if tf_version is None: - tf_version = tensorflow.version.GIT_VERSION - if tf_version[0] == 'v': - tf_version = tf_version[1:] - + + tf_version = tensorflow.version.VERSION + #if tf_version is None: + # tf_version = tensorflow.version.GIT_VERSION + if tf_version[0] == 'v': + tf_version = tf_version[1:] if tf_version[0] == '2': tf = tensorflow.compat.v1 else: @@ -93,7 +87,7 @@ class nn(): # Disable tensorflow warnings tf_logger = logging.getLogger('tensorflow') tf_logger.setLevel(logging.ERROR) - + if tf_version[0] == '2': tf.disable_v2_behavior() nn.tf = tf @@ -105,20 +99,21 @@ class nn(): import core.leras.optimizers import core.leras.models import core.leras.archis - + # Configure tensorflow session-config if len(device_config.devices) == 0: - nn.tf_default_device = "/CPU:0" config = tf.ConfigProto(device_count={'GPU': 0}) + nn.tf_default_device_name = '/CPU:0' else: - nn.tf_default_device = "/GPU:0" - config = tf.ConfigProto() + nn.tf_default_device_name = f'/{device_config.devices[0].tf_dev_type}:0' + + config = tf.ConfigProto(allow_soft_placement=True) config.gpu_options.visible_device_list = ','.join([str(device.index) for device in device_config.devices]) - + config.gpu_options.force_gpu_compatible = True config.gpu_options.allow_growth = True nn.tf_sess_config = config - + if nn.tf_sess is None: nn.tf_sess = tf.Session(config=nn.tf_sess_config) @@ -202,14 +197,6 @@ class nn(): nn.tf_sess.close() nn.tf_sess = None - @staticmethod - def get_current_device(): - # Undocumented access to last tf.device(...) - objs = nn.tf.get_default_graph()._device_function_stack.peek_objs() - if len(objs) != 0: - return objs[0].display_name - return nn.tf_default_device - @staticmethod def ask_choose_device_idxs(choose_only_one=False, allow_cpu=True, suggest_best_multi_gpu=False, suggest_all_gpu=False): devices = Devices.getDevices() @@ -273,7 +260,7 @@ class nn(): @staticmethod def ask_choose_device(*args, **kwargs): return nn.DeviceConfig.GPUIndexes( nn.ask_choose_device_idxs(*args,**kwargs) ) - + def __init__ (self, devices=None): devices = devices or [] diff --git a/core/leras/ops/__init__.py b/core/leras/ops/__init__.py index 500a22a..ef4efd6 100644 --- a/core/leras/ops/__init__.py +++ b/core/leras/ops/__init__.py @@ -204,7 +204,7 @@ def random_binomial(shape, p=0.0, dtype=None, seed=None): seed = np.random.randint(10e6) return array_ops.where( random_ops.random_uniform(shape, dtype=tf.float16, seed=seed) < p, - array_ops.ones(shape, dtype=dtype), array_ops.zeros(shape, dtype=dtype)) + array_ops.ones(shape, dtype=dtype), array_ops.zeros(shape, dtype=dtype)) nn.random_binomial = random_binomial def gaussian_blur(input, radius=2.0): @@ -237,6 +237,19 @@ def gaussian_blur(input, radius=2.0): return x nn.gaussian_blur = gaussian_blur +def get_gaussian_weights(batch_size, in_ch, resolution, num_scale=5, sigma=(0.5, 1., 2., 4., 8.)): + w = np.empty((num_scale, batch_size, in_ch, resolution, resolution)) + for i in range(num_scale): + gaussian = np.exp(-1.*np.arange(-(resolution/2-0.5), resolution/2+0.5)**2/(2*sigma[i]**2)) + gaussian = np.outer(gaussian, gaussian.reshape((resolution, 1))) # extend to 2D + gaussian = gaussian/np.sum(gaussian) # normalization + gaussian = np.reshape(gaussian, (1, 1, resolution, resolution)) # reshape to 3D + gaussian = np.tile(gaussian, (batch_size, in_ch, 1, 1)) + w[i, :, :, :, :] = gaussian + return w + +nn.get_gaussian_weights = get_gaussian_weights + def style_loss(target, style, gaussian_blur_radius=0.0, loss_weight=1.0, step_size=1): def sd(content, style, loss_weight): content_nc = content.shape[ nn.conv2d_ch_axis ] @@ -333,7 +346,9 @@ def depth_to_space(x, size): x = tf.reshape(x, (-1, oh, ow, oc, )) return x else: - return tf.depth_to_space(x, size, data_format=nn.data_format) + cfg = nn.getCurrentDeviceConfig() + if not cfg.cpu_only: + return tf.depth_to_space(x, size, data_format=nn.data_format) b,c,h,w = x.shape.as_list() oh, ow = h * size, w * size oc = c // (size * size) @@ -344,11 +359,6 @@ def depth_to_space(x, size): return x nn.depth_to_space = depth_to_space -def pixel_norm(x, power = 1.0): - return x * power * tf.rsqrt(tf.reduce_mean(tf.square(x), axis=nn.conv2d_spatial_axes, keepdims=True) + 1e-06) -nn.pixel_norm = pixel_norm - - def rgb_to_lab(srgb): srgb_pixels = tf.reshape(srgb, [-1, 3]) linear_mask = tf.cast(srgb_pixels <= 0.04045, dtype=tf.float32) @@ -385,12 +395,17 @@ def total_variation_mse(images): """ pixel_dif1 = images[:, 1:, :, :] - images[:, :-1, :, :] pixel_dif2 = images[:, :, 1:, :] - images[:, :, :-1, :] - + tot_var = ( tf.reduce_sum(tf.square(pixel_dif1), axis=[1,2,3]) + tf.reduce_sum(tf.square(pixel_dif2), axis=[1,2,3]) ) return tot_var nn.total_variation_mse = total_variation_mse + +def pixel_norm(x, axes): + return x * tf.rsqrt(tf.reduce_mean(tf.square(x), axis=axes, keepdims=True) + 1e-06) +nn.pixel_norm = pixel_norm + """ def tf_suppress_lower_mean(t, eps=0.00001): if t.shape.ndims != 1: @@ -400,4 +415,4 @@ def tf_suppress_lower_mean(t, eps=0.00001): q = tf.clip_by_value(q-t_mean_eps, 0, eps) q = q * (t/eps) return q -""" \ No newline at end of file +""" diff --git a/core/mathlib/__init__.py b/core/mathlib/__init__.py index a11e725..7e5fa13 100644 --- a/core/mathlib/__init__.py +++ b/core/mathlib/__init__.py @@ -1,7 +1,12 @@ -import numpy as np import math + +import cv2 +import numpy as np +import numpy.linalg as npla + from .umeyama import umeyama + def get_power_of_two(x): i = 0 while (1 << i) < x: @@ -23,3 +28,70 @@ def rotationMatrixToEulerAngles(R) : def polygon_area(x,y): return 0.5*np.abs(np.dot(x,np.roll(y,1))-np.dot(y,np.roll(x,1))) + +def rotate_point(origin, point, deg): + """ + Rotate a point counterclockwise by a given angle around a given origin. + + The angle should be given in radians. + """ + ox, oy = origin + px, py = point + + rad = deg * math.pi / 180.0 + qx = ox + math.cos(rad) * (px - ox) - math.sin(rad) * (py - oy) + qy = oy + math.sin(rad) * (px - ox) + math.cos(rad) * (py - oy) + return np.float32([qx, qy]) + +def transform_points(points, mat, invert=False): + if invert: + mat = cv2.invertAffineTransform (mat) + points = np.expand_dims(points, axis=1) + points = cv2.transform(points, mat, points.shape) + points = np.squeeze(points) + return points + + +def transform_mat(mat, res, tx, ty, rotation, scale): + """ + transform mat in local space of res + scale -> translate -> rotate + + tx,ty float + rotation int degrees + scale float + """ + + + lt, rt, lb, ct = transform_points ( np.float32([(0,0),(res,0),(0,res),(res / 2, res/2) ]),mat, True) + + hor_v = (rt-lt).astype(np.float32) + hor_size = npla.norm(hor_v) + hor_v /= hor_size + + ver_v = (lb-lt).astype(np.float32) + ver_size = npla.norm(ver_v) + ver_v /= ver_size + + bt_diag_vec = (rt-ct).astype(np.float32) + half_diag_len = npla.norm(bt_diag_vec) + bt_diag_vec /= half_diag_len + + tb_diag_vec = np.float32( [ -bt_diag_vec[1], bt_diag_vec[0] ] ) + + rt = ct + bt_diag_vec*half_diag_len*scale + lb = ct - bt_diag_vec*half_diag_len*scale + lt = ct - tb_diag_vec*half_diag_len*scale + + rt[0] += tx*hor_size + lb[0] += tx*hor_size + lt[0] += tx*hor_size + rt[1] += ty*ver_size + lb[1] += ty*ver_size + lt[1] += ty*ver_size + + rt = rotate_point(ct, rt, rotation) + lb = rotate_point(ct, lb, rotation) + lt = rotate_point(ct, lt, rotation) + + return cv2.getAffineTransform( np.float32([lt, rt, lb]), np.float32([ [0,0], [res,0], [0,res] ]) ) diff --git a/doc/Alipay_donation.jpg b/doc/Alipay_donation.jpg deleted file mode 100644 index 48781e1..0000000 Binary files a/doc/Alipay_donation.jpg and /dev/null differ diff --git a/doc/dfl_cover.png b/doc/dfl_cover.png new file mode 100644 index 0000000..53f2b7f Binary files /dev/null and b/doc/dfl_cover.png differ diff --git a/doc/features/background-power/README.md b/doc/features/background-power/README.md new file mode 100644 index 0000000..0c83e52 --- /dev/null +++ b/doc/features/background-power/README.md @@ -0,0 +1,32 @@ +# Background Power option + +Allows you to train the model to include the background, which may help with areas around the mask. +Unlike **Background Style Power**, this does not use any additional VRAM, and does not require lowering the batch size. + +- [DESCRIPTION](#description) +- [USAGE](#usage) +- [DIFFERENCE WITH BACKGROUND STYLE POWER](#difference-with-background-style-power) + +*Examples trained with background power `0.3`:* + +![](example.jpeg)![](example2.jpeg) + +## DESCRIPTION + +Applies the same loss calculation used for the area *inside* the mask, to the area *outside* the mask, multiplied with +the chosen background power value. + +E.g. (simplified): Source Loss = Masked area image difference + Background Power * Non-masked area image difference + +## USAGE + +`[0.0] Background power ( 0.0..1.0 ?:help ) : 0.3` + +## DIFFERENCE WITH BACKGROUND STYLE POWER + +**Background Style Power** applies a loss to the source by comparing the background of the dest to that of the +predicted src/dest (5th column). This operation requires additional VRAM, due to the face that the predicted src/dest +outputs are not normally used in training (other then being viewable in the preview window). + +**Background Power** does *not* use the src/dest images whatsoever, instead comparing the background of the predicted +source to that of the original source, and the same for the background of the dest images. diff --git a/doc/features/background-power/example.jpeg b/doc/features/background-power/example.jpeg new file mode 100644 index 0000000..004d149 Binary files /dev/null and b/doc/features/background-power/example.jpeg differ diff --git a/doc/features/background-power/example2.jpeg b/doc/features/background-power/example2.jpeg new file mode 100644 index 0000000..7d00de9 Binary files /dev/null and b/doc/features/background-power/example2.jpeg differ diff --git a/doc/features/gan-options/README.md b/doc/features/gan-options/README.md new file mode 100644 index 0000000..45d1adb --- /dev/null +++ b/doc/features/gan-options/README.md @@ -0,0 +1,50 @@ +# GAN Options + +Allows you to use one-sided label smoothing and noisy labels when training the discriminator. + +- [ONE-SIDED LABEL SMOOTHING](#one-sided-label-smoothing) +- [NOISY LABELS](#noisy-labels) + +## ONE-SIDED LABEL SMOOTHING + +![](tutorial-on-theory-and-application-of-generative-adversarial-networks-54-638.jpg) + +> Deep networks may suffer from overconfidence. For example, it uses very few features to classify an object. To +> mitigate the problem, deep learning uses regulation and dropout to avoid overconfidence. +> +> In GAN, if the discriminator depends on a small set of features to detect real images, the generator may just produce +> these features only to exploit the discriminator. The optimization may turn too greedy and produces no long term +> benefit. In GAN, overconfidence hurts badly. To avoid the problem, we penalize the discriminator when the prediction +> for any real images go beyond 0.9 (D(real image)>0.9). This is done by setting our target label value to be 0.9 +> instead of 1.0. + - [GAN — Ways to improve GAN performance](https://towardsdatascience.com/gan-ways-to-improve-gan-performance-acf37f9f59b) + +By setting the label smoothing value to any value > 0, the target label value used with the discriminator will be: +``` +target label value = 1 - (label smoothing value) +``` +### USAGE + +``` +[0.1] GAN label smoothing ( 0 - 0.5 ?:help ) : 0.1 +``` + +## NOISY LABELS + +> make the labels the noisy for the discriminator: occasionally flip the labels when training the discriminator + - [How to Train a GAN? Tips and tricks to make GANs work](https://github.com/soumith/ganhacks/blob/master/README.md#6-use-soft-and-noisy-labels) + +By setting the noisy labels value to any value > 0, then the target labels used with the discriminator will be flipped +("fake" => "real" / "real" => "fake") with probability p (where p is the noisy label value). + +E.g., if the value is 0.05, then ~5% of the labels will be flipped when training the discriminator + +### USAGE +``` +[0.05] GAN noisy labels ( 0 - 0.5 ?:help ) : 0.05 +``` + + + + + diff --git a/doc/features/gan-options/tutorial-on-theory-and-application-of-generative-adversarial-networks-54-638.jpg b/doc/features/gan-options/tutorial-on-theory-and-application-of-generative-adversarial-networks-54-638.jpg new file mode 100644 index 0000000..e3698e9 Binary files /dev/null and b/doc/features/gan-options/tutorial-on-theory-and-application-of-generative-adversarial-networks-54-638.jpg differ diff --git a/doc/features/ms-ssim/README.md b/doc/features/ms-ssim/README.md new file mode 100644 index 0000000..41039d3 --- /dev/null +++ b/doc/features/ms-ssim/README.md @@ -0,0 +1,43 @@ +# Multiscale SSIM (MS-SSIM) + +Allows you to train using the MS-SSIM (multiscale structural similarity index measure) as the main loss metric, +a perceptually more accurate measure of image quality than MSE (mean squared error). + +As an added benefit, you may see a decrease in ms/iteration (when using the same batch size) with Multiscale loss +enabled. You may also be able to train with a larger batch size with it enabled. + +- [DESCRIPTION](#description) +- [USAGE](#usage) + +## DESCRIPTION + +[SSIM](https://en.wikipedia.org/wiki/Structural_similarity) is metric for comparing the perceptial quality of an image: +> SSIM is a perception-based model that considers image degradation as perceived change in structural information, +> while also incorporating important perceptual phenomena, including both luminance masking and contrast masking terms. +> [...] +> Structural information is the idea that the pixels have strong inter-dependencies especially when they are spatially +> close. These dependencies carry important information about the structure of the objects in the visual scene. +> Luminance masking is a phenomenon whereby image distortions (in this context) tend to be less visible in bright +> regions, while contrast masking is a phenomenon whereby distortions become less visible where there is significant +> activity or "texture" in the image. + +The current loss metric is a combination of SSIM (structural similarity index measure) and +[MSE](https://en.wikipedia.org/wiki/Mean_squared_error) (mean squared error). + +[Multiscale SSIM](https://en.wikipedia.org/wiki/Structural_similarity#Multi-Scale_SSIM) is a variant of SSIM that +improves upon SSIM by comparing the similarity at multiple scales (e.g.: full-size, half-size, 1/4 size, etc.) +By using MS-SSIM as our main loss metric, we should expect the image similarity to improve across each scale, improving +both the large scale and small scale detail of the predicted images. + +Original paper: [Wang, Zhou, Eero P. Simoncelli, and Alan C. Bovik. +"Multiscale structural similarity for image quality assessment." +Signals, Systems and Computers, 2004.](https://www.cns.nyu.edu/pub/eero/wang03b.pdf) + +## USAGE + +``` +[n] Use multiscale loss? ( y/n ?:help ) : y +``` + + + diff --git a/doc/features/random-color/README.md b/doc/features/random-color/README.md new file mode 100644 index 0000000..d1aeac1 --- /dev/null +++ b/doc/features/random-color/README.md @@ -0,0 +1,25 @@ +# Random Color option + +Helps train the model to generalize perceptual color and lightness, and improves color transfer between src and dst. + +- [DESCRIPTION](#description) +- [USAGE](#usage) + +![](example.jpeg) + +## DESCRIPTION + +Converts images to [CIE L\*a\*b* colorspace](https://en.wikipedia.org/wiki/CIELAB_color_space), +and then randomly rotates around the `L*` axis. While the perceptual lightness stays constant, only the `a*` and `b*` +color channels are modified. After rotation, converts back to BGR (blue/green/red) colorspace. + +If visualized using the [CIE L\*a\*b* cylindical model](https://en.wikipedia.org/wiki/CIELAB_color_space#Cylindrical_model), +this is a random rotation of `h°` (hue angle, angle of the hue in the CIELAB color wheel), +maintaining the same `C*` (chroma, relative saturation). + +## USAGE + +``` +[n] Random color ( y/n ?:help ) : y +``` + diff --git a/doc/features/random-color/example.jpeg b/doc/features/random-color/example.jpeg new file mode 100644 index 0000000..2a69632 Binary files /dev/null and b/doc/features/random-color/example.jpeg differ diff --git a/doc/features/webui/README.md b/doc/features/webui/README.md new file mode 100644 index 0000000..c181fca --- /dev/null +++ b/doc/features/webui/README.md @@ -0,0 +1,45 @@ +# Web UI + +View and interact with the training preview window with your web browser. +Allows you to view and control the preview remotely, and train on headless machines. + +- [INSTALLATION](#installation) +- [DESCRIPTION](#description) +- [USAGE](#usage) +- [SSH PORT FORWARDING](#ssh-port-forwarding) + +![](example.png) + +## INSTALLATION + +Requires additional Python dependencies to be installed: +- [Flask](https://palletsprojects.com/p/flask/), + version [1.1.1](https://pypi.org/project/Flask/1.1.1/) +- [Flask-SocketIO](https://github.com/miguelgrinberg/Flask-SocketIO/), + version [4.2.1](https://pypi.org/project/Flask-SocketIO/4.2.1/) + +``` +pip install Flask==1.1.1 +pip install Flask-SocketIO==4.2.1 +``` + +## DESCRIPTION + +Launches a Flask web application which sends commands to the training thread +(save/exit/fetch new preview, etc.), and displays live updates for the log output +e.g.: `[09:50:53][#106913][0503ms][0.3109][0.2476]`, and updates the graph/preview image. + +## USAGE + +Enable the Web UI by appending `--flask-preview` to the `train` command. +Once training begins, Web UI will start, and can be accessed at http://localhost:5000/ + +## SSH PORT FORWARDING + +When running on a remote/headless box, view the Web UI in your local browser simply by +adding the ssh option `-L 5000:localhost:5000`. Once connected, the Web UI can be viewed +locally at http://localhost:5000/ + +Several Android/iOS SSH apps (such as [JuiceSSH](https://juicessh.com/) +exist which support port forwarding, allowing you to interact with the preview pane +from anywhere with your phone. diff --git a/doc/features/webui/example.png b/doc/features/webui/example.png new file mode 100644 index 0000000..4cc7c05 Binary files /dev/null and b/doc/features/webui/example.png differ diff --git a/doc/fixes/predicted_src_mask/README.md b/doc/fixes/predicted_src_mask/README.md new file mode 100644 index 0000000..1247442 --- /dev/null +++ b/doc/fixes/predicted_src_mask/README.md @@ -0,0 +1,5 @@ +# Example of bug: +![](preview_image_bug.jpeg) + +# Demonstration of fix: +![](preview_image_fix.jpeg) diff --git a/doc/fixes/predicted_src_mask/preview_image_bug.jpeg b/doc/fixes/predicted_src_mask/preview_image_bug.jpeg new file mode 100644 index 0000000..52767fa Binary files /dev/null and b/doc/fixes/predicted_src_mask/preview_image_bug.jpeg differ diff --git a/doc/fixes/predicted_src_mask/preview_image_fix.jpeg b/doc/fixes/predicted_src_mask/preview_image_fix.jpeg new file mode 100644 index 0000000..43a8b19 Binary files /dev/null and b/doc/fixes/predicted_src_mask/preview_image_fix.jpeg differ diff --git a/doc/logo_directx.png b/doc/logo_directx.png new file mode 100644 index 0000000..f9fb10a Binary files /dev/null and b/doc/logo_directx.png differ diff --git a/facelib/FaceEnhancer.py b/facelib/FaceEnhancer.py index 0b5ced3..1dc0dd9 100644 --- a/facelib/FaceEnhancer.py +++ b/facelib/FaceEnhancer.py @@ -161,11 +161,11 @@ class FaceEnhancer(object): if not model_path.exists(): raise Exception("Unable to load FaceEnhancer.npy") - with tf.device ('/CPU:0' if place_model_on_cpu else '/GPU:0'): + with tf.device ('/CPU:0' if place_model_on_cpu else nn.tf_default_device_name): self.model = FaceEnhancer() self.model.load_weights (model_path) - with tf.device ('/CPU:0' if run_on_cpu else '/GPU:0'): + with tf.device ('/CPU:0' if run_on_cpu else nn.tf_default_device_name): self.model.build_for_run ([ (tf.float32, nn.get4Dshape (192,192,3) ), (tf.float32, (None,1,) ), (tf.float32, (None,1,) ), diff --git a/facelib/XSegNet.py b/facelib/XSegNet.py index 761ab94..5621a65 100644 --- a/facelib/XSegNet.py +++ b/facelib/XSegNet.py @@ -39,7 +39,7 @@ class XSegNet(object): self.target_t = tf.placeholder (nn.floatx, nn.get4Dshape(resolution,resolution,1) ) # Initializing model classes - with tf.device ('/CPU:0' if place_model_on_cpu else '/GPU:0'): + with tf.device ('/CPU:0' if place_model_on_cpu else nn.tf_default_device_name): self.model = nn.XSeg(3, 32, 1, name=name) self.model_weights = self.model.get_weights() if training: @@ -53,7 +53,7 @@ class XSegNet(object): self.model_filename_list += [ [self.model, f'{model_name}.npy'] ] if not training: - with tf.device ('/CPU:0' if run_on_cpu else '/GPU:0'): + with tf.device ('/CPU:0' if run_on_cpu else nn.tf_default_device_name): _, pred = self.model(self.input_t) def net_run(input_np): diff --git a/flaskr/__init__.py b/flaskr/__init__.py new file mode 100644 index 0000000..e69de29 diff --git a/flaskr/app.py b/flaskr/app.py new file mode 100644 index 0000000..b56943e --- /dev/null +++ b/flaskr/app.py @@ -0,0 +1,102 @@ +from pathlib import Path + +from flask import Flask, send_file, Response, render_template, render_template_string, request, g +from flask_socketio import SocketIO, emit +import logging + + +def create_flask_app(s2c, c2s, s2flask, kwargs): + app = Flask(__name__, template_folder="templates", static_folder="static") + log = logging.getLogger('werkzeug') + log.disabled = True + model_path = Path(kwargs.get('saved_models_path', '')) + filename = 'preview.png' + preview_file = str(model_path / filename) + + def gen(): + frame = open(preview_file, 'rb').read() + while True: + try: + frame = open(preview_file, 'rb').read() + except: + pass + yield b'--frame\r\nContent-Type: image/png\r\n\r\n' + yield frame + yield b'\r\n\r\n' + + def send(queue, op): + queue.put({'op': op}) + + def send_and_wait(queue, op): + while not s2flask.empty(): + s2flask.get() + queue.put({'op': op}) + while s2flask.empty(): + pass + s2flask.get() + + @app.route('/save', methods=['POST']) + def save(): + send(s2c, 'save') + return '', 204 + + @app.route('/exit', methods=['POST']) + def exit(): + send(c2s, 'close') + request.environ.get('werkzeug.server.shutdown')() + return '', 204 + + @app.route('/update', methods=['POST']) + def update(): + send(c2s, 'update') + return '', 204 + + @app.route('/next_preview', methods=['POST']) + def next_preview(): + send(c2s, 'next_preview') + return '', 204 + + @app.route('/change_history_range', methods=['POST']) + def change_history_range(): + send(c2s, 'change_history_range') + return '', 204 + + @app.route('/zoom_prev', methods=['POST']) + def zoom_prev(): + send(c2s, 'zoom_prev') + return '', 204 + + @app.route('/zoom_next', methods=['POST']) + def zoom_next(): + send(c2s, 'zoom_next') + return '', 204 + + @app.route('/') + def index(): + return render_template('index.html') + + # @app.route('/preview_image') + # def preview_image(): + # return Response(gen(), mimetype='multipart/x-mixed-replace;boundary=frame') + + @app.route('/preview_image') + def preview_image(): + return send_file(preview_file, mimetype='image/png', cache_timeout=-1) + + socketio = SocketIO(app) + + @socketio.on('connect', namespace='/') + def test_connect(): + emit('my response', {'data': 'Connected'}) + + @socketio.on('disconnect', namespace='/test') + def test_disconnect(): + print('Client disconnected') + + return socketio, app + + + + + + diff --git a/flaskr/static/favicon.ico b/flaskr/static/favicon.ico new file mode 100644 index 0000000..46aec07 Binary files /dev/null and b/flaskr/static/favicon.ico differ diff --git a/flaskr/templates/index.html b/flaskr/templates/index.html new file mode 100644 index 0000000..4ab78dd --- /dev/null +++ b/flaskr/templates/index.html @@ -0,0 +1,95 @@ + + + + + + + Training Preview + + + + + +
Training Preview
+
+
+ + + + + + + +
+ + + diff --git a/main.py b/main.py index 73ec718..f7616bb 100644 --- a/main.py +++ b/main.py @@ -23,7 +23,7 @@ if __name__ == "__main__": setattr(namespace, self.dest, os.path.abspath(os.path.expanduser(values))) exit_code = 0 - + parser = argparse.ArgumentParser() subparsers = parser.add_subparsers() @@ -52,9 +52,9 @@ if __name__ == "__main__": p.add_argument('--output-debug', action="store_true", dest="output_debug", default=None, help="Writes debug images to _debug\ directory.") p.add_argument('--no-output-debug', action="store_false", dest="output_debug", default=None, help="Don't writes debug images to _debug\ directory.") p.add_argument('--face-type', dest="face_type", choices=['half_face', 'full_face', 'whole_face', 'head', 'mark_only'], default=None) - p.add_argument('--max-faces-from-image', type=int, dest="max_faces_from_image", default=None, help="Max faces from image.") + p.add_argument('--max-faces-from-image', type=int, dest="max_faces_from_image", default=None, help="Max faces from image.") p.add_argument('--image-size', type=int, dest="image_size", default=None, help="Output image size.") - p.add_argument('--jpeg-quality', type=int, dest="jpeg_quality", default=None, help="Jpeg quality.") + p.add_argument('--jpeg-quality', type=int, dest="jpeg_quality", default=None, help="Jpeg quality.") p.add_argument('--manual-fix', action="store_true", dest="manual_fix", default=False, help="Enables manual extract only frames where faces were not recognized.") p.add_argument('--manual-output-debug-fix', action="store_true", dest="manual_output_debug_fix", default=False, help="Performs manual reextract input-dir frames which were deleted from [output_dir]_debug\ dir.") p.add_argument('--manual-window-size', type=int, dest="manual_window_size", default=1368, help="Manual fix window size. Default: 1368.") @@ -128,7 +128,9 @@ if __name__ == "__main__": 'execute_programs' : [ [int(x[0]), x[1] ] for x in arguments.execute_program ], 'debug' : arguments.debug, 'tensorboard_dir' : arguments.tensorboard_dir, - 'start_tensorboard' : arguments.start_tensorboard + 'start_tensorboard' : arguments.start_tensorboard, + 'dump_ckpt' : arguments.dump_ckpt, + 'flask_preview' : arguments.flask_preview, } from mainscripts import Trainer Trainer.main(**kwargs) @@ -150,6 +152,10 @@ if __name__ == "__main__": p.add_argument('--start-tensorboard', action="store_true", dest="start_tensorboard", default=False, help="Automatically start the tensorboard server preconfigured to the tensorboard-logdir") + p.add_argument('--dump-ckpt', action="store_true", dest="dump_ckpt", default=False, help="Dump the model to ckpt format.") + p.add_argument('--flask-preview', action="store_true", dest="flask_preview", default=False, + help="Launches a flask server to view the previews in a web browser") + p.add_argument('--execute-program', dest="execute_program", default=[], action='append', nargs='+') p.set_defaults (func=process_train) @@ -257,6 +263,16 @@ if __name__ == "__main__": p.set_defaults(func=process_faceset_enhancer) + + p = facesettool_parser.add_parser ("resize", help="Resize DFL faceset.") + p.add_argument('--input-dir', required=True, action=fixPathAction, dest="input_dir", help="Input directory of aligned faces.") + + def process_faceset_resizer(arguments): + osex.set_process_lowest_prio() + from mainscripts import FacesetResizer + FacesetResizer.process_folder ( Path(arguments.input_dir) ) + p.set_defaults(func=process_faceset_resizer) + def process_dev_test(arguments): osex.set_process_lowest_prio() from mainscripts import dev_misc @@ -265,10 +281,10 @@ if __name__ == "__main__": p = subparsers.add_parser( "dev_test", help="") p.add_argument('--input-dir', required=True, action=fixPathAction, dest="input_dir") p.set_defaults (func=process_dev_test) - + # ========== XSeg xseg_parser = subparsers.add_parser( "xseg", help="XSeg tools.").add_subparsers() - + p = xseg_parser.add_parser( "editor", help="XSeg editor.") def process_xsegeditor(arguments): @@ -276,11 +292,11 @@ if __name__ == "__main__": from XSegEditor import XSegEditor global exit_code exit_code = XSegEditor.start (Path(arguments.input_dir)) - + p.add_argument('--input-dir', required=True, action=fixPathAction, dest="input_dir") p.set_defaults (func=process_xsegeditor) - + p = xseg_parser.add_parser( "apply", help="Apply trained XSeg model to the extracted faces.") def process_xsegapply(arguments): @@ -290,8 +306,8 @@ if __name__ == "__main__": p.add_argument('--input-dir', required=True, action=fixPathAction, dest="input_dir") p.add_argument('--model-dir', required=True, action=fixPathAction, dest="model_dir") p.set_defaults (func=process_xsegapply) - - + + p = xseg_parser.add_parser( "remove", help="Remove applied XSeg masks from the extracted faces.") def process_xsegremove(arguments): osex.set_process_lowest_prio() @@ -299,8 +315,8 @@ if __name__ == "__main__": XSegUtil.remove_xseg (Path(arguments.input_dir) ) p.add_argument('--input-dir', required=True, action=fixPathAction, dest="input_dir") p.set_defaults (func=process_xsegremove) - - + + p = xseg_parser.add_parser( "remove_labels", help="Remove XSeg labels from the extracted faces.") def process_xsegremovelabels(arguments): osex.set_process_lowest_prio() @@ -308,8 +324,8 @@ if __name__ == "__main__": XSegUtil.remove_xseg_labels (Path(arguments.input_dir) ) p.add_argument('--input-dir', required=True, action=fixPathAction, dest="input_dir") p.set_defaults (func=process_xsegremovelabels) - - + + p = xseg_parser.add_parser( "fetch", help="Copies faces containing XSeg polygons in _xseg dir.") def process_xsegfetch(arguments): @@ -318,7 +334,7 @@ if __name__ == "__main__": XSegUtil.fetch_xseg (Path(arguments.input_dir) ) p.add_argument('--input-dir', required=True, action=fixPathAction, dest="input_dir") p.set_defaults (func=process_xsegfetch) - + def bad_args(arguments): parser.print_help() exit(0) @@ -329,9 +345,9 @@ if __name__ == "__main__": if exit_code == 0: print ("Done.") - + exit(exit_code) - + ''' import code code.interact(local=dict(globals(), **locals())) diff --git a/mainscripts/Extractor.py b/mainscripts/Extractor.py index e1a1fa7..f9be605 100644 --- a/mainscripts/Extractor.py +++ b/mainscripts/Extractor.py @@ -97,9 +97,6 @@ class ExtractSubprocessor(Subprocessor): h, w, c = image.shape - dflimg = DFLIMG.load (filepath) - extract_from_dflimg = (h == w and (dflimg is not None and dflimg.has_data()) ) - if 'rects' in self.type or self.type == 'all': data = ExtractSubprocessor.Cli.rects_stage (data=data, image=image, @@ -110,7 +107,6 @@ class ExtractSubprocessor(Subprocessor): if 'landmarks' in self.type or self.type == 'all': data = ExtractSubprocessor.Cli.landmarks_stage (data=data, image=image, - extract_from_dflimg=extract_from_dflimg, landmarks_extractor=self.landmarks_extractor, rects_extractor=self.rects_extractor, ) @@ -121,7 +117,6 @@ class ExtractSubprocessor(Subprocessor): face_type=self.face_type, image_size=self.image_size, jpeg_quality=self.jpeg_quality, - extract_from_dflimg=extract_from_dflimg, output_debug_path=self.output_debug_path, final_output_path=self.final_output_path, ) @@ -161,7 +156,6 @@ class ExtractSubprocessor(Subprocessor): @staticmethod def landmarks_stage(data, image, - extract_from_dflimg, landmarks_extractor, rects_extractor, ): @@ -176,7 +170,7 @@ class ExtractSubprocessor(Subprocessor): elif data.rects_rotation == 270: rotated_image = image.swapaxes( 0,1 )[::-1,:,:] - data.landmarks = landmarks_extractor.extract (rotated_image, data.rects, rects_extractor if (not extract_from_dflimg and data.landmarks_accurate) else None, is_bgr=True) + data.landmarks = landmarks_extractor.extract (rotated_image, data.rects, rects_extractor if (data.landmarks_accurate) else None, is_bgr=True) if data.rects_rotation != 0: for i, (rect, lmrks) in enumerate(zip(data.rects, data.landmarks)): new_rect, new_lmrks = rect, lmrks @@ -207,7 +201,6 @@ class ExtractSubprocessor(Subprocessor): face_type, image_size, jpeg_quality, - extract_from_dflimg = False, output_debug_path=None, final_output_path=None, ): @@ -219,72 +212,53 @@ class ExtractSubprocessor(Subprocessor): if output_debug_path is not None: debug_image = image.copy() - if extract_from_dflimg and len(rects) != 1: - #if re-extracting from dflimg and more than 1 or zero faces detected - dont process and just copy it - print("extract_from_dflimg and len(rects) != 1", filepath ) - output_filepath = final_output_path / filepath.name - if filepath != str(output_file): - shutil.copy ( str(filepath), str(output_filepath) ) - data.final_output_files.append (output_filepath) - else: - face_idx = 0 - for rect, image_landmarks in zip( rects, landmarks ): + face_idx = 0 + for rect, image_landmarks in zip( rects, landmarks ): + if image_landmarks is None: + continue - if extract_from_dflimg and face_idx > 1: - #cannot extract more than 1 face from dflimg - break + rect = np.array(rect) - if image_landmarks is None: + if face_type == FaceType.MARK_ONLY: + image_to_face_mat = None + face_image = image + face_image_landmarks = image_landmarks + else: + image_to_face_mat = LandmarksProcessor.get_transform_mat (image_landmarks, image_size, face_type) + + face_image = cv2.warpAffine(image, image_to_face_mat, (image_size, image_size), cv2.INTER_LANCZOS4) + face_image_landmarks = LandmarksProcessor.transform_points (image_landmarks, image_to_face_mat) + + landmarks_bbox = LandmarksProcessor.transform_points ( [ (0,0), (0,image_size-1), (image_size-1, image_size-1), (image_size-1,0) ], image_to_face_mat, True) + + rect_area = mathlib.polygon_area(np.array(rect[[0,2,2,0]]).astype(np.float32), np.array(rect[[1,1,3,3]]).astype(np.float32)) + landmarks_area = mathlib.polygon_area(landmarks_bbox[:,0].astype(np.float32), landmarks_bbox[:,1].astype(np.float32) ) + + if not data.manual and face_type <= FaceType.FULL_NO_ALIGN and landmarks_area > 4*rect_area: #get rid of faces which umeyama-landmark-area > 4*detector-rect-area continue - rect = np.array(rect) + if output_debug_path is not None: + LandmarksProcessor.draw_rect_landmarks (debug_image, rect, image_landmarks, face_type, image_size, transparent_mask=True) - if face_type == FaceType.MARK_ONLY: - image_to_face_mat = None - face_image = image - face_image_landmarks = image_landmarks - else: - image_to_face_mat = LandmarksProcessor.get_transform_mat (image_landmarks, image_size, face_type) + output_path = final_output_path + if data.force_output_path is not None: + output_path = data.force_output_path - face_image = cv2.warpAffine(image, image_to_face_mat, (image_size, image_size), cv2.INTER_LANCZOS4) - face_image_landmarks = LandmarksProcessor.transform_points (image_landmarks, image_to_face_mat) + output_filepath = output_path / f"{filepath.stem}_{face_idx}.jpg" + cv2_imwrite(output_filepath, face_image, [int(cv2.IMWRITE_JPEG_QUALITY), jpeg_quality ] ) - landmarks_bbox = LandmarksProcessor.transform_points ( [ (0,0), (0,image_size-1), (image_size-1, image_size-1), (image_size-1,0) ], image_to_face_mat, True) + dflimg = DFLJPG.load(output_filepath) + dflimg.set_face_type(FaceType.toString(face_type)) + dflimg.set_landmarks(face_image_landmarks.tolist()) + dflimg.set_source_filename(filepath.name) + dflimg.set_source_rect(rect) + dflimg.set_source_landmarks(image_landmarks.tolist()) + dflimg.set_image_to_face_mat(image_to_face_mat) + dflimg.save() - rect_area = mathlib.polygon_area(np.array(rect[[0,2,2,0]]).astype(np.float32), np.array(rect[[1,1,3,3]]).astype(np.float32)) - landmarks_area = mathlib.polygon_area(landmarks_bbox[:,0].astype(np.float32), landmarks_bbox[:,1].astype(np.float32) ) - - if not data.manual and face_type <= FaceType.FULL_NO_ALIGN and landmarks_area > 4*rect_area: #get rid of faces which umeyama-landmark-area > 4*detector-rect-area - continue - - if output_debug_path is not None: - LandmarksProcessor.draw_rect_landmarks (debug_image, rect, image_landmarks, face_type, image_size, transparent_mask=True) - - output_path = final_output_path - if data.force_output_path is not None: - output_path = data.force_output_path - - if extract_from_dflimg and filepath.suffix == '.jpg': - #if extracting from dflimg and jpg copy it in order not to lose quality - output_filepath = output_path / filepath.name - if filepath != output_filepath: - shutil.copy ( str(filepath), str(output_filepath) ) - else: - output_filepath = output_path / f"{filepath.stem}_{face_idx}.jpg" - cv2_imwrite(output_filepath, face_image, [int(cv2.IMWRITE_JPEG_QUALITY), jpeg_quality ] ) - - dflimg = DFLJPG.load(output_filepath) - dflimg.set_face_type(FaceType.toString(face_type)) - dflimg.set_landmarks(face_image_landmarks.tolist()) - dflimg.set_source_filename(filepath.name) - dflimg.set_source_rect(rect) - dflimg.set_source_landmarks(image_landmarks.tolist()) - dflimg.set_image_to_face_mat(image_to_face_mat) - dflimg.save() - - data.final_output_files.append (output_filepath) - face_idx += 1 - data.faces_detected = face_idx + data.final_output_files.append (output_filepath) + face_idx += 1 + data.faces_detected = face_idx if output_debug_path is not None: cv2_imwrite( output_debug_path / (filepath.stem+'.jpg'), debug_image, [int(cv2.IMWRITE_JPEG_QUALITY), 50] ) diff --git a/mainscripts/FacesetResizer.py b/mainscripts/FacesetResizer.py new file mode 100644 index 0000000..acc4ea4 --- /dev/null +++ b/mainscripts/FacesetResizer.py @@ -0,0 +1,209 @@ +import multiprocessing +import shutil + +import cv2 +from core import pathex +from core.cv2ex import * +from core.interact import interact as io +from core.joblib import Subprocessor +from DFLIMG import * +from facelib import FaceType, LandmarksProcessor + + +class FacesetResizerSubprocessor(Subprocessor): + + #override + def __init__(self, image_paths, output_dirpath, image_size, face_type=None): + self.image_paths = image_paths + self.output_dirpath = output_dirpath + self.image_size = image_size + self.face_type = face_type + self.result = [] + + super().__init__('FacesetResizer', FacesetResizerSubprocessor.Cli, 600) + + #override + def on_clients_initialized(self): + io.progress_bar (None, len (self.image_paths)) + + #override + def on_clients_finalized(self): + io.progress_bar_close() + + #override + def process_info_generator(self): + base_dict = {'output_dirpath':self.output_dirpath, 'image_size':self.image_size, 'face_type':self.face_type} + + for device_idx in range( min(8, multiprocessing.cpu_count()) ): + client_dict = base_dict.copy() + device_name = f'CPU #{device_idx}' + client_dict['device_name'] = device_name + yield device_name, {}, client_dict + + #override + def get_data(self, host_dict): + if len (self.image_paths) > 0: + return self.image_paths.pop(0) + + #override + def on_data_return (self, host_dict, data): + self.image_paths.insert(0, data) + + #override + def on_result (self, host_dict, data, result): + io.progress_bar_inc(1) + if result[0] == 1: + self.result +=[ (result[1], result[2]) ] + + #override + def get_result(self): + return self.result + + class Cli(Subprocessor.Cli): + + #override + def on_initialize(self, client_dict): + self.output_dirpath = client_dict['output_dirpath'] + self.image_size = client_dict['image_size'] + self.face_type = client_dict['face_type'] + self.log_info (f"Running on { client_dict['device_name'] }") + + #override + def process_data(self, filepath): + try: + dflimg = DFLIMG.load (filepath) + if dflimg is None or not dflimg.has_data(): + self.log_err (f"{filepath.name} is not a dfl image file") + else: + img = cv2_imread(filepath) + h,w = img.shape[:2] + if h != w: + raise Exception(f'w != h in {filepath}') + + image_size = self.image_size + face_type = self.face_type + output_filepath = self.output_dirpath / filepath.name + + if face_type is not None: + lmrks = dflimg.get_landmarks() + mat = LandmarksProcessor.get_transform_mat(lmrks, image_size, face_type) + + img = cv2.warpAffine(img, mat, (image_size, image_size), flags=cv2.INTER_LANCZOS4 ) + img = np.clip(img, 0, 255).astype(np.uint8) + + cv2_imwrite ( str(output_filepath), img, [int(cv2.IMWRITE_JPEG_QUALITY), 100] ) + + dfl_dict = dflimg.get_dict() + dflimg = DFLIMG.load (output_filepath) + dflimg.set_dict(dfl_dict) + + xseg_mask = dflimg.get_xseg_mask() + if xseg_mask is not None: + xseg_res = 256 + + xseg_lmrks = lmrks.copy() + xseg_lmrks *= (xseg_res / w) + xseg_mat = LandmarksProcessor.get_transform_mat(xseg_lmrks, xseg_res, face_type) + + xseg_mask = cv2.warpAffine(xseg_mask, xseg_mat, (xseg_res, xseg_res), flags=cv2.INTER_LANCZOS4 ) + xseg_mask[xseg_mask < 0.5] = 0 + xseg_mask[xseg_mask >= 0.5] = 1 + + dflimg.set_xseg_mask(xseg_mask) + + seg_ie_polys = dflimg.get_seg_ie_polys() + + for poly in seg_ie_polys.get_polys(): + poly_pts = poly.get_pts() + poly_pts = LandmarksProcessor.transform_points(poly_pts, mat) + poly.set_points(poly_pts) + + dflimg.set_seg_ie_polys(seg_ie_polys) + + lmrks = LandmarksProcessor.transform_points(lmrks, mat) + dflimg.set_landmarks(lmrks) + + image_to_face_mat = dflimg.get_image_to_face_mat() + if image_to_face_mat is not None: + image_to_face_mat = LandmarksProcessor.get_transform_mat ( dflimg.get_source_landmarks(), image_size, face_type ) + dflimg.set_image_to_face_mat(image_to_face_mat) + dflimg.set_face_type( FaceType.toString(face_type) ) + dflimg.save() + + else: + dfl_dict = dflimg.get_dict() + + scale = w / image_size + + img = cv2.resize(img, (image_size, image_size), interpolation=cv2.INTER_LANCZOS4) + + cv2_imwrite ( str(output_filepath), img, [int(cv2.IMWRITE_JPEG_QUALITY), 100] ) + + dflimg = DFLIMG.load (output_filepath) + dflimg.set_dict(dfl_dict) + + lmrks = dflimg.get_landmarks() + lmrks /= scale + dflimg.set_landmarks(lmrks) + + seg_ie_polys = dflimg.get_seg_ie_polys() + seg_ie_polys.mult_points( 1.0 / scale) + dflimg.set_seg_ie_polys(seg_ie_polys) + + image_to_face_mat = dflimg.get_image_to_face_mat() + + if image_to_face_mat is not None: + face_type = FaceType.fromString ( dflimg.get_face_type() ) + image_to_face_mat = LandmarksProcessor.get_transform_mat ( dflimg.get_source_landmarks(), image_size, face_type ) + dflimg.set_image_to_face_mat(image_to_face_mat) + dflimg.save() + + return (1, filepath, output_filepath) + except: + self.log_err (f"Exception occured while processing file {filepath}. Error: {traceback.format_exc()}") + + return (0, filepath, None) + +def process_folder ( dirpath): + + image_size = io.input_int(f"New image size", 512, valid_range=[256,2048]) + + face_type = io.input_str ("Change face type", 'same', ['h','mf','f','wf','head','same']).lower() + if face_type == 'same': + face_type = None + else: + face_type = {'h' : FaceType.HALF, + 'mf' : FaceType.MID_FULL, + 'f' : FaceType.FULL, + 'wf' : FaceType.WHOLE_FACE, + 'head' : FaceType.HEAD}[face_type] + + + output_dirpath = dirpath.parent / (dirpath.name + '_resized') + output_dirpath.mkdir (exist_ok=True, parents=True) + + dirpath_parts = '/'.join( dirpath.parts[-2:]) + output_dirpath_parts = '/'.join( output_dirpath.parts[-2:] ) + io.log_info (f"Resizing faceset in {dirpath_parts}") + io.log_info ( f"Processing to {output_dirpath_parts}") + + output_images_paths = pathex.get_image_paths(output_dirpath) + if len(output_images_paths) > 0: + for filename in output_images_paths: + Path(filename).unlink() + + image_paths = [Path(x) for x in pathex.get_image_paths( dirpath )] + result = FacesetResizerSubprocessor ( image_paths, output_dirpath, image_size, face_type).run() + + is_merge = io.input_bool (f"\r\nMerge {output_dirpath_parts} to {dirpath_parts} ?", True) + if is_merge: + io.log_info (f"Copying processed files to {dirpath_parts}") + + for (filepath, output_filepath) in result: + try: + shutil.copy (output_filepath, filepath) + except: + pass + + io.log_info (f"Removing {output_dirpath_parts}") + shutil.rmtree(output_dirpath) diff --git a/mainscripts/Trainer.py b/mainscripts/Trainer.py index bd1c6c7..7b4b40b 100644 --- a/mainscripts/Trainer.py +++ b/mainscripts/Trainer.py @@ -1,8 +1,11 @@ -import sys +import os +import sys import traceback import queue import threading import time +from enum import Enum + import numpy as np import itertools from pathlib import Path @@ -48,6 +51,7 @@ def log_tensorboard_model_previews(iter, model, train_summary_writer): log_tensorboard_previews(iter, model.get_static_previews(), 'static_preview', train_summary_writer) def trainerThread (s2c, c2s, e, + socketio=None, model_class_name = None, saved_models_path = None, training_data_src_path = None, @@ -63,6 +67,7 @@ def trainerThread (s2c, c2s, e, debug=False, tensorboard_dir=None, start_tensorboard=False, + dump_ckpt=False, **kwargs): while True: try: @@ -79,9 +84,12 @@ def trainerThread (s2c, c2s, e, if not saved_models_path.exists(): saved_models_path.mkdir(exist_ok=True, parents=True) - + + if dump_ckpt: + cpu_only=True + model = models.import_model(model_class_name)( - is_training=True, + is_training=not dump_ckpt, saved_models_path=saved_models_path, training_data_src_path=training_data_src_path, training_data_dst_path=training_data_dst_path, @@ -92,9 +100,13 @@ def trainerThread (s2c, c2s, e, force_gpu_idxs=force_gpu_idxs, cpu_only=cpu_only, silent_start=silent_start, - debug=debug, - ) + debug=debug) + if dump_ckpt: + e.set() + model.dump_ckpt() + break + is_reached_goal = model.is_reached_iter_goal() if tensorboard_dir is not None: @@ -107,70 +119,55 @@ def trainerThread (s2c, c2s, e, }) shared_state = { 'after_save' : False } + shared_state = {'after_save': False} loss_string = "" - save_iter = model.get_iter() + save_iter = model.get_iter() + def model_save(): if not debug and not is_reached_goal: - io.log_info ("Saving....", end='\r') + io.log_info("Saving....", end='\r') model.save() shared_state['after_save'] = True - + def model_backup(): if not debug and not is_reached_goal: - model.create_backup() - - def log_step(step, step_time, src_loss, dst_loss): - c2s.put({ - 'op': 'tb', - 'action': 'step', - 'step': step, - 'step_time': step_time, - 'src_loss': src_loss, - 'dst_loss': dst_loss - }) - - def log_previews(step, previews, static_previews): - c2s.put({ - 'op': 'tb', - 'action': 'preview', - 'step': step, - 'previews': previews, - 'static_previews': static_previews - }) + model.create_backup() def send_preview(): if not debug: previews = model.get_previews() - c2s.put ( {'op':'show', 'previews': previews, 'iter':model.get_iter(), 'loss_history': model.get_loss_history().copy() } ) + c2s.put({'op': 'show', 'previews': previews, 'iter': model.get_iter(), + 'loss_history': model.get_loss_history().copy()}) else: - previews = [( 'debug, press update for new', model.debug_one_iter())] - c2s.put ( {'op':'show', 'previews': previews} ) - e.set() #Set the GUI Thread as Ready + previews = [('debug, press update for new', model.debug_one_iter())] + c2s.put({'op': 'show', 'previews': previews}) + e.set() # Set the GUI Thread as Ready if model.get_target_iter() != 0: if is_reached_goal: io.log_info('Model already trained to target iteration. You can use preview.') else: - io.log_info('Starting. Target iteration: %d. Press "Enter" to stop training and save model.' % ( model.get_target_iter() ) ) + io.log_info('Starting. Target iteration: %d. Press "Enter" to stop training and save model.' % ( + model.get_target_iter())) else: io.log_info('Starting. Press "Enter" to stop training and save model.') last_save_time = time.time() last_preview_time = time.time() - execute_programs = [ [x[0], x[1], time.time() ] for x in execute_programs ] + execute_programs = [[x[0], x[1], time.time()] for x in execute_programs] - for i in itertools.count(0,1): + for i in itertools.count(0, 1): if not debug: cur_time = time.time() for x in execute_programs: prog_time, prog, last_time = x exec_prog = False - if prog_time > 0 and (cur_time - start_time) >= prog_time: + if 0 < prog_time <= (cur_time - start_time): x[0] = 0 exec_prog = True - elif prog_time < 0 and (cur_time - last_time) >= -prog_time: + elif prog_time < 0 and (cur_time - last_time) >= -prog_time: x[2] = cur_time exec_prog = True @@ -178,18 +175,20 @@ def trainerThread (s2c, c2s, e, try: exec(prog) except Exception as e: - print("Unable to execute program: %s" % (prog) ) + print("Unable to execute program: %s" % prog) if not is_reached_goal: if model.get_iter() == 0: io.log_info("") - io.log_info("Trying to do the first iteration. If an error occurs, reduce the model parameters.") + io.log_info( + "Trying to do the first iteration. If an error occurs, reduce the model parameters.") io.log_info("") - + if sys.platform[0:3] == 'win': io.log_info("!!!") - io.log_info("Windows 10 users IMPORTANT notice. You should set this setting in order to work correctly.") + io.log_info( + "Windows 10 users IMPORTANT notice. You should set this setting in order to work correctly.") io.log_info("https://i.imgur.com/B7cmDCB.jpg") io.log_info("!!!") @@ -198,19 +197,19 @@ def trainerThread (s2c, c2s, e, loss_history = model.get_loss_history() time_str = time.strftime("[%H:%M:%S]") if iter_time >= 10: - loss_string = "{0}[#{1:06d}][{2:.5s}s]".format ( time_str, iter, '{:0.4f}'.format(iter_time) ) + loss_string = "{0}[#{1:06d}][{2:.5s}s]".format(time_str, iter, '{:0.4f}'.format(iter_time)) else: - loss_string = "{0}[#{1:06d}][{2:04d}ms]".format ( time_str, iter, int(iter_time*1000) ) + loss_string = "{0}[#{1:06d}][{2:04d}ms]".format(time_str, iter, int(iter_time * 1000)) if shared_state['after_save']: shared_state['after_save'] = False - - mean_loss = np.mean ( loss_history[save_iter:iter], axis=0) + + mean_loss = np.mean(loss_history[save_iter:iter], axis=0) for loss_value in mean_loss: loss_string += "[%.4f]" % (loss_value) - io.log_info (loss_string) + io.log_info(loss_string) save_iter = iter else: @@ -218,9 +217,12 @@ def trainerThread (s2c, c2s, e, loss_string += "[%.4f]" % (loss_value) if io.is_colab(): - io.log_info ('\r' + loss_string, end='') + io.log_info('\r' + loss_string, end='') else: - io.log_info (loss_string, end='\r') + io.log_info(loss_string, end='\r') + + if socketio is not None: + socketio.emit('loss', loss_string) loss_entry = loss_history[-1] log_step(iter, iter_time, loss_entry[0], loss_entry[1] if len(loss_entry) > 1 else None) @@ -229,10 +231,10 @@ def trainerThread (s2c, c2s, e, model_save() if model.get_target_iter() != 0 and model.is_reached_iter_goal(): - io.log_info ('Reached target iteration.') + io.log_info('Reached target iteration.') model_save() is_reached_goal = True - io.log_info ('You can use preview now.') + io.log_info('You can use preview now.') if not is_reached_goal and (time.time() - last_preview_time) >= tensorboard_preview_interval_min*60: last_preview_time += tensorboard_preview_interval_min*60 @@ -245,7 +247,7 @@ def trainerThread (s2c, c2s, e, model_save() send_preview() - if i==0: + if i == 0: if is_reached_goal: model.pass_one_iter() send_preview() @@ -254,8 +256,8 @@ def trainerThread (s2c, c2s, e, time.sleep(0.005) while not s2c.empty(): - input = s2c.get() - op = input['op'] + item = s2c.get() + op = item['op'] if op == 'save': model_save() elif op == 'backup': @@ -272,12 +274,10 @@ def trainerThread (s2c, c2s, e, if i == -1: break - - model.finalize() except Exception as e: - print ('Error: %s' % (str(e))) + print('Error: %s' % (str(e))) traceback.print_exc() break c2s.put ( {'op':'close'} ) @@ -330,20 +330,209 @@ def handle_tensorboard_op(input): log_tensorboard_previews(step, previews, 'preview', train_summary_writer) if static_previews is not None: log_tensorboard_previews(step, static_previews, 'static_preview', train_summary_writer) + c2s.put({'op': 'close'}) + + +class Zoom(Enum): + ZOOM_25 = (1 / 4, '25%') + ZOOM_33 = (1 / 3, '33%') + ZOOM_50 = (1 / 2, '50%') + ZOOM_67 = (2 / 3, '67%') + ZOOM_75 = (3 / 4, '75%') + ZOOM_80 = (4 / 5, '80%') + ZOOM_90 = (9 / 10, '90%') + ZOOM_100 = (1, '100%') + ZOOM_110 = (11 / 10, '110%') + ZOOM_125 = (5 / 4, '125%') + ZOOM_150 = (3 / 2, '150%') + ZOOM_175 = (7 / 4, '175%') + ZOOM_200 = (2, '200%') + ZOOM_250 = (5 / 2, '250%') + ZOOM_300 = (3, '300%') + ZOOM_400 = (4, '400%') + ZOOM_500 = (5, '500%') + + def __init__(self, scale, label): + self.scale = scale + self.label = label + + def prev(self): + cls = self.__class__ + members = list(cls) + index = members.index(self) - 1 + if index < 0: + return self + return members[index] + + def next(self): + cls = self.__class__ + members = list(cls) + index = members.index(self) + 1 + if index >= len(members): + return self + return members[index] + + +def scale_previews(previews, zoom=Zoom.ZOOM_100): + scaled = [] + for preview in previews: + preview_name, preview_rgb = preview + scale_factor = zoom.scale + if scale_factor < 1: + scaled.append((preview_name, cv2.resize(preview_rgb, (0, 0), + fx=scale_factor, + fy=scale_factor, + interpolation=cv2.INTER_AREA))) + elif scale_factor > 1: + scaled.append((preview_name, cv2.resize(preview_rgb, (0, 0), + fx=scale_factor, + fy=scale_factor, + interpolation=cv2.INTER_LANCZOS4))) + else: + scaled.append((preview_name, preview_rgb)) + return scaled + + +def create_preview_pane_image(previews, selected_preview, loss_history, + show_last_history_iters_count, iteration, batch_size, zoom=Zoom.ZOOM_100): + scaled_previews = scale_previews(previews, zoom) + selected_preview_name = scaled_previews[selected_preview][0] + selected_preview_rgb = scaled_previews[selected_preview][1] + h, w, c = selected_preview_rgb.shape + + # HEAD + head_lines = [ + '[s]:save [enter]:exit [-/+]:zoom: %s' % zoom.label, + '[p]:update [space]:next preview [l]:change history range', + 'Preview: "%s" [%d/%d]' % (selected_preview_name, selected_preview + 1, len(previews)) + ] + head_line_height = int(15 * zoom.scale) + head_height = len(head_lines) * head_line_height + head = np.ones((head_height, w, c)) * 0.1 + + for i in range(0, len(head_lines)): + t = i * head_line_height + b = (i + 1) * head_line_height + head[t:b, 0:w] += imagelib.get_text_image((head_line_height, w, c), head_lines[i], color=[0.8] * c) + + final = head + + if loss_history is not None: + if show_last_history_iters_count == 0: + loss_history_to_show = loss_history + else: + loss_history_to_show = loss_history[-show_last_history_iters_count:] + lh_height = int(100 * zoom.scale) + lh_img = models.ModelBase.get_loss_history_preview(loss_history_to_show, iteration, w, c, lh_height) + final = np.concatenate([final, lh_img], axis=0) + + final = np.concatenate([final, selected_preview_rgb], axis=0) + final = np.clip(final, 0, 1) + return (final * 255).astype(np.uint8) + def main(**kwargs): - io.log_info ("Running trainer.\r\n") + io.log_info("Running trainer.\r\n") no_preview = kwargs.get('no_preview', False) + flask_preview = kwargs.get('flask_preview', False) s2c = queue.Queue() c2s = queue.Queue() e = threading.Event() - thread = threading.Thread(target=trainerThread, args=(s2c, c2s, e), kwargs=kwargs ) - thread.start() - e.wait() #Wait for inital load to occur. + previews = None + loss_history = None + selected_preview = 0 + update_preview = False + is_waiting_preview = False + show_last_history_iters_count = 0 + iteration = 0 + batch_size = 1 + zoom = Zoom.ZOOM_100 + + if flask_preview: + from flaskr.app import create_flask_app + s2flask = queue.Queue() + socketio, flask_app = create_flask_app(s2c, c2s, s2flask, kwargs) + + thread = threading.Thread(target=trainerThread, args=(s2c, c2s, e, socketio), kwargs=kwargs) + thread.start() + + e.wait() # Wait for inital load to occur. + + flask_t = threading.Thread(target=socketio.run, args=(flask_app,), + kwargs={'debug': True, 'use_reloader': False}) + flask_t.start() + + while True: + if not c2s.empty(): + item = c2s.get() + op = item['op'] + if op == 'show': + is_waiting_preview = False + loss_history = item['loss_history'] if 'loss_history' in item.keys() else None + previews = item['previews'] if 'previews' in item.keys() else None + iteration = item['iter'] if 'iter' in item.keys() else 0 + # batch_size = input['batch_size'] if 'iter' in input.keys() else 1 + if previews is not None: + update_preview = True + elif op == 'update': + if not is_waiting_preview: + is_waiting_preview = True + s2c.put({'op': 'preview'}) + elif op == 'next_preview': + selected_preview = (selected_preview + 1) % len(previews) + update_preview = True + elif op == 'change_history_range': + if show_last_history_iters_count == 0: + show_last_history_iters_count = 5000 + elif show_last_history_iters_count == 5000: + show_last_history_iters_count = 10000 + elif show_last_history_iters_count == 10000: + show_last_history_iters_count = 50000 + elif show_last_history_iters_count == 50000: + show_last_history_iters_count = 100000 + elif show_last_history_iters_count == 100000: + show_last_history_iters_count = 0 + update_preview = True + elif op == 'close': + s2c.put({'op': 'close'}) + break + elif op == 'zoom_prev': + zoom = zoom.prev() + update_preview = True + elif op == 'zoom_next': + zoom = zoom.next() + update_preview = True + + if update_preview: + update_preview = False + selected_preview = selected_preview % len(previews) + preview_pane_image = create_preview_pane_image(previews, + selected_preview, + loss_history, + show_last_history_iters_count, + iteration, + batch_size, + zoom) + # io.show_image(wnd_name, preview_pane_image) + model_path = Path(kwargs.get('saved_models_path', '')) + filename = 'preview.png' + preview_file = str(model_path / filename) + cv2.imwrite(preview_file, preview_pane_image) + s2flask.put({'op': 'show'}) + socketio.emit('preview', {'iter': iteration, 'loss': loss_history[-1]}) + try: + io.process_messages(0.01) + except KeyboardInterrupt: + s2c.put({'op': 'close'}) + else: + thread = threading.Thread(target=trainerThread, args=(s2c, c2s, e), kwargs=kwargs) + thread.start() + + e.wait() # Wait for inital load to occur. if no_preview: while True: @@ -357,7 +546,7 @@ def main(**kwargs): try: io.process_messages(0.1) except KeyboardInterrupt: - s2c.put ( {'op': 'close'} ) + s2c.put({'op': 'close'}) else: wnd_name = "Training preview" io.named_window(wnd_name) @@ -373,33 +562,33 @@ def main(**kwargs): iter = 0 while True: if not c2s.empty(): - input = c2s.get() - op = input['op'] + item = c2s.get() + op = item['op'] if op == 'show': is_waiting_preview = False - loss_history = input['loss_history'] if 'loss_history' in input.keys() else None - previews = input['previews'] if 'previews' in input.keys() else None - iter = input['iter'] if 'iter' in input.keys() else 0 + loss_history = item['loss_history'] if 'loss_history' in item.keys() else None + previews = item['previews'] if 'previews' in item.keys() else None + iter = item['iter'] if 'iter' in item.keys() else 0 if previews is not None: max_w = 0 max_h = 0 for (preview_name, preview_rgb) in previews: (h, w, c) = preview_rgb.shape - max_h = max (max_h, h) - max_w = max (max_w, w) + max_h = max(max_h, h) + max_w = max(max_w, w) max_size = 800 if max_h > max_size: - max_w = int( max_w / (max_h / max_size) ) + max_w = int(max_w / (max_h / max_size)) max_h = max_size - #make all previews size equal + # make all previews size equal for preview in previews[:]: (preview_name, preview_rgb) = preview (h, w, c) = preview_rgb.shape if h != max_h or w != max_w: previews.remove(preview) - previews.append ( (preview_name, cv2.resize(preview_rgb, (max_w, max_h))) ) + previews.append((preview_name, cv2.resize(preview_rgb, (max_w, max_h)))) selected_preview = selected_preview % len(previews) update_preview = True elif op == 'tb': @@ -412,22 +601,22 @@ def main(**kwargs): selected_preview_name = previews[selected_preview][0] selected_preview_rgb = previews[selected_preview][1] - (h,w,c) = selected_preview_rgb.shape + (h, w, c) = selected_preview_rgb.shape # HEAD head_lines = [ '[s]:save [b]:backup [enter]:exit', '[p]:update [space]:next preview [l]:change history range', - 'Preview: "%s" [%d/%d]' % (selected_preview_name,selected_preview+1, len(previews) ) - ] + 'Preview: "%s" [%d/%d]' % (selected_preview_name, selected_preview + 1, len(previews)) + ] head_line_height = 15 head_height = len(head_lines) * head_line_height - head = np.ones ( (head_height,w,c) ) * 0.1 + head = np.ones((head_height, w, c)) * 0.1 for i in range(0, len(head_lines)): - t = i*head_line_height - b = (i+1)*head_line_height - head[t:b, 0:w] += imagelib.get_text_image ( (head_line_height,w,c) , head_lines[i], color=[0.8]*c ) + t = i * head_line_height + b = (i + 1) * head_line_height + head[t:b, 0:w] += imagelib.get_text_image((head_line_height, w, c), head_lines[i], color=[0.8] * c) final = head @@ -438,27 +627,28 @@ def main(**kwargs): loss_history_to_show = loss_history[-show_last_history_iters_count:] lh_img = models.ModelBase.get_loss_history_preview(loss_history_to_show, iter, w, c) - final = np.concatenate ( [final, lh_img], axis=0 ) + final = np.concatenate([final, lh_img], axis=0) - final = np.concatenate ( [final, selected_preview_rgb], axis=0 ) + final = np.concatenate([final, selected_preview_rgb], axis=0) final = np.clip(final, 0, 1) - io.show_image( wnd_name, (final*255).astype(np.uint8) ) + io.show_image(wnd_name, (final * 255).astype(np.uint8)) is_showing = True key_events = io.get_key_events(wnd_name) - key, chr_key, ctrl_pressed, alt_pressed, shift_pressed = key_events[-1] if len(key_events) > 0 else (0,0,False,False,False) + key, chr_key, ctrl_pressed, alt_pressed, shift_pressed = key_events[-1] if len(key_events) > 0 else ( + 0, 0, False, False, False) if key == ord('\n') or key == ord('\r'): - s2c.put ( {'op': 'close'} ) + s2c.put({'op': 'close'}) elif key == ord('s'): - s2c.put ( {'op': 'save'} ) + s2c.put({'op': 'save'}) elif key == ord('b'): - s2c.put ( {'op': 'backup'} ) + s2c.put({'op': 'backup'}) elif key == ord('p'): if not is_waiting_preview: is_waiting_preview = True - s2c.put ( {'op': 'preview'} ) + s2c.put({'op': 'preview'}) elif key == ord('l'): if show_last_history_iters_count == 0: show_last_history_iters_count = 5000 @@ -478,6 +668,6 @@ def main(**kwargs): try: io.process_messages(0.1) except KeyboardInterrupt: - s2c.put ( {'op': 'close'} ) + s2c.put({'op': 'close'}) - io.destroy_all_windows() \ No newline at end of file + io.destroy_all_windows() diff --git a/mainscripts/XSegUtil.py b/mainscripts/XSegUtil.py index 9f9f893..c75a14a 100644 --- a/mainscripts/XSegUtil.py +++ b/mainscripts/XSegUtil.py @@ -10,8 +10,8 @@ from core.cv2ex import * from core.interact import interact as io from core.leras import nn from DFLIMG import * -from facelib import XSegNet - +from facelib import XSegNet, LandmarksProcessor, FaceType +import pickle def apply_xseg(input_path, model_path): if not input_path.exists(): @@ -20,17 +20,42 @@ def apply_xseg(input_path, model_path): if not model_path.exists(): raise ValueError(f'{model_path} not found. Please ensure it exists.') + face_type = None + + model_dat = model_path / 'XSeg_data.dat' + if model_dat.exists(): + dat = pickle.loads( model_dat.read_bytes() ) + dat_options = dat.get('options', None) + if dat_options is not None: + face_type = dat_options.get('face_type', None) + + + + if face_type is None: + face_type = io.input_str ("XSeg model face type", 'same', ['h','mf','f','wf','head','same'], help_message="Specify face type of trained XSeg model. For example if XSeg model trained as WF, but faceset is HEAD, specify WF to apply xseg only on WF part of HEAD. Default is 'same'").lower() + if face_type == 'same': + face_type = None + + if face_type is not None: + face_type = {'h' : FaceType.HALF, + 'mf' : FaceType.MID_FULL, + 'f' : FaceType.FULL, + 'wf' : FaceType.WHOLE_FACE, + 'head' : FaceType.HEAD}[face_type] + io.log_info(f'Applying trained XSeg model to {input_path.name}/ folder.') device_config = nn.DeviceConfig.ask_choose_device(choose_only_one=True) nn.initialize(device_config) + + xseg = XSegNet(name='XSeg', load_weights=True, weights_file_root=model_path, data_format=nn.data_format, raise_on_no_model_files=True) - res = xseg.get_resolution() + xseg_res = xseg.get_resolution() images_paths = pathex.get_image_paths(input_path, return_Path_class=True) @@ -42,15 +67,36 @@ def apply_xseg(input_path, model_path): img = cv2_imread(filepath).astype(np.float32) / 255.0 h,w,c = img.shape - if w != res: - img = cv2.resize( img, (res,res), interpolation=cv2.INTER_CUBIC ) - if len(img.shape) == 2: - img = img[...,None] - - mask = xseg.extract(img) - mask[mask < 0.5]=0 - mask[mask >= 0.5]=1 + img_face_type = FaceType.fromString( dflimg.get_face_type() ) + if face_type is not None and img_face_type != face_type: + lmrks = dflimg.get_source_landmarks() + + fmat = LandmarksProcessor.get_transform_mat(lmrks, w, face_type) + imat = LandmarksProcessor.get_transform_mat(lmrks, w, img_face_type) + + g_p = LandmarksProcessor.transform_points (np.float32([(0,0),(w,0),(0,w) ]), fmat, True) + g_p2 = LandmarksProcessor.transform_points (g_p, imat) + + mat = cv2.getAffineTransform( g_p2, np.float32([(0,0),(w,0),(0,w) ]) ) + + img = cv2.warpAffine(img, mat, (w, w), cv2.INTER_LANCZOS4) + img = cv2.resize(img, (xseg_res, xseg_res), interpolation=cv2.INTER_LANCZOS4) + else: + if w != xseg_res: + img = cv2.resize( img, (xseg_res,xseg_res), interpolation=cv2.INTER_LANCZOS4 ) + + if len(img.shape) == 2: + img = img[...,None] + + mask = xseg.extract(img) + + if face_type is not None and img_face_type != face_type: + mask = cv2.resize(mask, (w, w), interpolation=cv2.INTER_LANCZOS4) + mask = cv2.warpAffine( mask, mat, (w,w), np.zeros( (h,w,c), dtype=np.float), cv2.WARP_INVERSE_MAP | cv2.INTER_LANCZOS4) + mask = cv2.resize(mask, (xseg_res, xseg_res), interpolation=cv2.INTER_LANCZOS4) + mask[mask < 0.5]=0 + mask[mask >= 0.5]=1 dflimg.set_xseg_mask(mask) dflimg.save() @@ -67,7 +113,8 @@ def fetch_xseg(input_path): images_paths = pathex.get_image_paths(input_path, return_Path_class=True) - files_copied = 0 + + files_copied = [] for filepath in io.progress_bar_generator(images_paths, "Processing"): dflimg = DFLIMG.load(filepath) if dflimg is None or not dflimg.has_data(): @@ -77,10 +124,16 @@ def fetch_xseg(input_path): ie_polys = dflimg.get_seg_ie_polys() if ie_polys.has_polys(): - files_copied += 1 + files_copied.append(filepath) shutil.copy ( str(filepath), str(output_path / filepath.name) ) - io.log_info(f'Files copied: {files_copied}') + io.log_info(f'Files copied: {len(files_copied)}') + + is_delete = io.input_bool (f"\r\nDelete original files?", True) + if is_delete: + for filepath in files_copied: + Path(filepath).unlink() + def remove_xseg(input_path): if not input_path.exists(): diff --git a/merger/MergeMasked.py b/merger/MergeMasked.py index 1d3de4a..7d2caa4 100644 --- a/merger/MergeMasked.py +++ b/merger/MergeMasked.py @@ -57,7 +57,9 @@ def MergeMaskedFace (predictor_func, predictor_input_shape, prd_face_mask_a_0 = cv2.resize (prd_face_mask_a_0, (output_size, output_size), interpolation=cv2.INTER_CUBIC) prd_face_dst_mask_a_0 = cv2.resize (prd_face_dst_mask_a_0, (output_size, output_size), interpolation=cv2.INTER_CUBIC) - if cfg.mask_mode == 1: #dst + if cfg.mask_mode == 0: #full + wrk_face_mask_a_0 = np.ones_like(dst_face_mask_a_0) + elif cfg.mask_mode == 1: #dst wrk_face_mask_a_0 = cv2.resize (dst_face_mask_a_0, (output_size,output_size), interpolation=cv2.INTER_CUBIC) elif cfg.mask_mode == 2: #learned-prd wrk_face_mask_a_0 = prd_face_mask_a_0 @@ -142,7 +144,9 @@ def MergeMaskedFace (predictor_func, predictor_input_shape, elif 'raw' in cfg.mode: if cfg.mode == 'raw-rgb': - out_img = cv2.warpAffine( prd_face_bgr, face_output_mat, img_size, img_bgr.copy(), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC ) + out_img_face = cv2.warpAffine( prd_face_bgr, face_output_mat, img_size, np.empty_like(img_bgr), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC) + out_img_face_mask = cv2.warpAffine( np.ones_like(prd_face_bgr), face_output_mat, img_size, np.empty_like(img_bgr), cv2.WARP_INVERSE_MAP | cv2.INTER_CUBIC) + out_img = img_bgr*(1-out_img_face_mask) + out_img_face*out_img_face_mask out_merging_mask_a = img_face_mask_a elif cfg.mode == 'raw-predict': out_img = prd_face_bgr diff --git a/merger/MergerConfig.py b/merger/MergerConfig.py index 432bdf1..eba1493 100644 --- a/merger/MergerConfig.py +++ b/merger/MergerConfig.py @@ -81,7 +81,8 @@ mode_dict = {0:'original', mode_str_dict = { mode_dict[key] : key for key in mode_dict.keys() } -mask_mode_dict = {1:'dst', +mask_mode_dict = {0:'full', + 1:'dst', 2:'learned-prd', 3:'learned-dst', 4:'learned-prd*learned-dst', diff --git a/models/ModelBase.py b/models/ModelBase.py index 3cb88c5..a547173 100644 --- a/models/ModelBase.py +++ b/models/ModelBase.py @@ -8,6 +8,7 @@ import pickle import shutil import tempfile import time +import datetime from pathlib import Path import cv2 @@ -180,12 +181,15 @@ class ModelBase(object): if self.is_first_run(): # save as default options only for first run model initialize self.default_options_path.write_bytes( pickle.dumps (self.options) ) - + self.session_name = self.options.get('session_name', "") self.autobackup_hour = self.options.get('autobackup_hour', 0) + self.maximum_n_backups = self.options.get('maximum_n_backups', 24) self.write_preview_history = self.options.get('write_preview_history', False) self.target_iter = self.options.get('target_iter',0) self.random_flip = self.options.get('random_flip',True) - + self.random_src_flip = self.options.get('random_src_flip', False) + self.random_dst_flip = self.options.get('random_dst_flip', True) + self.on_initialize() self.options['batch_size'] = self.batch_size @@ -276,13 +280,21 @@ class ModelBase(object): def ask_override(self): return self.is_training and self.iter != 0 and io.input_in_time ("Press enter in 2 seconds to override model settings.", 5 if io.is_colab() else 2 ) + def ask_session_name(self, default_value=""): + default_session_name = self.options['session_name'] = self.load_or_def_option('session_name', default_value) + self.options['session_name'] = io.input_str("Session name", default_session_name, help_message="String to refer back to in summary.txt and in autobackup foldername") + def ask_autobackup_hour(self, default_value=0): default_autobackup_hour = self.options['autobackup_hour'] = self.load_or_def_option('autobackup_hour', default_value) - self.options['autobackup_hour'] = io.input_int(f"Autobackup every N hour", default_autobackup_hour, add_info="0..24", help_message="Autobackup model files with preview every N hour. Latest backup located in model/<>_autobackups/01") + self.options['autobackup_hour'] = io.input_int(f"Autobackup every N hour", default_autobackup_hour, add_info="0..24", help_message="Autobackup model files with preview every N hour. Latest backup is the last folder when sorted by name ascending located in model/<>_autobackups") + + def ask_maximum_n_backups(self, default_value=24): + default_maximum_n_backups = self.options['maximum_n_backups'] = self.load_or_def_option('maximum_n_backups', default_value) + self.options['maximum_n_backups'] = io.input_int(f"Maximum N backups", default_maximum_n_backups, help_message="Maximum amount of backups that are located in model/<>_autobackups. Inputting 0 here would allow it to autobackup as many times as it occurs.") def ask_write_preview_history(self, default_value=False): default_write_preview_history = self.load_or_def_option('write_preview_history', default_value) - self.options['write_preview_history'] = io.input_bool(f"Write preview history", default_write_preview_history, help_message="Preview history will be writed to _history folder.") + self.options['write_preview_history'] = io.input_bool(f"Write preview history", default_write_preview_history, help_message="Preview history will be written to _history folder.") if self.options['write_preview_history']: if io.is_support_windows(): @@ -297,6 +309,14 @@ class ModelBase(object): def ask_random_flip(self): default_random_flip = self.load_or_def_option('random_flip', True) self.options['random_flip'] = io.input_bool("Flip faces randomly", default_random_flip, help_message="Predicted face will look more naturally without this option, but src faceset should cover all face directions as dst faceset.") + + def ask_random_src_flip(self): + default_random_src_flip = self.load_or_def_option('random_src_flip', False) + self.options['random_src_flip'] = io.input_bool("Flip SRC faces randomly", default_random_src_flip, help_message="Random horizontal flip SRC faceset. Covers more angles, but the face may look less naturally.") + + def ask_random_dst_flip(self): + default_random_dst_flip = self.load_or_def_option('random_dst_flip', True) + self.options['random_dst_flip'] = io.input_bool("Flip DST faces randomly", default_random_dst_flip, help_message="Random horizontal flip DST faceset. Makes generalization of src->dst better, if src random flip is not enabled.") def ask_batch_size(self, suggest_batch_size=None, range=None): default_batch_size = self.load_or_def_option('batch_size', suggest_batch_size or self.batch_size) @@ -405,33 +425,32 @@ class ModelBase(object): bckp_filename_list = [ self.get_strpath_storage_for_file(filename) for _, filename in self.get_model_filename_list() ] bckp_filename_list += [ str(self.get_summary_path()), str(self.model_data_path) ] - for i in range(24,0,-1): - idx_str = '%.2d' % i - next_idx_str = '%.2d' % (i+1) + # Create new backup + session_suffix = f'_{self.session_name}' if self.session_name else '' + idx_str = datetime.datetime.now().strftime('%Y%m%dT%H%M%S') + session_suffix + idx_backup_path = self.autobackups_path / idx_str + idx_backup_path.mkdir() + for filename in bckp_filename_list: + shutil.copy(str(filename), str(idx_backup_path / Path(filename).name))\ - idx_backup_path = self.autobackups_path / idx_str - next_idx_packup_path = self.autobackups_path / next_idx_str + previews = self.get_previews() - if idx_backup_path.exists(): - if i == 24: - pathex.delete_all_files(idx_backup_path) - else: - next_idx_packup_path.mkdir(exist_ok=True) - pathex.move_all_files (idx_backup_path, next_idx_packup_path) + # Generate previews and save in new backup + plist = [] + for i in range(len(previews)): + name, bgr = previews[i] + plist += [ (bgr, idx_backup_path / ( ('preview_%s.jpg') % (name)) ) ] - if i == 1: - idx_backup_path.mkdir(exist_ok=True) - for filename in bckp_filename_list: - shutil.copy ( str(filename), str(idx_backup_path / Path(filename).name) ) + if len(plist) != 0: + self.get_preview_history_writer().post(plist, self.loss_history, self.iter) - previews = self.get_previews() - plist = [] - for i in range(len(previews)): - name, bgr = previews[i] - plist += [ (bgr, idx_backup_path / ( ('preview_%s.jpg') % (name)) ) ] - - if len(plist) != 0: - self.get_preview_history_writer().post(plist, self.loss_history, self.iter) + # Check if we've exceeded the max number of backups + if self.maximum_n_backups != 0: + all_backups = sorted([x for x in self.autobackups_path.iterdir() if x.is_dir()]) + while len(all_backups) > self.maximum_n_backups: + oldest_backup = all_backups.pop(0) + pathex.delete_all_files(oldest_backup) + oldest_backup.rmdir() def debug_one_iter(self): images = [] @@ -535,7 +554,7 @@ class ModelBase(object): def get_summary_text(self): visible_options = self.options.copy() visible_options.update(self.options_show_override) - + ###Generate text summary of model hyperparameters #Find the longest key name and value string. Used as column widths. width_name = max([len(k) for k in visible_options.keys()] + [17]) + 1 # Single space buffer to left edge. Minimum of 17, the length of the longest static string used "Current iteration" @@ -574,10 +593,9 @@ class ModelBase(object): return summary_text @staticmethod - def get_loss_history_preview(loss_history, iter, w, c): + def get_loss_history_preview(loss_history, iter, w, c, lh_height=100): loss_history = np.array (loss_history.copy()) - lh_height = 100 lh_img = np.ones ( (lh_height,w,c) ) * 0.1 if len(loss_history) != 0: diff --git a/models/Model_AMP/Model.py b/models/Model_AMP/Model.py new file mode 100644 index 0000000..edbe795 --- /dev/null +++ b/models/Model_AMP/Model.py @@ -0,0 +1,892 @@ +import multiprocessing +import operator +from functools import partial + +import numpy as np + +from core import mathlib +from core.interact import interact as io +from core.leras import nn +from facelib import FaceType +from models import ModelBase +from samplelib import * +from core.cv2ex import * + +class AMPModel(ModelBase): + + #override + def on_initialize_options(self): + device_config = nn.getCurrentDeviceConfig() + + lowest_vram = 2 + if len(device_config.devices) != 0: + lowest_vram = device_config.devices.get_worst_device().total_mem_gb + + if lowest_vram >= 4: + suggest_batch_size = 8 + else: + suggest_batch_size = 4 + + yn_str = {True:'y',False:'n'} + min_res = 64 + max_res = 640 + + default_resolution = self.options['resolution'] = self.load_or_def_option('resolution', 224) + default_face_type = self.options['face_type'] = self.load_or_def_option('face_type', 'wf') + default_models_opt_on_gpu = self.options['models_opt_on_gpu'] = self.load_or_def_option('models_opt_on_gpu', True) + + default_ae_dims = self.options['ae_dims'] = self.load_or_def_option('ae_dims', 256) + default_e_dims = self.options['e_dims'] = self.load_or_def_option('e_dims', 64) + default_d_dims = self.options['d_dims'] = self.options.get('d_dims', None) + default_d_mask_dims = self.options['d_mask_dims'] = self.options.get('d_mask_dims', None) + default_morph_factor = self.options['morph_factor'] = self.options.get('morph_factor', 0.33) + default_masked_training = self.options['masked_training'] = self.load_or_def_option('masked_training', True) + default_eyes_mouth_prio = self.options['eyes_mouth_prio'] = self.load_or_def_option('eyes_mouth_prio', True) + default_uniform_yaw = self.options['uniform_yaw'] = self.load_or_def_option('uniform_yaw', False) + + lr_dropout = self.load_or_def_option('lr_dropout', 'n') + lr_dropout = {True:'y', False:'n'}.get(lr_dropout, lr_dropout) #backward comp + default_lr_dropout = self.options['lr_dropout'] = lr_dropout + + default_loss_function = self.options['loss_function'] = self.load_or_def_option('loss_function', 'SSIM') + + default_random_warp = self.options['random_warp'] = self.load_or_def_option('random_warp', True) + default_random_downsample = self.options['random_downsample'] = self.load_or_def_option('random_downsample', False) + default_random_noise = self.options['random_noise'] = self.load_or_def_option('random_noise', False) + default_random_blur = self.options['random_blur'] = self.load_or_def_option('random_blur', False) + default_random_jpeg = self.options['random_jpeg'] = self.load_or_def_option('random_jpeg', False) + + default_background_power = self.options['background_power'] = self.load_or_def_option('background_power', 0.0) + default_ct_mode = self.options['ct_mode'] = self.load_or_def_option('ct_mode', 'none') + default_random_color = self.options['random_color'] = self.load_or_def_option('random_color', False) + default_clipgrad = self.options['clipgrad'] = self.load_or_def_option('clipgrad', False) + default_pretrain = self.options['pretrain'] = self.load_or_def_option('pretrain', False) + + + ask_override = self.ask_override() + if self.is_first_run() or ask_override: + self.ask_autobackup_hour() + self.ask_write_preview_history() + self.ask_target_iter() + self.ask_random_src_flip() + self.ask_random_dst_flip() + self.ask_batch_size(suggest_batch_size) + + if self.is_first_run(): + resolution = io.input_int("Resolution", default_resolution, add_info="64-640", help_message="More resolution requires more VRAM and time to train. Value will be adjusted to multiple of 32 .") + resolution = np.clip ( (resolution // 32) * 32, min_res, max_res) + self.options['resolution'] = resolution + self.options['face_type'] = io.input_str ("Face type", default_face_type, ['wf','head'], help_message="whole face / head").lower() + + + default_d_dims = self.options['d_dims'] = self.load_or_def_option('d_dims', 64) + + default_d_mask_dims = default_d_dims // 3 + default_d_mask_dims += default_d_mask_dims % 2 + default_d_mask_dims = self.options['d_mask_dims'] = self.load_or_def_option('d_mask_dims', default_d_mask_dims) + + if self.is_first_run(): + self.options['ae_dims'] = np.clip ( io.input_int("AutoEncoder dimensions", default_ae_dims, add_info="32-1024", help_message="All face information will packed to AE dims. If amount of AE dims are not enough, then for example closed eyes will not be recognized. More dims are better, but require more VRAM. You can fine-tune model size to fit your GPU." ), 32, 1024 ) + + e_dims = np.clip ( io.input_int("Encoder dimensions", default_e_dims, add_info="16-256", help_message="More dims help to recognize more facial features and achieve sharper result, but require more VRAM. You can fine-tune model size to fit your GPU." ), 16, 256 ) + self.options['e_dims'] = e_dims + e_dims % 2 + + d_dims = np.clip ( io.input_int("Decoder dimensions", default_d_dims, add_info="16-256", help_message="More dims help to recognize more facial features and achieve sharper result, but require more VRAM. You can fine-tune model size to fit your GPU." ), 16, 256 ) + self.options['d_dims'] = d_dims + d_dims % 2 + + d_mask_dims = np.clip ( io.input_int("Decoder mask dimensions", default_d_mask_dims, add_info="16-256", help_message="Typical mask dimensions = decoder dimensions / 3. If you manually cut out obstacles from the dst mask, you can increase this parameter to achieve better quality." ), 16, 256 ) + self.options['d_mask_dims'] = d_mask_dims + d_mask_dims % 2 + + if self.is_first_run() or ask_override: + morph_factor = np.clip ( io.input_number ("Morph factor.", default_morph_factor, add_info="0.1 .. 0.5", help_message="The smaller the value, the more src-like facial expressions will appear. The larger the value, the less space there is to train a large dst faceset in the neural network. Typical fine value is 0.33"), 0.1, 0.5 ) + self.options['morph_factor'] = morph_factor + + if self.options['face_type'] == 'wf' or self.options['face_type'] == 'head': + self.options['masked_training'] = io.input_bool ("Masked training", default_masked_training, help_message="This option is available only for 'whole_face' or 'head' type. Masked training clips training area to full_face mask or XSeg mask, thus network will train the faces properly.") + + self.options['eyes_mouth_prio'] = io.input_bool ("Eyes and mouth priority", default_eyes_mouth_prio, help_message='Helps to fix eye problems during training like "alien eyes" and wrong eyes direction. Also makes the detail of the teeth higher.') + self.options['uniform_yaw'] = io.input_bool ("Uniform yaw distribution of samples", default_uniform_yaw, help_message='Helps to fix blurry side faces due to small amount of them in the faceset.') + + default_gan_power = self.options['gan_power'] = self.load_or_def_option('gan_power', 0.0) + default_gan_patch_size = self.options['gan_patch_size'] = self.load_or_def_option('gan_patch_size', self.options['resolution'] // 8) + default_gan_dims = self.options['gan_dims'] = self.load_or_def_option('gan_dims', 16) + + if self.is_first_run() or ask_override: + self.options['models_opt_on_gpu'] = io.input_bool ("Place models and optimizer on GPU", default_models_opt_on_gpu, help_message="When you train on one GPU, by default model and optimizer weights are placed on GPU to accelerate the process. You can place they on CPU to free up extra VRAM, thus set bigger dimensions.") + + self.options['lr_dropout'] = io.input_str (f"Use learning rate dropout", default_lr_dropout, ['n','y','cpu'], help_message="When the face is trained enough, you can enable this option to get extra sharpness and reduce subpixel shake for less amount of iterations. Enabled it before `disable random warp` and before GAN. \nn - disabled.\ny - enabled\ncpu - enabled on CPU. This allows not to use extra VRAM, sacrificing 20% time of iteration.") + + self.options['loss_function'] = io.input_str(f"Loss function", default_loss_function, ['SSIM', 'MS-SSIM', 'MS-SSIM+L1'], + help_message="Change loss function used for image quality assessment.") + + self.options['random_warp'] = io.input_bool ("Enable random warp of samples", default_random_warp, help_message="Random warp is required to generalize facial expressions of both faces. When the face is trained enough, you can disable it to get extra sharpness and reduce subpixel shake for less amount of iterations.") + + self.options['random_downsample'] = io.input_bool("Enable random downsample of samples", default_random_downsample, help_message="") + self.options['random_noise'] = io.input_bool("Enable random noise added to samples", default_random_noise, help_message="") + self.options['random_blur'] = io.input_bool("Enable random blur of samples", default_random_blur, help_message="") + self.options['random_jpeg'] = io.input_bool("Enable random jpeg compression of samples", default_random_jpeg, help_message="") + + self.options['gan_power'] = np.clip ( io.input_number ("GAN power", default_gan_power, add_info="0.0 .. 1.0", help_message="Forces the neural network to learn small details of the face. Enable it only when the face is trained enough with lr_dropout(on) and random_warp(off), and don't disable. The higher the value, the higher the chances of artifacts. Typical fine value is 0.1"), 0.0, 1.0 ) + + if self.options['gan_power'] != 0.0: + gan_patch_size = np.clip ( io.input_int("GAN patch size", default_gan_patch_size, add_info="3-640", help_message="The higher patch size, the higher the quality, the more VRAM is required. You can get sharper edges even at the lowest setting. Typical fine value is resolution / 8." ), 3, 640 ) + self.options['gan_patch_size'] = gan_patch_size + + gan_dims = np.clip ( io.input_int("GAN dimensions", default_gan_dims, add_info="4-64", help_message="The dimensions of the GAN network. The higher dimensions, the more VRAM is required. You can get sharper edges even at the lowest setting. Typical fine value is 16." ), 4, 64 ) + self.options['gan_dims'] = gan_dims + + self.options['background_power'] = np.clip ( io.input_number("Background power", default_background_power, add_info="0.0..1.0", help_message="Learn the area outside of the mask. Helps smooth out area near the mask boundaries. Can be used at any time"), 0.0, 1.0 ) + + self.options['ct_mode'] = io.input_str (f"Color transfer for src faceset", default_ct_mode, ['none','rct','lct','mkl','idt','sot', 'fs-aug'], help_message="Change color distribution of src samples close to dst samples. Try all modes to find the best.") + self.options['random_color'] = io.input_bool ("Random color", default_random_color, help_message="Samples are randomly rotated around the L axis in LAB colorspace, helps generalize training") + self.options['clipgrad'] = io.input_bool ("Enable gradient clipping", default_clipgrad, help_message="Gradient clipping reduces chance of model collapse, sacrificing speed of training.") + + self.options['pretrain'] = io.input_bool ("Enable pretraining mode", default_pretrain, help_message="Pretrain the model with large amount of various faces. After that, model can be used to train the fakes more quickly. Forces random_warp=N, random_flips=Y, gan_power=0.0, lr_dropout=N, uniform_yaw=Y") + + self.gan_model_changed = (default_gan_patch_size != self.options['gan_patch_size']) or (default_gan_dims != self.options['gan_dims']) + self.pretrain_just_disabled = (default_pretrain == True and self.options['pretrain'] == False) + + #override + def on_initialize(self): + device_config = nn.getCurrentDeviceConfig() + devices = device_config.devices + self.model_data_format = "NCHW" + nn.initialize(data_format=self.model_data_format) + tf = nn.tf + + self.resolution = resolution = self.options['resolution'] + + lowest_dense_res = self.lowest_dense_res = resolution // 32 + + class Downscale(nn.ModelBase): + def __init__(self, in_ch, out_ch, kernel_size=5, *kwargs ): + self.in_ch = in_ch + self.out_ch = out_ch + self.kernel_size = kernel_size + super().__init__(*kwargs) + + def on_build(self, *args, **kwargs ): + self.conv1 = nn.Conv2D( self.in_ch, self.out_ch, kernel_size=self.kernel_size, strides=2, padding='SAME') + + def forward(self, x): + x = self.conv1(x) + x = tf.nn.leaky_relu(x, 0.1) + return x + + def get_out_ch(self): + return self.out_ch + + class Upscale(nn.ModelBase): + def on_build(self, in_ch, out_ch, kernel_size=3 ): + self.conv1 = nn.Conv2D( in_ch, out_ch*4, kernel_size=kernel_size, padding='SAME') + + def forward(self, x): + x = self.conv1(x) + x = tf.nn.leaky_relu(x, 0.1) + x = nn.depth_to_space(x, 2) + return x + + class ResidualBlock(nn.ModelBase): + def on_build(self, ch, kernel_size=3 ): + self.conv1 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME') + self.conv2 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME') + + def forward(self, inp): + x = self.conv1(inp) + x = tf.nn.leaky_relu(x, 0.2) + x = self.conv2(x) + x = tf.nn.leaky_relu(inp+x, 0.2) + return x + + class Encoder(nn.ModelBase): + def on_build(self, in_ch, e_ch, ae_ch): + self.down1 = Downscale(in_ch, e_ch, kernel_size=5) + self.res1 = ResidualBlock(e_ch) + self.down2 = Downscale(e_ch, e_ch*2, kernel_size=5) + self.down3 = Downscale(e_ch*2, e_ch*4, kernel_size=5) + self.down4 = Downscale(e_ch*4, e_ch*8, kernel_size=5) + self.down5 = Downscale(e_ch*8, e_ch*8, kernel_size=5) + self.res5 = ResidualBlock(e_ch*8) + self.dense1 = nn.Dense( lowest_dense_res*lowest_dense_res*e_ch*8, ae_ch ) + + def forward(self, inp): + x = inp + x = self.down1(x) + x = self.res1(x) + x = self.down2(x) + x = self.down3(x) + x = self.down4(x) + x = self.down5(x) + x = self.res5(x) + x = nn.flatten(x) + x = nn.pixel_norm(x, axes=-1) + x = self.dense1(x) + return x + + + class Inter(nn.ModelBase): + def __init__(self, ae_ch, ae_out_ch, **kwargs): + self.ae_ch, self.ae_out_ch = ae_ch, ae_out_ch + super().__init__(**kwargs) + + def on_build(self): + ae_ch, ae_out_ch = self.ae_ch, self.ae_out_ch + self.dense2 = nn.Dense( ae_ch, lowest_dense_res * lowest_dense_res * ae_out_ch ) + + def forward(self, inp): + x = inp + x = self.dense2(x) + x = nn.reshape_4D (x, lowest_dense_res, lowest_dense_res, self.ae_out_ch) + return x + + def get_out_ch(self): + return self.ae_out_ch + + class Decoder(nn.ModelBase): + def on_build(self, in_ch, d_ch, d_mask_ch ): + self.upscale0 = Upscale(in_ch, d_ch*8, kernel_size=3) + self.upscale1 = Upscale(d_ch*8, d_ch*8, kernel_size=3) + self.upscale2 = Upscale(d_ch*8, d_ch*4, kernel_size=3) + self.upscale3 = Upscale(d_ch*4, d_ch*2, kernel_size=3) + + self.res0 = ResidualBlock(d_ch*8, kernel_size=3) + self.res1 = ResidualBlock(d_ch*8, kernel_size=3) + self.res2 = ResidualBlock(d_ch*4, kernel_size=3) + self.res3 = ResidualBlock(d_ch*2, kernel_size=3) + + self.upscalem0 = Upscale(in_ch, d_mask_ch*8, kernel_size=3) + self.upscalem1 = Upscale(d_mask_ch*8, d_mask_ch*8, kernel_size=3) + self.upscalem2 = Upscale(d_mask_ch*8, d_mask_ch*4, kernel_size=3) + self.upscalem3 = Upscale(d_mask_ch*4, d_mask_ch*2, kernel_size=3) + self.upscalem4 = Upscale(d_mask_ch*2, d_mask_ch*1, kernel_size=3) + self.out_convm = nn.Conv2D( d_mask_ch*1, 1, kernel_size=1, padding='SAME') + + self.out_conv = nn.Conv2D( d_ch*2, 3, kernel_size=1, padding='SAME') + self.out_conv1 = nn.Conv2D( d_ch*2, 3, kernel_size=3, padding='SAME') + self.out_conv2 = nn.Conv2D( d_ch*2, 3, kernel_size=3, padding='SAME') + self.out_conv3 = nn.Conv2D( d_ch*2, 3, kernel_size=3, padding='SAME') + + def forward(self, inp): + z = inp + + x = self.upscale0(z) + x = self.res0(x) + x = self.upscale1(x) + x = self.res1(x) + x = self.upscale2(x) + x = self.res2(x) + x = self.upscale3(x) + x = self.res3(x) + + x = tf.nn.sigmoid( nn.depth_to_space(tf.concat( (self.out_conv(x), + self.out_conv1(x), + self.out_conv2(x), + self.out_conv3(x)), nn.conv2d_ch_axis), 2) ) + + m = self.upscalem0(z) + m = self.upscalem1(m) + m = self.upscalem2(m) + m = self.upscalem3(m) + m = self.upscalem4(m) + m = tf.nn.sigmoid(self.out_convm(m)) + return x, m + + self.face_type = {'wf' : FaceType.WHOLE_FACE, + 'head' : FaceType.HEAD}[ self.options['face_type'] ] + + if 'eyes_prio' in self.options: + self.options.pop('eyes_prio') + + eyes_mouth_prio = self.options['eyes_mouth_prio'] + + ae_dims = self.ae_dims = self.options['ae_dims'] + e_dims = self.options['e_dims'] + d_dims = self.options['d_dims'] + d_mask_dims = self.options['d_mask_dims'] + morph_factor = self.options['morph_factor'] + + pretrain = self.pretrain = self.options['pretrain'] + if self.pretrain_just_disabled: + self.set_iter(0) + + self.gan_power = gan_power = 0.0 if self.pretrain else self.options['gan_power'] + random_warp = False if self.pretrain else self.options['random_warp'] + random_src_flip = self.random_src_flip if not self.pretrain else True + random_dst_flip = self.random_dst_flip if not self.pretrain else True + + if self.pretrain: + self.options_show_override['gan_power'] = 0.0 + self.options_show_override['random_warp'] = False + self.options_show_override['lr_dropout'] = 'n' + self.options_show_override['uniform_yaw'] = True + + masked_training = self.options['masked_training'] + ct_mode = self.options['ct_mode'] + if ct_mode == 'none': + ct_mode = None + + models_opt_on_gpu = False if len(devices) == 0 else self.options['models_opt_on_gpu'] + models_opt_device = nn.tf_default_device_name if models_opt_on_gpu and self.is_training else '/CPU:0' + optimizer_vars_on_cpu = models_opt_device=='/CPU:0' + + input_ch=3 + bgr_shape = self.bgr_shape = nn.get4Dshape(resolution,resolution,input_ch) + mask_shape = nn.get4Dshape(resolution,resolution,1) + self.model_filename_list = [] + + with tf.device ('/CPU:0'): + #Place holders on CPU + self.warped_src = tf.placeholder (nn.floatx, bgr_shape, name='warped_src') + self.warped_dst = tf.placeholder (nn.floatx, bgr_shape, name='warped_dst') + + self.target_src = tf.placeholder (nn.floatx, bgr_shape, name='target_src') + self.target_dst = tf.placeholder (nn.floatx, bgr_shape, name='target_dst') + + self.target_srcm = tf.placeholder (nn.floatx, mask_shape, name='target_srcm') + self.target_srcm_em = tf.placeholder (nn.floatx, mask_shape, name='target_srcm_em') + self.target_dstm = tf.placeholder (nn.floatx, mask_shape, name='target_dstm') + self.target_dstm_em = tf.placeholder (nn.floatx, mask_shape, name='target_dstm_em') + + self.morph_value_t = tf.placeholder (nn.floatx, (1,), name='morph_value_t') + + # Initializing model classes + + with tf.device (models_opt_device): + self.encoder = Encoder(in_ch=input_ch, e_ch=e_dims, ae_ch=ae_dims, name='encoder') + self.inter_src = Inter(ae_ch=ae_dims, ae_out_ch=ae_dims, name='inter_src') + self.inter_dst = Inter(ae_ch=ae_dims, ae_out_ch=ae_dims, name='inter_dst') + self.decoder = Decoder(in_ch=ae_dims, d_ch=d_dims, d_mask_ch=d_mask_dims, name='decoder') + + self.model_filename_list += [ [self.encoder, 'encoder.npy'], + [self.inter_src, 'inter_src.npy'], + [self.inter_dst , 'inter_dst.npy'], + [self.decoder , 'decoder.npy'] ] + + if self.is_training: + if gan_power != 0: + self.GAN = nn.UNetPatchDiscriminator(patch_size=self.options['gan_patch_size'], in_ch=input_ch, base_ch=self.options['gan_dims'], name="GAN") + self.model_filename_list += [ [self.GAN, 'GAN.npy'] ] + + # Initialize optimizers + lr=5e-5 + lr_dropout = 0.3 if self.options['lr_dropout'] in ['y','cpu'] and not self.pretrain else 1.0 + + clipnorm = 1.0 if self.options['clipgrad'] else 0.0 + + self.all_weights = self.encoder.get_weights() + self.inter_src.get_weights() + self.inter_dst.get_weights() + self.decoder.get_weights() + if pretrain: + self.trainable_weights = self.encoder.get_weights() + self.inter_dst.get_weights() + self.decoder.get_weights() + else: + self.trainable_weights = self.encoder.get_weights() + self.inter_src.get_weights() + self.inter_dst.get_weights() + self.decoder.get_weights() + + self.src_dst_opt = nn.AdaBelief(lr=lr, lr_dropout=lr_dropout, clipnorm=clipnorm, name='src_dst_opt') + self.src_dst_opt.initialize_variables (self.all_weights, vars_on_cpu=optimizer_vars_on_cpu, lr_dropout_on_cpu=self.options['lr_dropout']=='cpu') + self.model_filename_list += [ (self.src_dst_opt, 'src_dst_opt.npy') ] + + if gan_power != 0: + self.GAN_opt = nn.AdaBelief(lr=lr, lr_dropout=lr_dropout, clipnorm=clipnorm, name='GAN_opt') + self.GAN_opt.initialize_variables ( self.GAN.get_weights(), vars_on_cpu=optimizer_vars_on_cpu, lr_dropout_on_cpu=self.options['lr_dropout']=='cpu')#+self.D_src_x2.get_weights() + self.model_filename_list += [ (self.GAN_opt, 'GAN_opt.npy') ] + + if self.is_training: + # Adjust batch size for multiple GPU + gpu_count = max(1, len(devices) ) + bs_per_gpu = max(1, self.get_batch_size() // gpu_count) + self.set_batch_size( gpu_count*bs_per_gpu) + + # Compute losses per GPU + gpu_pred_src_src_list = [] + gpu_pred_dst_dst_list = [] + gpu_pred_src_dst_list = [] + gpu_pred_src_srcm_list = [] + gpu_pred_dst_dstm_list = [] + gpu_pred_src_dstm_list = [] + + gpu_src_losses = [] + gpu_dst_losses = [] + gpu_G_loss_gvs = [] + gpu_GAN_loss_gvs = [] + gpu_D_code_loss_gvs = [] + gpu_D_src_dst_loss_gvs = [] + + for gpu_id in range(gpu_count): + with tf.device( f'/{devices[gpu_id].tf_dev_type}:{gpu_id}' if len(devices) != 0 else f'/CPU:0' ): + with tf.device(f'/CPU:0'): + # slice on CPU, otherwise all batch data will be transfered to GPU first + batch_slice = slice( gpu_id*bs_per_gpu, (gpu_id+1)*bs_per_gpu ) + gpu_warped_src = self.warped_src [batch_slice,:,:,:] + gpu_warped_dst = self.warped_dst [batch_slice,:,:,:] + gpu_target_src = self.target_src [batch_slice,:,:,:] + gpu_target_dst = self.target_dst [batch_slice,:,:,:] + gpu_target_srcm = self.target_srcm[batch_slice,:,:,:] + gpu_target_srcm_em = self.target_srcm_em[batch_slice,:,:,:] + gpu_target_dstm = self.target_dstm[batch_slice,:,:,:] + gpu_target_dstm_em = self.target_dstm_em[batch_slice,:,:,:] + + # process model tensors + gpu_src_code = self.encoder (gpu_warped_src) + gpu_dst_code = self.encoder (gpu_warped_dst) + + if pretrain: + gpu_src_inter_src_code = self.inter_src (gpu_src_code) + gpu_dst_inter_dst_code = self.inter_dst (gpu_dst_code) + gpu_src_code = gpu_src_inter_src_code * nn.random_binomial( [bs_per_gpu, gpu_src_inter_src_code.shape.as_list()[1], 1,1] , p=morph_factor) + gpu_dst_code = gpu_src_dst_code = gpu_dst_inter_dst_code * nn.random_binomial( [bs_per_gpu, gpu_dst_inter_dst_code.shape.as_list()[1], 1,1] , p=0.25) + else: + gpu_src_inter_src_code = self.inter_src (gpu_src_code) + gpu_src_inter_dst_code = self.inter_dst (gpu_src_code) + gpu_dst_inter_src_code = self.inter_src (gpu_dst_code) + gpu_dst_inter_dst_code = self.inter_dst (gpu_dst_code) + + inter_rnd_binomial = nn.random_binomial( [bs_per_gpu, gpu_src_inter_src_code.shape.as_list()[1], 1,1] , p=morph_factor) + gpu_src_code = gpu_src_inter_src_code * inter_rnd_binomial + gpu_src_inter_dst_code * (1-inter_rnd_binomial) + gpu_dst_code = gpu_dst_inter_dst_code + + ae_dims_slice = tf.cast(ae_dims*self.morph_value_t[0], tf.int32) + gpu_src_dst_code = tf.concat( (tf.slice(gpu_dst_inter_src_code, [0,0,0,0], [-1, ae_dims_slice , lowest_dense_res, lowest_dense_res]), + tf.slice(gpu_dst_inter_dst_code, [0,ae_dims_slice,0,0], [-1,ae_dims-ae_dims_slice, lowest_dense_res,lowest_dense_res]) ), 1 ) + + gpu_pred_src_src, gpu_pred_src_srcm = self.decoder(gpu_src_code) + gpu_pred_dst_dst, gpu_pred_dst_dstm = self.decoder(gpu_dst_code) + gpu_pred_src_dst, gpu_pred_src_dstm = self.decoder(gpu_src_dst_code) + + gpu_pred_src_src_list.append(gpu_pred_src_src) + gpu_pred_dst_dst_list.append(gpu_pred_dst_dst) + gpu_pred_src_dst_list.append(gpu_pred_src_dst) + + gpu_pred_src_srcm_list.append(gpu_pred_src_srcm) + gpu_pred_dst_dstm_list.append(gpu_pred_dst_dstm) + gpu_pred_src_dstm_list.append(gpu_pred_src_dstm) + + gpu_target_srcm_blur = nn.gaussian_blur(gpu_target_srcm, max(1, resolution // 32) ) + gpu_target_srcm_blur = tf.clip_by_value(gpu_target_srcm_blur, 0, 0.5) * 2 + + gpu_target_dstm_blur = nn.gaussian_blur(gpu_target_dstm, max(1, resolution // 32) ) + gpu_target_dstm_blur = tf.clip_by_value(gpu_target_dstm_blur, 0, 0.5) * 2 + + gpu_target_dst_anti_masked = gpu_target_dst*(1.0-gpu_target_dstm_blur) + gpu_target_src_anti_masked = gpu_target_src*(1.0-gpu_target_srcm_blur) + gpu_target_src_masked_opt = gpu_target_src*gpu_target_srcm_blur if masked_training else gpu_target_src + gpu_target_dst_masked_opt = gpu_target_dst*gpu_target_dstm_blur if masked_training else gpu_target_dst + + gpu_pred_src_src_masked_opt = gpu_pred_src_src*gpu_target_srcm_blur if masked_training else gpu_pred_src_src + gpu_pred_src_src_anti_masked = gpu_pred_src_src*(1.0-gpu_target_srcm_blur) + gpu_pred_dst_dst_masked_opt = gpu_pred_dst_dst*gpu_target_dstm_blur if masked_training else gpu_pred_dst_dst + gpu_pred_dst_dst_anti_masked = gpu_pred_dst_dst*(1.0-gpu_target_dstm_blur) + + if self.options['loss_function'] == 'MS-SSIM': + gpu_dst_loss = 10 * nn.MsSsim(bs_per_gpu, input_ch, resolution)(gpu_target_dst_masked_opt, gpu_pred_dst_dst_masked_opt, max_val=1.0) + gpu_dst_loss += tf.reduce_mean ( 10*tf.square ( gpu_target_dst_masked_opt - gpu_pred_dst_dst_masked_opt ), axis=[1,2,3]) + elif self.options['loss_function'] == 'MS-SSIM+L1': + gpu_dst_loss = 10 * nn.MsSsim(bs_per_gpu, input_ch, resolution, use_l1=True)(gpu_target_dst_masked_opt, gpu_pred_dst_dst_masked_opt, max_val=1.0) + else: + if resolution < 256: + gpu_dst_loss = tf.reduce_mean ( 10*nn.dssim(gpu_target_dst_masked_opt, gpu_pred_dst_dst_masked_opt, max_val=1.0, filter_size=int(resolution/11.6) ), axis=[1]) + else: + gpu_dst_loss = tf.reduce_mean ( 5*nn.dssim(gpu_target_dst_masked_opt, gpu_pred_dst_dst_masked_opt, max_val=1.0, filter_size=int(resolution/11.6) ), axis=[1]) + gpu_dst_loss += tf.reduce_mean ( 5*nn.dssim(gpu_target_dst_masked_opt, gpu_pred_dst_dst_masked_opt, max_val=1.0, filter_size=int(resolution/23.2) ), axis=[1]) + gpu_dst_loss += tf.reduce_mean ( 10*tf.square( gpu_target_dst_masked_opt- gpu_pred_dst_dst_masked_opt ), axis=[1,2,3]) + if eyes_mouth_prio: + gpu_dst_loss += tf.reduce_mean ( 300*tf.abs ( gpu_target_dst*gpu_target_dstm_em - gpu_pred_dst_dst*gpu_target_dstm_em ), axis=[1,2,3]) + gpu_dst_loss += tf.reduce_mean ( 10*tf.square( gpu_target_dstm - gpu_pred_dst_dstm ),axis=[1,2,3] ) + gpu_dst_loss += 0.1*tf.reduce_mean(tf.square(gpu_pred_dst_dst_anti_masked-gpu_target_dst_anti_masked),axis=[1,2,3] ) + + if self.options['background_power'] > 0: + bg_factor = self.options['background_power'] + + if self.options['loss_function'] == 'MS-SSIM': + gpu_dst_loss += bg_factor * 10 * nn.MsSsim(bs_per_gpu, input_ch, resolution)(gpu_target_dst, gpu_pred_dst_dst, max_val=1.0) + gpu_dst_loss += bg_factor * tf.reduce_mean ( 10*tf.square ( gpu_target_dst - gpu_pred_dst_dst ), axis=[1,2,3]) + elif self.options['loss_function'] == 'MS-SSIM+L1': + gpu_dst_loss += bg_factor * 10 * nn.MsSsim(bs_per_gpu, input_ch, resolution, use_l1=True)(gpu_target_dst, gpu_pred_dst_dst, max_val=1.0) + else: + if resolution < 256: + gpu_dst_loss += bg_factor * tf.reduce_mean ( 10*nn.dssim(gpu_target_dst, gpu_pred_dst_dst, max_val=1.0, filter_size=int(resolution/11.6)), axis=[1]) + else: + gpu_dst_loss += bg_factor * tf.reduce_mean ( 5*nn.dssim(gpu_target_dst, gpu_pred_dst_dst, max_val=1.0, filter_size=int(resolution/11.6)), axis=[1]) + gpu_dst_loss += bg_factor * tf.reduce_mean ( 5*nn.dssim(gpu_target_dst, gpu_pred_dst_dst, max_val=1.0, filter_size=int(resolution/23.2)), axis=[1]) + gpu_dst_loss += bg_factor * tf.reduce_mean ( 10*tf.square ( gpu_target_dst - gpu_pred_dst_dst ), axis=[1,2,3]) + + gpu_dst_losses += [gpu_dst_loss] + + if not pretrain: + if self.options['loss_function'] == 'MS-SSIM': + gpu_src_loss = 10 * nn.MsSsim(bs_per_gpu, input_ch, resolution)(gpu_target_src_masked_opt, gpu_pred_src_src_masked_opt, max_val=1.0) + gpu_src_loss += tf.reduce_mean ( 10*tf.square ( gpu_target_src_masked_opt - gpu_pred_src_src_masked_opt ), axis=[1,2,3]) + elif self.options['loss_function'] == 'MS-SSIM+L1': + gpu_src_loss = 10 * nn.MsSsim(bs_per_gpu, input_ch, resolution, use_l1=True)(gpu_target_src_masked_opt, gpu_pred_src_src_masked_opt, max_val=1.0) + else: + if resolution < 256: + gpu_src_loss = tf.reduce_mean ( 10*nn.dssim(gpu_target_src_masked_opt, gpu_pred_src_src_masked_opt, max_val=1.0, filter_size=int(resolution/11.6)), axis=[1]) + else: + gpu_src_loss = tf.reduce_mean ( 5*nn.dssim(gpu_target_src_masked_opt, gpu_pred_src_src_masked_opt, max_val=1.0, filter_size=int(resolution/11.6)), axis=[1]) + gpu_src_loss += tf.reduce_mean ( 5*nn.dssim(gpu_target_src_masked_opt, gpu_pred_src_src_masked_opt, max_val=1.0, filter_size=int(resolution/23.2)), axis=[1]) + gpu_src_loss += tf.reduce_mean ( 10*tf.square ( gpu_target_src_masked_opt - gpu_pred_src_src_masked_opt ), axis=[1,2,3]) + + if eyes_mouth_prio: + gpu_src_loss += tf.reduce_mean ( 300*tf.abs ( gpu_target_src*gpu_target_srcm_em - gpu_pred_src_src*gpu_target_srcm_em ), axis=[1,2,3]) + + gpu_src_loss += tf.reduce_mean ( 10*tf.square( gpu_target_srcm - gpu_pred_src_srcm ),axis=[1,2,3] ) + + if self.options['background_power'] > 0: + bg_factor = self.options['background_power'] + + if self.options['loss_function'] == 'MS-SSIM': + gpu_src_loss += bg_factor * 10 * nn.MsSsim(bs_per_gpu, input_ch, resolution)(gpu_target_src, gpu_pred_src_src, max_val=1.0) + gpu_src_loss += bg_factor * tf.reduce_mean ( 10*tf.square ( gpu_target_src - gpu_pred_src_src ), axis=[1,2,3]) + elif self.options['loss_function'] == 'MS-SSIM+L1': + gpu_src_loss += bg_factor * 10 * nn.MsSsim(bs_per_gpu, input_ch, resolution, use_l1=True)(gpu_target_src, gpu_pred_src_src, max_val=1.0) + else: + if resolution < 256: + gpu_src_loss += bg_factor * tf.reduce_mean ( 10*nn.dssim(gpu_target_src, gpu_pred_src_src, max_val=1.0, filter_size=int(resolution/11.6)), axis=[1]) + else: + gpu_src_loss += bg_factor * tf.reduce_mean ( 5*nn.dssim(gpu_target_src, gpu_pred_src_src, max_val=1.0, filter_size=int(resolution/11.6)), axis=[1]) + gpu_src_loss += bg_factor * tf.reduce_mean ( 5*nn.dssim(gpu_target_src, gpu_pred_src_src, max_val=1.0, filter_size=int(resolution/23.2)), axis=[1]) + gpu_src_loss += bg_factor * tf.reduce_mean ( 10*tf.square ( gpu_target_src - gpu_pred_src_src ), axis=[1,2,3]) + else: + gpu_src_loss = gpu_dst_loss + + gpu_src_losses += [gpu_src_loss] + + if pretrain: + gpu_G_loss = gpu_dst_loss + else: + gpu_G_loss = gpu_src_loss + gpu_dst_loss + + def DLossOnes(logits): + return tf.reduce_mean( tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.ones_like(logits), logits=logits), axis=[1,2,3]) + + def DLossZeros(logits): + return tf.reduce_mean( tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.zeros_like(logits), logits=logits), axis=[1,2,3]) + + if gan_power != 0: + gpu_pred_src_src_d, gpu_pred_src_src_d2 = self.GAN(gpu_pred_src_src_masked_opt) + gpu_pred_dst_dst_d, gpu_pred_dst_dst_d2 = self.GAN(gpu_pred_dst_dst_masked_opt) + gpu_target_src_d, gpu_target_src_d2 = self.GAN(gpu_target_src_masked_opt) + gpu_target_dst_d, gpu_target_dst_d2 = self.GAN(gpu_target_dst_masked_opt) + + gpu_D_src_dst_loss = (DLossOnes (gpu_target_src_d) + DLossOnes (gpu_target_src_d2) + \ + DLossZeros(gpu_pred_src_src_d) + DLossZeros(gpu_pred_src_src_d2) + \ + DLossOnes (gpu_target_dst_d) + DLossOnes (gpu_target_dst_d2) + \ + DLossZeros(gpu_pred_dst_dst_d) + DLossZeros(gpu_pred_dst_dst_d2) + ) * ( 1.0 / 8) + + gpu_D_src_dst_loss_gvs += [ nn.gradients (gpu_D_src_dst_loss, self.GAN.get_weights() ) ] + + gpu_G_loss += (DLossOnes(gpu_pred_src_src_d) + DLossOnes(gpu_pred_src_src_d2) + \ + DLossOnes(gpu_pred_dst_dst_d) + DLossOnes(gpu_pred_dst_dst_d2) + ) * gan_power + + if masked_training: + # Minimal src-src-bg rec with total_variation_mse to suppress random bright dots from gan + gpu_G_loss += 0.000001*nn.total_variation_mse(gpu_pred_src_src) + gpu_G_loss += 0.02*tf.reduce_mean(tf.square(gpu_pred_src_src_anti_masked-gpu_target_src_anti_masked),axis=[1,2,3] ) + + gpu_G_loss_gvs += [ nn.gradients ( gpu_G_loss, self.trainable_weights ) ] + + + # Average losses and gradients, and create optimizer update ops + with tf.device(f'/CPU:0'): + pred_src_src = nn.concat(gpu_pred_src_src_list, 0) + pred_dst_dst = nn.concat(gpu_pred_dst_dst_list, 0) + pred_src_dst = nn.concat(gpu_pred_src_dst_list, 0) + pred_src_srcm = nn.concat(gpu_pred_src_srcm_list, 0) + pred_dst_dstm = nn.concat(gpu_pred_dst_dstm_list, 0) + pred_src_dstm = nn.concat(gpu_pred_src_dstm_list, 0) + + with tf.device (models_opt_device): + src_loss = tf.concat(gpu_src_losses, 0) + dst_loss = tf.concat(gpu_dst_losses, 0) + src_dst_loss_gv_op = self.src_dst_opt.get_update_op (nn.average_gv_list (gpu_G_loss_gvs)) + + if gan_power != 0: + src_D_src_dst_loss_gv_op = self.GAN_opt.get_update_op (nn.average_gv_list(gpu_D_src_dst_loss_gvs) ) + #GAN_loss_gv_op = self.src_dst_opt.get_update_op (nn.average_gv_list(gpu_GAN_loss_gvs) ) + + + # Initializing training and view functions + def src_dst_train(warped_src, target_src, target_srcm, target_srcm_em, \ + warped_dst, target_dst, target_dstm, target_dstm_em, ): + s, d, _ = nn.tf_sess.run ( [ src_loss, dst_loss, src_dst_loss_gv_op], + feed_dict={self.warped_src :warped_src, + self.target_src :target_src, + self.target_srcm:target_srcm, + self.target_srcm_em:target_srcm_em, + self.warped_dst :warped_dst, + self.target_dst :target_dst, + self.target_dstm:target_dstm, + self.target_dstm_em:target_dstm_em, + }) + return s, d + self.src_dst_train = src_dst_train + + if gan_power != 0: + def D_src_dst_train(warped_src, target_src, target_srcm, target_srcm_em, \ + warped_dst, target_dst, target_dstm, target_dstm_em, ): + nn.tf_sess.run ([src_D_src_dst_loss_gv_op], feed_dict={self.warped_src :warped_src, + self.target_src :target_src, + self.target_srcm:target_srcm, + self.target_srcm_em:target_srcm_em, + self.warped_dst :warped_dst, + self.target_dst :target_dst, + self.target_dstm:target_dstm, + self.target_dstm_em:target_dstm_em}) + self.D_src_dst_train = D_src_dst_train + + + def AE_view(warped_src, warped_dst, morph_value): + return nn.tf_sess.run ( [pred_src_src, pred_dst_dst, pred_dst_dstm, pred_src_dst, pred_src_dstm], + feed_dict={self.warped_src:warped_src, self.warped_dst:warped_dst, self.morph_value_t:[morph_value] }) + + self.AE_view = AE_view + else: + #Initializing merge function + with tf.device( nn.tf_default_device_name if len(devices) != 0 else f'/CPU:0'): + gpu_dst_code = self.encoder (self.warped_dst) + gpu_dst_inter_src_code = self.inter_src ( gpu_dst_code) + gpu_dst_inter_dst_code = self.inter_dst ( gpu_dst_code) + + ae_dims_slice = tf.cast(ae_dims*self.morph_value_t[0], tf.int32) + gpu_src_dst_code = tf.concat( ( tf.slice(gpu_dst_inter_src_code, [0,0,0,0], [-1, ae_dims_slice , lowest_dense_res, lowest_dense_res]), + tf.slice(gpu_dst_inter_dst_code, [0,ae_dims_slice,0,0], [-1,ae_dims-ae_dims_slice, lowest_dense_res,lowest_dense_res]) ), 1 ) + + gpu_pred_src_dst, gpu_pred_src_dstm = self.decoder(gpu_src_dst_code) + _, gpu_pred_dst_dstm = self.decoder(gpu_dst_inter_dst_code) + + def AE_merge(warped_dst, morph_value): + return nn.tf_sess.run ( [gpu_pred_src_dst, gpu_pred_dst_dstm, gpu_pred_src_dstm], feed_dict={self.warped_dst:warped_dst, self.morph_value_t:[morph_value] }) + + self.AE_merge = AE_merge + + # Loading/initializing all models/optimizers weights + for model, filename in io.progress_bar_generator(self.model_filename_list, "Initializing models"): + if self.pretrain_just_disabled: + do_init = False + if model == self.inter_src or model == self.inter_dst: + do_init = True + else: + do_init = self.is_first_run() + if self.is_training and gan_power != 0 and model == self.GAN: + if self.gan_model_changed: + do_init = True + + if not do_init: + do_init = not model.load_weights( self.get_strpath_storage_for_file(filename) ) + if do_init: + model.init_weights() + + + ############### + + # initializing sample generators + if self.is_training: + training_data_src_path = self.training_data_src_path if not self.pretrain else self.get_pretraining_data_path() + training_data_dst_path = self.training_data_dst_path if not self.pretrain else self.get_pretraining_data_path() + + random_ct_samples_path=training_data_dst_path if ct_mode is not None and not self.pretrain else None + + + cpu_count = min(multiprocessing.cpu_count(), 8) + src_generators_count = cpu_count // 2 + dst_generators_count = cpu_count // 2 + if ct_mode is not None: + src_generators_count = int(src_generators_count * 1.5) + + fs_aug = None + if ct_mode == 'fs-aug': + fs_aug = 'fs-aug' + + channel_type = SampleProcessor.ChannelType.LAB_RAND_TRANSFORM if self.options['random_color'] else SampleProcessor.ChannelType.BGR + + self.set_training_data_generators ([ + SampleGeneratorFace(training_data_src_path, random_ct_samples_path=random_ct_samples_path, debug=self.is_debug(), batch_size=self.get_batch_size(), + sample_process_options=SampleProcessor.Options(random_flip=random_src_flip), + output_sample_types = [ {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':random_warp, + 'random_downsample': self.options['random_downsample'], + 'random_noise': self.options['random_noise'], + 'random_blur': self.options['random_blur'], + 'random_jpeg': self.options['random_jpeg'], + 'transform':True, 'channel_type' : channel_type, 'ct_mode': ct_mode, + 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, + {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':False , 'transform':True, 'channel_type' : channel_type, 'ct_mode': ct_mode, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, + {'sample_type': SampleProcessor.SampleType.FACE_MASK, 'warp':False , 'transform':True, 'channel_type' : SampleProcessor.ChannelType.G, 'face_mask_type' : SampleProcessor.FaceMaskType.FULL_FACE, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, + {'sample_type': SampleProcessor.SampleType.FACE_MASK, 'warp':False , 'transform':True, 'channel_type' : SampleProcessor.ChannelType.G, 'face_mask_type' : SampleProcessor.FaceMaskType.FULL_FACE_EYES, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, + ], + uniform_yaw_distribution=self.options['uniform_yaw'] or self.pretrain, + generators_count=src_generators_count ), + + SampleGeneratorFace(training_data_dst_path, debug=self.is_debug(), batch_size=self.get_batch_size(), + sample_process_options=SampleProcessor.Options(random_flip=random_dst_flip), + output_sample_types = [ {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':random_warp, + 'random_downsample': self.options['random_downsample'], + 'random_noise': self.options['random_noise'], + 'random_blur': self.options['random_blur'], + 'random_jpeg': self.options['random_jpeg'], + 'transform':True, 'channel_type' : channel_type, 'ct_mode': fs_aug, + 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, + {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':False , 'transform':True, 'channel_type' : channel_type, 'ct_mode': fs_aug, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, + {'sample_type': SampleProcessor.SampleType.FACE_MASK, 'warp':False , 'transform':True, 'channel_type' : SampleProcessor.ChannelType.G, 'face_mask_type' : SampleProcessor.FaceMaskType.FULL_FACE, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, + {'sample_type': SampleProcessor.SampleType.FACE_MASK, 'warp':False , 'transform':True, 'channel_type' : SampleProcessor.ChannelType.G, 'face_mask_type' : SampleProcessor.FaceMaskType.FULL_FACE_EYES, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, + ], + uniform_yaw_distribution=self.options['uniform_yaw'] or self.pretrain, + generators_count=dst_generators_count ) + ]) + + self.last_src_samples_loss = [] + self.last_dst_samples_loss = [] + if self.pretrain_just_disabled: + self.update_sample_for_preview(force_new=True) + + + def dump_ckpt(self): + tf = nn.tf + with tf.device (nn.tf_default_device_name): + warped_dst = tf.placeholder (nn.floatx, (None, self.resolution, self.resolution, 3), name='in_face') + warped_dst = tf.transpose(warped_dst, (0,3,1,2)) + morph_value = tf.placeholder (nn.floatx, (1,), name='morph_value') + + gpu_dst_code = self.encoder (warped_dst) + gpu_dst_inter_src_code = self.inter_src ( gpu_dst_code) + gpu_dst_inter_dst_code = self.inter_dst ( gpu_dst_code) + + ae_dims_slice = tf.cast(self.ae_dims*morph_value[0], tf.int32) + gpu_src_dst_code = tf.concat( (tf.slice(gpu_dst_inter_src_code, [0,0,0,0], [-1, ae_dims_slice , self.lowest_dense_res, self.lowest_dense_res]), + tf.slice(gpu_dst_inter_dst_code, [0,ae_dims_slice,0,0], [-1,self.ae_dims-ae_dims_slice, self.lowest_dense_res,self.lowest_dense_res]) ), 1 ) + + gpu_pred_src_dst, gpu_pred_src_dstm = self.decoder(gpu_src_dst_code) + _, gpu_pred_dst_dstm = self.decoder(gpu_dst_inter_dst_code) + + gpu_pred_src_dst = tf.transpose(gpu_pred_src_dst, (0,2,3,1)) + gpu_pred_dst_dstm = tf.transpose(gpu_pred_dst_dstm, (0,2,3,1)) + gpu_pred_src_dstm = tf.transpose(gpu_pred_src_dstm, (0,2,3,1)) + + tf.identity(gpu_pred_dst_dstm, name='out_face_mask') + tf.identity(gpu_pred_src_dst, name='out_celeb_face') + tf.identity(gpu_pred_src_dstm, name='out_celeb_face_mask') + + output_graph_def = tf.graph_util.convert_variables_to_constants( + nn.tf_sess, + tf.get_default_graph().as_graph_def(), + ['out_face_mask','out_celeb_face','out_celeb_face_mask'] + ) + + pb_filepath = self.get_strpath_storage_for_file('.pb') + with tf.gfile.GFile(pb_filepath, "wb") as f: + f.write(output_graph_def.SerializeToString()) + + #override + def get_model_filename_list(self): + return self.model_filename_list + + #override + def onSave(self): + for model, filename in io.progress_bar_generator(self.get_model_filename_list(), "Saving", leave=False): + model.save_weights ( self.get_strpath_storage_for_file(filename) ) + + #override + def should_save_preview_history(self): + return (not io.is_colab() and self.iter % ( 10*(max(1,self.resolution // 64)) ) == 0) or \ + (io.is_colab() and self.iter % 100 == 0) + + #override + def onTrainOneIter(self): + bs = self.get_batch_size() + + ( (warped_src, target_src, target_srcm, target_srcm_em), \ + (warped_dst, target_dst, target_dstm, target_dstm_em) ) = self.generate_next_samples() + + src_loss, dst_loss = self.src_dst_train (warped_src, target_src, target_srcm, target_srcm_em, warped_dst, target_dst, target_dstm, target_dstm_em) + + for i in range(bs): + self.last_src_samples_loss.append ( (target_src[i], target_srcm[i], target_srcm_em[i], src_loss[i] ) ) + self.last_dst_samples_loss.append ( (target_dst[i], target_dstm[i], target_dstm_em[i], dst_loss[i] ) ) + + if len(self.last_src_samples_loss) >= bs*16: + src_samples_loss = sorted(self.last_src_samples_loss, key=operator.itemgetter(3), reverse=True) + dst_samples_loss = sorted(self.last_dst_samples_loss, key=operator.itemgetter(3), reverse=True) + + target_src = np.stack( [ x[0] for x in src_samples_loss[:bs] ] ) + target_srcm = np.stack( [ x[1] for x in src_samples_loss[:bs] ] ) + target_srcm_em = np.stack( [ x[2] for x in src_samples_loss[:bs] ] ) + + target_dst = np.stack( [ x[0] for x in dst_samples_loss[:bs] ] ) + target_dstm = np.stack( [ x[1] for x in dst_samples_loss[:bs] ] ) + target_dstm_em = np.stack( [ x[2] for x in dst_samples_loss[:bs] ] ) + + src_loss, dst_loss = self.src_dst_train (target_src, target_src, target_srcm, target_srcm_em, target_dst, target_dst, target_dstm, target_dstm_em) + self.last_src_samples_loss = [] + self.last_dst_samples_loss = [] + + if self.gan_power != 0: + self.D_src_dst_train (warped_src, target_src, target_srcm, target_srcm_em, warped_dst, target_dst, target_dstm, target_dstm_em) + + return ( ('src_loss', np.mean(src_loss) ), ('dst_loss', np.mean(dst_loss) ), ) + + #override + def onGetPreview(self, samples): + ( (warped_src, target_src, target_srcm, target_srcm_em), + (warped_dst, target_dst, target_dstm, target_dstm_em) ) = samples + + S, D, SS, DD, DDM_000, _, _ = [ np.clip( nn.to_data_format(x,"NHWC", self.model_data_format), 0.0, 1.0) for x in ([target_src,target_dst] + self.AE_view (target_src, target_dst, 0.0) ) ] + + _, _, DDM_025, SD_025, SDM_025 = [ np.clip( nn.to_data_format(x,"NHWC", self.model_data_format), 0.0, 1.0) for x in self.AE_view (target_src, target_dst, 0.25) ] + _, _, DDM_050, SD_050, SDM_050 = [ np.clip( nn.to_data_format(x,"NHWC", self.model_data_format), 0.0, 1.0) for x in self.AE_view (target_src, target_dst, 0.50) ] + _, _, DDM_065, SD_065, SDM_065 = [ np.clip( nn.to_data_format(x,"NHWC", self.model_data_format), 0.0, 1.0) for x in self.AE_view (target_src, target_dst, 0.65) ] + _, _, DDM_075, SD_075, SDM_075 = [ np.clip( nn.to_data_format(x,"NHWC", self.model_data_format), 0.0, 1.0) for x in self.AE_view (target_src, target_dst, 0.75) ] + _, _, DDM_100, SD_100, SDM_100 = [ np.clip( nn.to_data_format(x,"NHWC", self.model_data_format), 0.0, 1.0) for x in self.AE_view (target_src, target_dst, 1.00) ] + + (DDM_000, + DDM_025, SDM_025, + DDM_050, SDM_050, + DDM_065, SDM_065, + DDM_075, SDM_075, + DDM_100, SDM_100) = [ np.repeat (x, (3,), -1) for x in (DDM_000, + DDM_025, SDM_025, + DDM_050, SDM_050, + DDM_065, SDM_065, + DDM_075, SDM_075, + DDM_100, SDM_100) ] + + target_srcm, target_dstm = [ nn.to_data_format(x,"NHWC", self.model_data_format) for x in ([target_srcm, target_dstm] )] + + n_samples = min(4, self.get_batch_size(), 800 // self.resolution ) + + result = [] + + i = np.random.randint(n_samples) + + st = [ np.concatenate ((S[i], D[i], DD[i]*DDM_000[i]), axis=1) ] + st += [ np.concatenate ((SS[i], DD[i], SD_075[i] ), axis=1) ] + + result += [ ('AMP morph 0.75', np.concatenate (st, axis=0 )), ] + + st = [ np.concatenate ((DD[i], SD_025[i], SD_050[i]), axis=1) ] + st += [ np.concatenate ((SD_065[i], SD_075[i], SD_100[i]), axis=1) ] + result += [ ('AMP morph list', np.concatenate (st, axis=0 )), ] + + + st = [ np.concatenate ((DD[i], SD_025[i]*DDM_025[i]*SDM_025[i], SD_050[i]*DDM_050[i]*SDM_050[i]), axis=1) ] + st += [ np.concatenate ((SD_065[i]*DDM_065[i]*SDM_065[i], SD_075[i]*DDM_075[i]*SDM_075[i], SD_100[i]*DDM_100[i]*SDM_100[i]), axis=1) ] + result += [ ('AMP morph list masked', np.concatenate (st, axis=0 )), ] + + return result + + def predictor_func (self, face, morph_value): + face = nn.to_data_format(face[None,...], self.model_data_format, "NHWC") + + bgr, mask_dst_dstm, mask_src_dstm = [ nn.to_data_format(x,"NHWC", self.model_data_format).astype(np.float32) for x in self.AE_merge (face, morph_value) ] + + return bgr[0], mask_src_dstm[0][...,0], mask_dst_dstm[0][...,0] + + #override + def get_MergerConfig(self): + morph_factor = np.clip ( io.input_number ("Morph factor", 0.75, add_info="0.0 .. 1.0"), 0.0, 1.0 ) + + def predictor_morph(face): + return self.predictor_func(face, morph_factor) + + + import merger + return predictor_morph, (self.options['resolution'], self.options['resolution'], 3), merger.MergerConfigMasked(face_type=self.face_type, default_mode = 'overlay') + +Model = AMPModel diff --git a/models/Model_AMP/__init__.py b/models/Model_AMP/__init__.py new file mode 100644 index 0000000..0188f11 --- /dev/null +++ b/models/Model_AMP/__init__.py @@ -0,0 +1 @@ +from .Model import Model diff --git a/models/Model_Quick96/Model.py b/models/Model_Quick96/Model.py index 3c39e46..fa139e5 100644 --- a/models/Model_Quick96/Model.py +++ b/models/Model_Quick96/Model.py @@ -31,7 +31,7 @@ class QModel(ModelBase): masked_training = True models_opt_on_gpu = len(devices) >= 1 and all([dev.total_mem_gb >= 4 for dev in devices]) - models_opt_device = '/GPU:0' if models_opt_on_gpu and self.is_training else '/CPU:0' + models_opt_device = nn.tf_default_device_name if models_opt_on_gpu and self.is_training else '/CPU:0' optimizer_vars_on_cpu = models_opt_device=='/CPU:0' input_ch = 3 @@ -96,7 +96,7 @@ class QModel(ModelBase): gpu_src_dst_loss_gvs = [] for gpu_id in range(gpu_count): - with tf.device( f'/GPU:{gpu_id}' if len(devices) != 0 else f'/CPU:0' ): + with tf.device( f'/{devices[gpu_id].tf_dev_type}:{gpu_id}' if len(devices) != 0 else f'/CPU:0' ): batch_slice = slice( gpu_id*bs_per_gpu, (gpu_id+1)*bs_per_gpu ) with tf.device(f'/CPU:0'): # slice on CPU, otherwise all batch data will be transfered to GPU first @@ -190,7 +190,7 @@ class QModel(ModelBase): self.AE_view = AE_view else: # Initializing merge function - with tf.device( f'/GPU:0' if len(devices) != 0 else f'/CPU:0'): + with tf.device( nn.tf_default_device_name if len(devices) != 0 else f'/CPU:0'): gpu_dst_code = self.inter(self.encoder(self.warped_dst)) gpu_pred_src_dst, gpu_pred_src_dstm = self.decoder_src(gpu_dst_code) _, gpu_pred_dst_dstm = self.decoder_dst(gpu_dst_code) diff --git a/models/Model_SAEHD/Model.py b/models/Model_SAEHD/Model.py index 02dbddf..ecac22b 100644 --- a/models/Model_SAEHD/Model.py +++ b/models/Model_SAEHD/Model.py @@ -53,20 +53,32 @@ class SAEHDModel(ModelBase): lr_dropout = {True:'y', False:'n'}.get(lr_dropout, lr_dropout) #backward comp default_lr_dropout = self.options['lr_dropout'] = lr_dropout + default_loss_function = self.options['loss_function'] = self.load_or_def_option('loss_function', 'SSIM') + default_random_warp = self.options['random_warp'] = self.load_or_def_option('random_warp', True) + default_random_downsample = self.options['random_downsample'] = self.load_or_def_option('random_downsample', False) + default_random_noise = self.options['random_noise'] = self.load_or_def_option('random_noise', False) + default_random_blur = self.options['random_blur'] = self.load_or_def_option('random_blur', False) + default_random_jpeg = self.options['random_jpeg'] = self.load_or_def_option('random_jpeg', False) + + default_background_power = self.options['background_power'] = self.load_or_def_option('background_power', 0.0) default_true_face_power = self.options['true_face_power'] = self.load_or_def_option('true_face_power', 0.0) default_face_style_power = self.options['face_style_power'] = self.load_or_def_option('face_style_power', 0.0) default_bg_style_power = self.options['bg_style_power'] = self.load_or_def_option('bg_style_power', 0.0) default_ct_mode = self.options['ct_mode'] = self.load_or_def_option('ct_mode', 'none') + default_random_color = self.options['random_color'] = self.load_or_def_option('random_color', False) default_clipgrad = self.options['clipgrad'] = self.load_or_def_option('clipgrad', False) default_pretrain = self.options['pretrain'] = self.load_or_def_option('pretrain', False) ask_override = self.ask_override() if self.is_first_run() or ask_override: + self.ask_session_name() self.ask_autobackup_hour() + self.ask_maximum_n_backups() self.ask_write_preview_history() self.ask_target_iter() - self.ask_random_flip() + self.ask_random_src_flip() + self.ask_random_dst_flip() self.ask_batch_size(suggest_batch_size) if self.is_first_run(): @@ -74,7 +86,7 @@ class SAEHDModel(ModelBase): resolution = np.clip ( (resolution // 16) * 16, min_res, max_res) self.options['resolution'] = resolution self.options['face_type'] = io.input_str ("Face type", default_face_type, ['h','mf','f','wf','head', 'custom'], help_message="Half / mid face / full face / whole face / head / custom. Half face has better resolution, but covers less area of cheeks. Mid face is 30% wider than half face. 'Whole face' covers full area of face include forehead. 'head' covers full head, but requires XSeg for src and dst faceset.").lower() - + while True: archi = io.input_str ("AE architecture", default_archi, help_message=\ """ @@ -135,11 +147,14 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... self.options['mouth_prio'] = io.input_bool ("Mouth priority", default_mouth_prio, help_message='Helps to fix mouth problems during training by forcing the neural network to train mouth with higher priority similar to eyes ') self.options['uniform_yaw'] = io.input_bool ("Uniform yaw distribution of samples", default_uniform_yaw, help_message='Helps to fix blurry side faces due to small amount of them in the faceset.') - + + default_gan_version = self.options['gan_version'] = self.load_or_def_option('gan_version', 2) default_gan_power = self.options['gan_power'] = self.load_or_def_option('gan_power', 0.0) default_gan_patch_size = self.options['gan_patch_size'] = self.load_or_def_option('gan_patch_size', self.options['resolution'] // 8) default_gan_dims = self.options['gan_dims'] = self.load_or_def_option('gan_dims', 16) - + default_gan_smoothing = self.options['gan_smoothing'] = self.load_or_def_option('gan_smoothing', 0.1) + default_gan_noise = self.options['gan_noise'] = self.load_or_def_option('gan_noise', 0.0) + if self.is_first_run() or ask_override: self.options['models_opt_on_gpu'] = io.input_bool ("Place models and optimizer on GPU", default_models_opt_on_gpu, help_message="When you train on one GPU, by default model and optimizer weights are placed on GPU to accelerate the process. You can place they on CPU to free up extra VRAM, thus set bigger dimensions.") @@ -147,33 +162,53 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... self.options['lr_dropout'] = io.input_str (f"Use learning rate dropout", default_lr_dropout, ['n','y','cpu'], help_message="When the face is trained enough, you can enable this option to get extra sharpness and reduce subpixel shake for less amount of iterations. Enabled it before `disable random warp` and before GAN. \nn - disabled.\ny - enabled\ncpu - enabled on CPU. This allows not to use extra VRAM, sacrificing 20% time of iteration.") + self.options['loss_function'] = io.input_str(f"Loss function", default_loss_function, ['SSIM', 'MS-SSIM', 'MS-SSIM+L1'], + help_message="Change loss function used for image quality assessment.") + self.options['random_warp'] = io.input_bool ("Enable random warp of samples", default_random_warp, help_message="Random warp is required to generalize facial expressions of both faces. When the face is trained enough, you can disable it to get extra sharpness and reduce subpixel shake for less amount of iterations.") - self.options['gan_power'] = np.clip ( io.input_number ("GAN power", default_gan_power, add_info="0.0 .. 1.0", help_message="Forces the neural network to learn small details of the face. Enable it only when the face is trained enough with lr_dropout(on) and random_warp(off), and don't disable. The higher the value, the higher the chances of artifacts. Typical fine value is 0.1"), 0.0, 1.0 ) - - if self.options['gan_power'] != 0.0: - gan_patch_size = np.clip ( io.input_int("GAN patch size", default_gan_patch_size, add_info="3-640", help_message="The higher patch size, the higher the quality, the more VRAM is required. You can get sharper edges even at the lowest setting. Typical fine value is resolution / 8." ), 3, 640 ) - self.options['gan_patch_size'] = gan_patch_size - - gan_dims = np.clip ( io.input_int("GAN dimensions", default_gan_dims, add_info="4-64", help_message="The dimensions of the GAN network. The higher dimensions, the more VRAM is required. You can get sharper edges even at the lowest setting. Typical fine value is 16." ), 4, 64 ) - self.options['gan_dims'] = gan_dims - + self.options['random_downsample'] = io.input_bool("Enable random downsample of samples", default_random_downsample, help_message="") + self.options['random_noise'] = io.input_bool("Enable random noise added to samples", default_random_noise, help_message="") + self.options['random_blur'] = io.input_bool("Enable random blur of samples", default_random_blur, help_message="") + self.options['random_jpeg'] = io.input_bool("Enable random jpeg compression of samples", default_random_jpeg, help_message="") + + self.options['gan_version'] = np.clip (io.input_int("GAN version", default_gan_version, add_info="2 or 3", help_message="Choose GAN version (v2: 7/16/2020, v3: 1/3/2021):"), 2, 3) + + if self.options['gan_version'] == 2: + self.options['gan_power'] = np.clip ( io.input_number ("GAN power", default_gan_power, add_info="0.0 .. 10.0", help_message="Train the network in Generative Adversarial manner. Forces the neural network to learn small details of the face. Enable it only when the face is trained enough and don't disable. Typical value is 0.1"), 0.0, 10.0 ) + else: + self.options['gan_power'] = np.clip ( io.input_number ("GAN power", default_gan_power, add_info="0.0 .. 1.0", help_message="Forces the neural network to learn small details of the face. Enable it only when the face is trained enough with lr_dropout(on) and random_warp(off), and don't disable. The higher the value, the higher the chances of artifacts. Typical fine value is 0.1"), 0.0, 1.0 ) + + if self.options['gan_power'] != 0.0: + if self.options['gan_version'] == 3: + gan_patch_size = np.clip ( io.input_int("GAN patch size", default_gan_patch_size, add_info="3-640", help_message="The higher patch size, the higher the quality, the more VRAM is required. You can get sharper edges even at the lowest setting. Typical fine value is resolution / 8." ), 3, 640 ) + self.options['gan_patch_size'] = gan_patch_size + + gan_dims = np.clip ( io.input_int("GAN dimensions", default_gan_dims, add_info="4-64", help_message="The dimensions of the GAN network. The higher dimensions, the more VRAM is required. You can get sharper edges even at the lowest setting. Typical fine value is 16." ), 4, 64 ) + self.options['gan_dims'] = gan_dims + + self.options['gan_smoothing'] = np.clip ( io.input_number("GAN label smoothing", default_gan_smoothing, add_info="0 - 0.5", help_message="Uses soft labels with values slightly off from 0/1 for GAN, has a regularizing effect"), 0, 0.5) + self.options['gan_noise'] = np.clip ( io.input_number("GAN noisy labels", default_gan_noise, add_info="0 - 0.5", help_message="Marks some images with the wrong label, helps prevent collapse"), 0, 0.5) + if 'df' in self.options['archi']: self.options['true_face_power'] = np.clip ( io.input_number ("'True face' power.", default_true_face_power, add_info="0.0000 .. 1.0", help_message="Experimental option. Discriminates result face to be more like src face. Higher value - stronger discrimination. Typical value is 0.01 . Comparison - https://i.imgur.com/czScS9q.png"), 0.0, 1.0 ) else: self.options['true_face_power'] = 0.0 + self.options['background_power'] = np.clip ( io.input_number("Background power", default_background_power, add_info="0.0..1.0", help_message="Learn the area outside of the mask. Helps smooth out area near the mask boundaries. Can be used at any time"), 0.0, 1.0 ) + self.options['face_style_power'] = np.clip ( io.input_number("Face style power", default_face_style_power, add_info="0.0..100.0", help_message="Learn the color of the predicted face to be the same as dst inside mask. If you want to use this option with 'whole_face' you have to use XSeg trained mask. Warning: Enable it only after 10k iters, when predicted face is clear enough to start learn style. Start from 0.001 value and check history changes. Enabling this option increases the chance of model collapse."), 0.0, 100.0 ) self.options['bg_style_power'] = np.clip ( io.input_number("Background style power", default_bg_style_power, add_info="0.0..100.0", help_message="Learn the area outside mask of the predicted face to be the same as dst. If you want to use this option with 'whole_face' you have to use XSeg trained mask. For whole_face you have to use XSeg trained mask. This can make face more like dst. Enabling this option increases the chance of model collapse. Typical value is 2.0"), 0.0, 100.0 ) self.options['ct_mode'] = io.input_str (f"Color transfer for src faceset", default_ct_mode, ['none','rct','lct','mkl','idt','sot', 'fs-aug'], help_message="Change color distribution of src samples close to dst samples. Try all modes to find the best. FS aug adds random color to dst and src") + self.options['random_color'] = io.input_bool ("Random color", default_random_color, help_message="Samples are randomly rotated around the L axis in LAB colorspace, helps generalize training") self.options['clipgrad'] = io.input_bool ("Enable gradient clipping", default_clipgrad, help_message="Gradient clipping reduces chance of model collapse, sacrificing speed of training.") - self.options['pretrain'] = io.input_bool ("Enable pretraining mode", default_pretrain, help_message="Pretrain the model with large amount of various faces. After that, model can be used to train the fakes more quickly.") + self.options['pretrain'] = io.input_bool ("Enable pretraining mode", default_pretrain, help_message="Pretrain the model with large amount of various faces. After that, model can be used to train the fakes more quickly. Forces random_warp=N, random_flips=Y, gan_power=0.0, lr_dropout=N, styles=0.0, uniform_yaw=Y") if self.options['pretrain'] and self.get_pretraining_data_path() is None: raise Exception("pretraining_data_path is not defined") - + self.gan_model_changed = (default_gan_patch_size != self.options['gan_patch_size']) or (default_gan_dims != self.options['gan_dims']) self.pretrain_just_disabled = (default_pretrain == True and self.options['pretrain'] == False) @@ -203,6 +238,8 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... archi_type, archi_opts = archi_split elif len(archi_split) == 1: archi_type, archi_opts = archi_split[0], None + + self.archi_type = archi_type ae_dims = self.options['ae_dims'] e_dims = self.options['e_dims'] @@ -216,7 +253,9 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... self.gan_power = gan_power = 0.0 if self.pretrain else self.options['gan_power'] random_warp = False if self.pretrain else self.options['random_warp'] - + random_src_flip = self.random_src_flip if not self.pretrain else True + random_dst_flip = self.random_dst_flip if not self.pretrain else True + if self.pretrain: self.options_show_override['gan_power'] = 0.0 self.options_show_override['random_warp'] = False @@ -229,28 +268,29 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... ct_mode = self.options['ct_mode'] if ct_mode == 'none': ct_mode = None - + + models_opt_on_gpu = False if len(devices) == 0 else self.options['models_opt_on_gpu'] - models_opt_device = '/GPU:0' if models_opt_on_gpu and self.is_training else '/CPU:0' + models_opt_device = nn.tf_default_device_name if models_opt_on_gpu and self.is_training else '/CPU:0' optimizer_vars_on_cpu = models_opt_device=='/CPU:0' input_ch=3 - bgr_shape = nn.get4Dshape(resolution,resolution,input_ch) + bgr_shape = self.bgr_shape = nn.get4Dshape(resolution,resolution,input_ch) mask_shape = nn.get4Dshape(resolution,resolution,1) self.model_filename_list = [] with tf.device ('/CPU:0'): #Place holders on CPU - self.warped_src = tf.placeholder (nn.floatx, bgr_shape) - self.warped_dst = tf.placeholder (nn.floatx, bgr_shape) + self.warped_src = tf.placeholder (nn.floatx, bgr_shape, name='warped_src') + self.warped_dst = tf.placeholder (nn.floatx, bgr_shape, name='warped_dst') - self.target_src = tf.placeholder (nn.floatx, bgr_shape) - self.target_dst = tf.placeholder (nn.floatx, bgr_shape) + self.target_src = tf.placeholder (nn.floatx, bgr_shape, name='target_src') + self.target_dst = tf.placeholder (nn.floatx, bgr_shape, name='target_dst') - self.target_srcm = tf.placeholder (nn.floatx, mask_shape) - self.target_srcm_em = tf.placeholder (nn.floatx, mask_shape) - self.target_dstm = tf.placeholder (nn.floatx, mask_shape) - self.target_dstm_em = tf.placeholder (nn.floatx, mask_shape) + self.target_srcm = tf.placeholder (nn.floatx, mask_shape, name='target_srcm') + self.target_srcm_em = tf.placeholder (nn.floatx, mask_shape, name='target_srcm_em') + self.target_dstm = tf.placeholder (nn.floatx, mask_shape, name='target_dstm') + self.target_dstm_em = tf.placeholder (nn.floatx, mask_shape, name='target_dstm_em') # Initializing model classes model_archi = nn.DeepFakeArchi(resolution, opts=archi_opts) @@ -294,8 +334,12 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... if self.is_training: if gan_power != 0: - self.D_src = nn.UNetPatchDiscriminator(patch_size=self.options['gan_patch_size'], in_ch=input_ch, base_ch=self.options['gan_dims'], name="D_src") - self.model_filename_list += [ [self.D_src, 'GAN.npy'] ] + if self.options['gan_version'] == 2: + self.D_src = nn.UNetPatchDiscriminatorV2(patch_size=resolution//16, in_ch=input_ch, name="D_src") + self.model_filename_list += [ [self.D_src, 'D_src_v2.npy'] ] + else: + self.D_src = nn.UNetPatchDiscriminator(patch_size=self.options['gan_patch_size'], in_ch=input_ch, base_ch=self.options['gan_dims'], name="D_src") + self.model_filename_list += [ [self.D_src, 'GAN.npy'] ] # Initialize optimizers lr=5e-5 @@ -320,9 +364,14 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... self.model_filename_list += [ (self.D_code_opt, 'D_code_opt.npy') ] if gan_power != 0: - self.D_src_dst_opt = OptimizerClass(lr=lr, lr_dropout=lr_dropout, clipnorm=clipnorm, name='GAN_opt') - self.D_src_dst_opt.initialize_variables ( self.D_src.get_weights(), vars_on_cpu=optimizer_vars_on_cpu, lr_dropout_on_cpu=self.options['lr_dropout']=='cpu')#+self.D_src_x2.get_weights() - self.model_filename_list += [ (self.D_src_dst_opt, 'GAN_opt.npy') ] + if self.options['gan_version'] == 2: + self.D_src_dst_opt = OptimizerClass(lr=lr, lr_dropout=lr_dropout, clipnorm=clipnorm, name='D_src_dst_opt') + self.D_src_dst_opt.initialize_variables ( self.D_src.get_weights(), vars_on_cpu=optimizer_vars_on_cpu, lr_dropout_on_cpu=self.options['lr_dropout']=='cpu')#+self.D_src_x2.get_weights() + self.model_filename_list += [ (self.D_src_dst_opt, 'D_src_v2_opt.npy') ] + else: + self.D_src_dst_opt = OptimizerClass(lr=lr, lr_dropout=lr_dropout, clipnorm=clipnorm, name='GAN_opt') + self.D_src_dst_opt.initialize_variables ( self.D_src.get_weights(), vars_on_cpu=optimizer_vars_on_cpu, lr_dropout_on_cpu=self.options['lr_dropout']=='cpu')#+self.D_src_x2.get_weights() + self.model_filename_list += [ (self.D_src_dst_opt, 'GAN_opt.npy') ] if self.is_training: # Adjust batch size for multiple GPU @@ -330,7 +379,6 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... bs_per_gpu = max(1, self.get_batch_size() // gpu_count) self.set_batch_size( gpu_count*bs_per_gpu) - # Compute losses per GPU gpu_pred_src_src_list = [] gpu_pred_dst_dst_list = [] @@ -344,9 +392,9 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... gpu_G_loss_gvs = [] gpu_D_code_loss_gvs = [] gpu_D_src_dst_loss_gvs = [] + for gpu_id in range(gpu_count): - with tf.device( f'/GPU:{gpu_id}' if len(devices) != 0 else f'/CPU:0' ): - + with tf.device( f'/{devices[gpu_id].tf_dev_type}:{gpu_id}' if len(devices) != 0 else f'/CPU:0' ): with tf.device(f'/CPU:0'): # slice on CPU, otherwise all batch data will be transfered to GPU first batch_slice = slice( gpu_id*bs_per_gpu, (gpu_id+1)*bs_per_gpu ) @@ -355,9 +403,9 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... gpu_target_src = self.target_src [batch_slice,:,:,:] gpu_target_dst = self.target_dst [batch_slice,:,:,:] gpu_target_srcm_all = self.target_srcm[batch_slice,:,:,:] - gpu_target_srcm_em = self.target_srcm_em[batch_slice,:,:,:] + gpu_target_srcm_em = self.target_srcm_em[batch_slice,:,:,:] gpu_target_dstm_all = self.target_dstm[batch_slice,:,:,:] - gpu_target_dstm_em = self.target_dstm_em[batch_slice,:,:,:] + gpu_target_dstm_em = self.target_dstm_em[batch_slice,:,:,:] # process model tensors if 'df' in archi_type: @@ -406,7 +454,7 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... gpu_target_dstm_style_blur = gpu_target_dstm_blur #default style mask is 0.5 on boundary gpu_target_dstm_blur = tf.clip_by_value(gpu_target_dstm_blur, 0, 0.5) * 2 - gpu_target_dst_masked = gpu_target_dst*gpu_target_dstm_blur + gpu_target_dst_masked = gpu_target_dst*gpu_target_dstm_blur gpu_target_dst_style_masked = gpu_target_dst*gpu_target_dstm_style_blur gpu_target_dst_style_anti_masked = gpu_target_dst*(1.0 - gpu_target_dstm_style_blur) @@ -421,12 +469,18 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... gpu_psd_target_dst_style_masked = gpu_pred_src_dst*gpu_target_dstm_style_blur gpu_psd_target_dst_style_anti_masked = gpu_pred_src_dst*(1.0 - gpu_target_dstm_style_blur) - if resolution < 256: - gpu_src_loss = tf.reduce_mean ( 10*nn.dssim(gpu_target_src_masked_opt, gpu_pred_src_src_masked_opt, max_val=1.0, filter_size=int(resolution/11.6)), axis=[1]) + if self.options['loss_function'] == 'MS-SSIM': + gpu_src_loss = 10 * nn.MsSsim(bs_per_gpu, input_ch, resolution)(gpu_target_src_masked_opt, gpu_pred_src_src_masked_opt, max_val=1.0) + gpu_src_loss += tf.reduce_mean ( 10*tf.square ( gpu_target_src_masked_opt - gpu_pred_src_src_masked_opt ), axis=[1,2,3]) + elif self.options['loss_function'] == 'MS-SSIM+L1': + gpu_src_loss = 10 * nn.MsSsim(bs_per_gpu, input_ch, resolution, use_l1=True)(gpu_target_src_masked_opt, gpu_pred_src_src_masked_opt, max_val=1.0) else: - gpu_src_loss = tf.reduce_mean ( 5*nn.dssim(gpu_target_src_masked_opt, gpu_pred_src_src_masked_opt, max_val=1.0, filter_size=int(resolution/11.6)), axis=[1]) - gpu_src_loss += tf.reduce_mean ( 5*nn.dssim(gpu_target_src_masked_opt, gpu_pred_src_src_masked_opt, max_val=1.0, filter_size=int(resolution/23.2)), axis=[1]) - gpu_src_loss += tf.reduce_mean ( 10*tf.square ( gpu_target_src_masked_opt - gpu_pred_src_src_masked_opt ), axis=[1,2,3]) + if resolution < 256: + gpu_src_loss = tf.reduce_mean ( 10*nn.dssim(gpu_target_src_masked_opt, gpu_pred_src_src_masked_opt, max_val=1.0, filter_size=int(resolution/11.6)), axis=[1]) + else: + gpu_src_loss = tf.reduce_mean ( 5*nn.dssim(gpu_target_src_masked_opt, gpu_pred_src_src_masked_opt, max_val=1.0, filter_size=int(resolution/11.6)), axis=[1]) + gpu_src_loss += tf.reduce_mean ( 5*nn.dssim(gpu_target_src_masked_opt, gpu_pred_src_src_masked_opt, max_val=1.0, filter_size=int(resolution/23.2)), axis=[1]) + gpu_src_loss += tf.reduce_mean ( 10*tf.square ( gpu_target_src_masked_opt - gpu_pred_src_src_masked_opt ), axis=[1,2,3]) if eyes_prio or mouth_prio: if eyes_prio and mouth_prio: @@ -440,6 +494,22 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... gpu_src_loss += tf.reduce_mean ( 10*tf.square( gpu_target_srcm - gpu_pred_src_srcm ),axis=[1,2,3] ) + if self.options['background_power'] > 0: + bg_factor = self.options['background_power'] + + if self.options['loss_function'] == 'MS-SSIM': + gpu_src_loss += bg_factor * 10 * nn.MsSsim(bs_per_gpu, input_ch, resolution)(gpu_target_src, gpu_pred_src_src, max_val=1.0) + gpu_src_loss += bg_factor * tf.reduce_mean ( 10*tf.square ( gpu_target_src - gpu_pred_src_src ), axis=[1,2,3]) + elif self.options['loss_function'] == 'MS-SSIM+L1': + gpu_src_loss += bg_factor * 10 * nn.MsSsim(bs_per_gpu, input_ch, resolution, use_l1=True)(gpu_target_src, gpu_pred_src_src, max_val=1.0) + else: + if resolution < 256: + gpu_src_loss += bg_factor * tf.reduce_mean ( 10*nn.dssim(gpu_target_src, gpu_pred_src_src, max_val=1.0, filter_size=int(resolution/11.6)), axis=[1]) + else: + gpu_src_loss += bg_factor * tf.reduce_mean ( 5*nn.dssim(gpu_target_src, gpu_pred_src_src, max_val=1.0, filter_size=int(resolution/11.6)), axis=[1]) + gpu_src_loss += bg_factor * tf.reduce_mean ( 5*nn.dssim(gpu_target_src, gpu_pred_src_src, max_val=1.0, filter_size=int(resolution/23.2)), axis=[1]) + gpu_src_loss += bg_factor * tf.reduce_mean ( 10*tf.square ( gpu_target_src - gpu_pred_src_src ), axis=[1,2,3]) + face_style_power = self.options['face_style_power'] / 100.0 if face_style_power != 0 and not self.pretrain: gpu_src_loss += nn.style_loss(gpu_psd_target_dst_style_masked, gpu_target_dst_style_masked, gaussian_blur_radius=resolution//16, loss_weight=10000*face_style_power) @@ -449,12 +519,18 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... gpu_src_loss += tf.reduce_mean( (10*bg_style_power)*nn.dssim( gpu_psd_target_dst_style_anti_masked, gpu_target_dst_style_anti_masked, max_val=1.0, filter_size=int(resolution/11.6)), axis=[1]) gpu_src_loss += tf.reduce_mean( (10*bg_style_power)*tf.square(gpu_psd_target_dst_style_anti_masked - gpu_target_dst_style_anti_masked), axis=[1,2,3] ) - if resolution < 256: - gpu_dst_loss = tf.reduce_mean ( 10*nn.dssim(gpu_target_dst_masked_opt, gpu_pred_dst_dst_masked_opt, max_val=1.0, filter_size=int(resolution/11.6) ), axis=[1]) + if self.options['loss_function'] == 'MS-SSIM': + gpu_dst_loss = 10 * nn.MsSsim(bs_per_gpu, input_ch, resolution)(gpu_target_dst_masked_opt, gpu_pred_dst_dst_masked_opt, max_val=1.0) + gpu_dst_loss += tf.reduce_mean ( 10*tf.square( gpu_target_dst_masked_opt- gpu_pred_dst_dst_masked_opt ), axis=[1,2,3]) + elif self.options['loss_function'] == 'MS-SSIM+L1': + gpu_dst_loss = 10 * nn.MsSsim(bs_per_gpu, input_ch, resolution, use_l1=True)(gpu_target_dst_masked_opt, gpu_pred_dst_dst_masked_opt, max_val=1.0) else: - gpu_dst_loss = tf.reduce_mean ( 5*nn.dssim(gpu_target_dst_masked_opt, gpu_pred_dst_dst_masked_opt, max_val=1.0, filter_size=int(resolution/11.6) ), axis=[1]) - gpu_dst_loss += tf.reduce_mean ( 5*nn.dssim(gpu_target_dst_masked_opt, gpu_pred_dst_dst_masked_opt, max_val=1.0, filter_size=int(resolution/23.2) ), axis=[1]) - gpu_dst_loss += tf.reduce_mean ( 10*tf.square( gpu_target_dst_masked_opt- gpu_pred_dst_dst_masked_opt ), axis=[1,2,3]) + if resolution < 256: + gpu_dst_loss = tf.reduce_mean ( 10*nn.dssim(gpu_target_dst_masked_opt, gpu_pred_dst_dst_masked_opt, max_val=1.0, filter_size=int(resolution/11.6) ), axis=[1]) + else: + gpu_dst_loss = tf.reduce_mean ( 5*nn.dssim(gpu_target_dst_masked_opt, gpu_pred_dst_dst_masked_opt, max_val=1.0, filter_size=int(resolution/11.6) ), axis=[1]) + gpu_dst_loss += tf.reduce_mean ( 5*nn.dssim(gpu_target_dst_masked_opt, gpu_pred_dst_dst_masked_opt, max_val=1.0, filter_size=int(resolution/23.2) ), axis=[1]) + gpu_dst_loss += tf.reduce_mean ( 10*tf.square( gpu_target_dst_masked_opt- gpu_pred_dst_dst_masked_opt ), axis=[1,2,3]) if eyes_prio or mouth_prio: @@ -467,6 +543,22 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... gpu_dst_loss += tf.reduce_mean ( 300*tf.abs ( gpu_target_dst*gpu_target_part_mask - gpu_pred_dst_dst*gpu_target_part_mask ), axis=[1,2,3]) + if self.options['background_power'] > 0: + bg_factor = self.options['background_power'] + + if self.options['loss_function'] == 'MS-SSIM': + gpu_dst_loss += bg_factor * 10 * nn.MsSsim(bs_per_gpu, input_ch, resolution)(gpu_target_dst, gpu_pred_dst_dst, max_val=1.0) + gpu_dst_loss += bg_factor * tf.reduce_mean ( 10*tf.square ( gpu_target_dst - gpu_pred_dst_dst ), axis=[1,2,3]) + elif self.options['loss_function'] == 'MS-SSIM+L1': + gpu_dst_loss += bg_factor * 10 * nn.MsSsim(bs_per_gpu, input_ch, resolution, use_l1=True)(gpu_target_dst, gpu_pred_dst_dst, max_val=1.0) + else: + if resolution < 256: + gpu_dst_loss += bg_factor * tf.reduce_mean ( 10*nn.dssim(gpu_target_dst, gpu_pred_dst_dst, max_val=1.0, filter_size=int(resolution/11.6)), axis=[1]) + else: + gpu_dst_loss += bg_factor * tf.reduce_mean ( 5*nn.dssim(gpu_target_dst, gpu_pred_dst_dst, max_val=1.0, filter_size=int(resolution/11.6)), axis=[1]) + gpu_dst_loss += bg_factor * tf.reduce_mean ( 5*nn.dssim(gpu_target_dst, gpu_pred_dst_dst, max_val=1.0, filter_size=int(resolution/23.2)), axis=[1]) + gpu_dst_loss += bg_factor * tf.reduce_mean ( 10*tf.square ( gpu_target_dst - gpu_pred_dst_dst ), axis=[1,2,3]) + gpu_dst_loss += tf.reduce_mean ( 10*tf.square( gpu_target_dstm - gpu_pred_dst_dstm ),axis=[1,2,3] ) gpu_src_losses += [gpu_src_loss] @@ -495,34 +587,49 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... gpu_pred_src_src_d, \ gpu_pred_src_src_d2 = self.D_src(gpu_pred_src_src_masked_opt) - gpu_pred_src_src_d_ones = tf.ones_like (gpu_pred_src_src_d) - gpu_pred_src_src_d_zeros = tf.zeros_like(gpu_pred_src_src_d) + def get_smooth_noisy_labels(label, tensor, smoothing=0.1, noise=0.05): + num_labels = self.batch_size + for d in tensor.get_shape().as_list()[1:]: + num_labels *= d - gpu_pred_src_src_d2_ones = tf.ones_like (gpu_pred_src_src_d2) - gpu_pred_src_src_d2_zeros = tf.zeros_like(gpu_pred_src_src_d2) + probs = tf.math.log([[noise, 1-noise]]) if label == 1 else tf.math.log([[1-noise, noise]]) + x = tf.random.categorical(probs, num_labels) + x = tf.cast(x, tf.float32) + x = tf.math.scalar_mul(1-smoothing, x) + # x = x + (smoothing/num_labels) + x = tf.reshape(x, (self.batch_size,) + tuple(tensor.get_shape().as_list()[1:])) + return x - gpu_target_src_d, \ - gpu_target_src_d2 = self.D_src(gpu_target_src_masked_opt) + smoothing = self.options['gan_smoothing'] + noise = self.options['gan_noise'] - gpu_target_src_d_ones = tf.ones_like(gpu_target_src_d) - gpu_target_src_d2_ones = tf.ones_like(gpu_target_src_d2) + gpu_pred_src_src_d_ones = tf.ones_like(gpu_pred_src_src_d) + gpu_pred_src_src_d2_ones = tf.ones_like(gpu_pred_src_src_d2) - gpu_D_src_dst_loss = (DLoss(gpu_target_src_d_ones , gpu_target_src_d) + \ - DLoss(gpu_pred_src_src_d_zeros , gpu_pred_src_src_d) ) * 0.5 + \ - (DLoss(gpu_target_src_d2_ones , gpu_target_src_d2) + \ - DLoss(gpu_pred_src_src_d2_zeros , gpu_pred_src_src_d2) ) * 0.5 + gpu_pred_src_src_d_smooth_zeros = get_smooth_noisy_labels(0, gpu_pred_src_src_d, smoothing=smoothing, noise=noise) + gpu_pred_src_src_d2_smooth_zeros = get_smooth_noisy_labels(0, gpu_pred_src_src_d2, smoothing=smoothing, noise=noise) + + gpu_target_src_d, gpu_target_src_d2 = self.D_src(gpu_target_src_masked_opt) + + gpu_target_src_d_smooth_ones = get_smooth_noisy_labels(1, gpu_target_src_d, smoothing=smoothing, noise=noise) + gpu_target_src_d2_smooth_ones = get_smooth_noisy_labels(1, gpu_target_src_d2, smoothing=smoothing, noise=noise) + + gpu_D_src_dst_loss = DLoss(gpu_target_src_d_smooth_ones, gpu_target_src_d) \ + + DLoss(gpu_pred_src_src_d_smooth_zeros, gpu_pred_src_src_d) \ + + DLoss(gpu_target_src_d2_smooth_ones, gpu_target_src_d2) \ + + DLoss(gpu_pred_src_src_d2_smooth_zeros, gpu_pred_src_src_d2) gpu_D_src_dst_loss_gvs += [ nn.gradients (gpu_D_src_dst_loss, self.D_src.get_weights() ) ]#+self.D_src_x2.get_weights() gpu_G_loss += gan_power*(DLoss(gpu_pred_src_src_d_ones, gpu_pred_src_src_d) + \ DLoss(gpu_pred_src_src_d2_ones, gpu_pred_src_src_d2)) - - + + if masked_training: # Minimal src-src-bg rec with total_variation_mse to suppress random bright dots from gan gpu_G_loss += 0.000001*nn.total_variation_mse(gpu_pred_src_src) gpu_G_loss += 0.02*tf.reduce_mean(tf.square(gpu_pred_src_src_anti_masked-gpu_target_src_anti_masked),axis=[1,2,3] ) - + gpu_G_loss_gvs += [ nn.gradients ( gpu_G_loss, self.src_dst_trainable_weights ) ] @@ -583,13 +690,13 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... def AE_view(warped_src, warped_dst): - return nn.tf_sess.run ( [pred_src_src, pred_dst_dst, pred_dst_dstm, pred_src_dst, pred_src_dstm], + return nn.tf_sess.run ( [pred_src_src, pred_src_srcm, pred_dst_dst, pred_dst_dstm, pred_src_dst, pred_src_dstm], feed_dict={self.warped_src:warped_src, self.warped_dst:warped_dst}) self.AE_view = AE_view else: # Initializing merge function - with tf.device( f'/GPU:0' if len(devices) != 0 else f'/CPU:0'): + with tf.device( nn.tf_default_device_name if len(devices) != 0 else f'/CPU:0'): if 'df' in archi_type: gpu_dst_code = self.inter(self.encoder(self.warped_dst)) gpu_pred_src_dst, gpu_pred_src_dstm = self.decoder_src(gpu_dst_code) @@ -632,7 +739,10 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... if do_init: model.init_weights() - + + + ############### + # initializing sample generators if self.is_training: training_data_src_path = self.training_data_src_path if not self.pretrain else self.get_pretraining_data_path() @@ -650,11 +760,19 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... if ct_mode == 'fs-aug': fs_aug = 'fs-aug' + channel_type = SampleProcessor.ChannelType.LAB_RAND_TRANSFORM if self.options['random_color'] else SampleProcessor.ChannelType.BGR + self.set_training_data_generators ([ SampleGeneratorFace(training_data_src_path, random_ct_samples_path=random_ct_samples_path, debug=self.is_debug(), batch_size=self.get_batch_size(), - sample_process_options=SampleProcessor.Options(random_flip=self.random_flip), - output_sample_types = [ {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':random_warp, 'transform':True, 'channel_type' : SampleProcessor.ChannelType.BGR, 'ct_mode': ct_mode, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, - {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':False , 'transform':True, 'channel_type' : SampleProcessor.ChannelType.BGR, 'ct_mode': ct_mode, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, + sample_process_options=SampleProcessor.Options(random_flip=random_src_flip), + output_sample_types = [ {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':random_warp, + 'random_downsample': self.options['random_downsample'], + 'random_noise': self.options['random_noise'], + 'random_blur': self.options['random_blur'], + 'random_jpeg': self.options['random_jpeg'], + 'transform':True, 'channel_type' : channel_type, 'ct_mode': ct_mode, + 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, + {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':False , 'transform':True, 'channel_type' : channel_type, 'ct_mode': ct_mode, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, {'sample_type': SampleProcessor.SampleType.FACE_MASK, 'warp':False , 'transform':True, 'channel_type' : SampleProcessor.ChannelType.G, 'face_mask_type' : SampleProcessor.FaceMaskType.FULL_FACE, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, {'sample_type': SampleProcessor.SampleType.FACE_MASK, 'warp':False , 'transform':True, 'channel_type' : SampleProcessor.ChannelType.G, 'face_mask_type' : SampleProcessor.FaceMaskType.FULL_FACE_EYES, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, ], @@ -662,9 +780,15 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... generators_count=src_generators_count ), SampleGeneratorFace(training_data_dst_path, debug=self.is_debug(), batch_size=self.get_batch_size(), - sample_process_options=SampleProcessor.Options(random_flip=self.random_flip), - output_sample_types = [ {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':random_warp, 'transform':True, 'channel_type' : SampleProcessor.ChannelType.BGR, 'ct_mode': fs_aug, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, - {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':False , 'transform':True, 'channel_type' : SampleProcessor.ChannelType.BGR, 'ct_mode': fs_aug, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, + sample_process_options=SampleProcessor.Options(random_flip=random_dst_flip), + output_sample_types = [ {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':random_warp, + 'random_downsample': self.options['random_downsample'], + 'random_noise': self.options['random_noise'], + 'random_blur': self.options['random_blur'], + 'random_jpeg': self.options['random_jpeg'], + 'transform':True, 'channel_type' : channel_type, 'ct_mode': fs_aug, + 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, + {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':False , 'transform':True, 'channel_type' : channel_type, 'ct_mode': fs_aug, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, {'sample_type': SampleProcessor.SampleType.FACE_MASK, 'warp':False , 'transform':True, 'channel_type' : SampleProcessor.ChannelType.G, 'face_mask_type' : SampleProcessor.FaceMaskType.FULL_FACE, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, {'sample_type': SampleProcessor.SampleType.FACE_MASK, 'warp':False , 'transform':True, 'channel_type' : SampleProcessor.ChannelType.G, 'face_mask_type' : SampleProcessor.FaceMaskType.FULL_FACE_EYES, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, ], @@ -677,7 +801,44 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... if self.pretrain_just_disabled: self.update_sample_for_preview(force_new=True) + + def dump_ckpt(self): + tf = nn.tf + + + with tf.device ('/CPU:0'): + warped_dst = tf.placeholder (nn.floatx, (None, self.resolution, self.resolution, 3), name='in_face') + warped_dst = tf.transpose(warped_dst, (0,3,1,2)) + + + if 'df' in self.archi_type: + gpu_dst_code = self.inter(self.encoder(warped_dst)) + gpu_pred_src_dst, gpu_pred_src_dstm = self.decoder_src(gpu_dst_code) + _, gpu_pred_dst_dstm = self.decoder_dst(gpu_dst_code) + elif 'liae' in self.archi_type: + gpu_dst_code = self.encoder (warped_dst) + gpu_dst_inter_B_code = self.inter_B (gpu_dst_code) + gpu_dst_inter_AB_code = self.inter_AB (gpu_dst_code) + gpu_dst_code = tf.concat([gpu_dst_inter_B_code,gpu_dst_inter_AB_code], nn.conv2d_ch_axis) + gpu_src_dst_code = tf.concat([gpu_dst_inter_AB_code,gpu_dst_inter_AB_code], nn.conv2d_ch_axis) + + gpu_pred_src_dst, gpu_pred_src_dstm = self.decoder(gpu_src_dst_code) + _, gpu_pred_dst_dstm = self.decoder(gpu_dst_code) + + gpu_pred_src_dst = tf.transpose(gpu_pred_src_dst, (0,2,3,1)) + gpu_pred_dst_dstm = tf.transpose(gpu_pred_dst_dstm, (0,2,3,1)) + gpu_pred_src_dstm = tf.transpose(gpu_pred_src_dstm, (0,2,3,1)) + + + saver = tf.train.Saver() + tf.identity(gpu_pred_dst_dstm, name='out_face_mask') + tf.identity(gpu_pred_src_dst, name='out_celeb_face') + tf.identity(gpu_pred_src_dstm, name='out_celeb_face_mask') + + saver.save(nn.tf_sess, self.get_strpath_storage_for_file('.ckpt') ) + + #override def get_model_filename_list(self): return self.model_filename_list @@ -737,8 +898,9 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... ( (warped_src, target_src, target_srcm, target_srcm_em), (warped_dst, target_dst, target_dstm, target_dstm_em) ) = samples - S, D, SS, DD, DDM, SD, SDM = [ np.clip( nn.to_data_format(x,"NHWC", self.model_data_format), 0.0, 1.0) for x in ([target_src,target_dst] + self.AE_view (target_src, target_dst) ) ] - DDM, SDM, = [ np.repeat (x, (3,), -1) for x in [DDM, SDM] ] + S, D, SS, SSM, DD, DDM, SD, SDM = [ np.clip( nn.to_data_format(x,"NHWC", self.model_data_format), 0.0, 1.0) for x in ([target_src,target_dst] + self.AE_view (target_src, target_dst) ) ] + SW, DW = [ np.clip( nn.to_data_format(x,"NHWC", self.model_data_format), 0.0, 1.0) for x in ([warped_src,warped_dst]) ] + SSM, DDM, SDM, = [ np.repeat (x, (3,), -1) for x in [SSM, DDM, SDM] ] target_srcm, target_dstm = [ nn.to_data_format(x,"NHWC", self.model_data_format) for x in ([target_srcm, target_dstm] )] @@ -753,12 +915,17 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... st.append ( np.concatenate ( ar, axis=1) ) result += [ ('SAEHD', np.concatenate (st, axis=0 )), ] + wt = [] + for i in range(n_samples): + ar = SW[i], SS[i], DW[i], DD[i], SD[i] + wt.append ( np.concatenate ( ar, axis=1) ) + result += [ ('SAEHD warped', np.concatenate (wt, axis=0 )), ] st_m = [] for i in range(n_samples): SD_mask = DDM[i]*SDM[i] if self.face_type < FaceType.HEAD else SDM[i] - ar = S[i]*target_srcm[i], SS[i], D[i]*target_dstm[i], DD[i]*DDM[i], SD[i]*SD_mask + ar = S[i]*target_srcm[i], SS[i]*SSM[i], D[i]*target_dstm[i], DD[i]*DDM[i], SD[i]*SD_mask st_m.append ( np.concatenate ( ar, axis=1) ) result += [ ('SAEHD masked', np.concatenate (st_m, axis=0 )), ] @@ -783,10 +950,27 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... st.append ( np.concatenate ( ar, axis=1) ) result += [ ('SAEHD pred', np.concatenate (st, axis=0 )), ] + wt = [] + for i in range(n_samples): + ar = SW[i], SS[i] + wt.append ( np.concatenate ( ar, axis=1) ) + result += [ ('SAEHD warped src-src', np.concatenate (wt, axis=0 )), ] + + wt = [] + for i in range(n_samples): + ar = DW[i], DD[i] + wt.append ( np.concatenate ( ar, axis=1) ) + result += [ ('SAEHD warped dst-dst', np.concatenate (wt, axis=0 )), ] + + wt = [] + for i in range(n_samples): + ar = DW[i], SD[i] + wt.append ( np.concatenate ( ar, axis=1) ) + result += [ ('SAEHD warped pred', np.concatenate (wt, axis=0 )), ] st_m = [] for i in range(n_samples): - ar = S[i]*target_srcm[i], SS[i] + ar = S[i]*target_srcm[i], SS[i]*SSM[i] st_m.append ( np.concatenate ( ar, axis=1) ) result += [ ('SAEHD masked src-src', np.concatenate (st_m, axis=0 )), ] diff --git a/models/Model_XSeg/Model.py b/models/Model_XSeg/Model.py index 4ab23fc..567a1f2 100644 --- a/models/Model_XSeg/Model.py +++ b/models/Model_XSeg/Model.py @@ -52,7 +52,7 @@ class XSegModel(ModelBase): 'head' : FaceType.HEAD}[ self.options['face_type'] ] place_model_on_cpu = len(devices) == 0 - models_opt_device = '/CPU:0' if place_model_on_cpu else '/GPU:0' + models_opt_device = '/CPU:0' if place_model_on_cpu else nn.tf_default_device_name bgr_shape = nn.get4Dshape(resolution,resolution,3) mask_shape = nn.get4Dshape(resolution,resolution,1) @@ -83,7 +83,7 @@ class XSegModel(ModelBase): for gpu_id in range(gpu_count): - with tf.device( f'/GPU:{gpu_id}' if len(devices) != 0 else f'/CPU:0' ): + with tf.device(f'/{devices[gpu_id].tf_dev_type}:{gpu_id}' if len(devices) != 0 else f'/CPU:0' ): with tf.device(f'/CPU:0'): # slice on CPU, otherwise all batch data will be transfered to GPU first batch_slice = slice( gpu_id*bs_per_gpu, (gpu_id+1)*bs_per_gpu ) @@ -95,6 +95,7 @@ class XSegModel(ModelBase): gpu_pred_list.append(gpu_pred_t) gpu_loss = tf.reduce_mean( tf.nn.sigmoid_cross_entropy_with_logits(labels=gpu_target_t, logits=gpu_pred_logits_t), axis=[1,2,3]) + gpu_losses += [gpu_loss] gpu_loss_gvs += [ nn.gradients ( gpu_loss, self.model.get_weights() ) ] diff --git a/requirements-cuda.txt b/requirements-cuda.txt index ca18994..8833aca 100644 --- a/requirements-cuda.txt +++ b/requirements-cuda.txt @@ -1,10 +1,12 @@ tqdm numpy==1.19.3 -h5py==2.9.0 +h5py==2.10.0 opencv-python==4.1.0.25 ffmpeg-python==0.1.17 scikit-image==0.14.2 scipy==1.4.1 colorama -tensorflow-gpu==2.3.1 -pyqt5 \ No newline at end of file +tensorflow-gpu==2.4.0 +pyqt5 +Flask==1.1.1 +flask-socketio==4.2.1 diff --git a/samplelib/PackedFaceset.py b/samplelib/PackedFaceset.py index 2ea01d9..e7ae1d4 100644 --- a/samplelib/PackedFaceset.py +++ b/samplelib/PackedFaceset.py @@ -84,17 +84,18 @@ class PackedFaceset(): of.write ( struct.pack("Q", offset) ) of.seek(0,2) of.close() + + if io.input_bool(f"Delete original files?", True): + for filename in io.progress_bar_generator(image_paths, "Deleting files"): + Path(filename).unlink() - for filename in io.progress_bar_generator(image_paths, "Deleting files"): - Path(filename).unlink() - - if as_person_faceset: - for dir_name in io.progress_bar_generator(dir_names, "Deleting dirs"): - dir_path = samples_path / dir_name - try: - shutil.rmtree(dir_path) - except: - io.log_info (f"unable to remove: {dir_path} ") + if as_person_faceset: + for dir_name in io.progress_bar_generator(dir_names, "Deleting dirs"): + dir_path = samples_path / dir_name + try: + shutil.rmtree(dir_path) + except: + io.log_info (f"unable to remove: {dir_path} ") @staticmethod def unpack(samples_path): diff --git a/samplelib/SampleGeneratorFaceXSeg.py b/samplelib/SampleGeneratorFaceXSeg.py index 9c7cf6e..7e38e64 100644 --- a/samplelib/SampleGeneratorFaceXSeg.py +++ b/samplelib/SampleGeneratorFaceXSeg.py @@ -6,7 +6,7 @@ from enum import IntEnum import cv2 import numpy as np - +from pathlib import Path from core import imagelib, mplib, pathex from core.imagelib import sd from core.cv2ex import * @@ -31,7 +31,7 @@ class SampleGeneratorFaceXSeg(SampleGeneratorBase): if len(seg_sample_idxs) == 0: raise Exception(f"No segmented faces found.") else: - io.log_info(f"Using {len(seg_sample_idxs)} xseg labeled samples.") + io.log_info(f"Using {len(seg_sample_idxs)} xseg labeled samples.") else: io.log_info(f"Using {len(seg_sample_idxs)} segmented samples.") @@ -40,11 +40,11 @@ class SampleGeneratorFaceXSeg(SampleGeneratorBase): else: self.generators_count = max(1, generators_count) + args = (samples, seg_sample_idxs, resolution, face_type, data_format) if self.debug: - self.generators = [ThisThreadGenerator ( self.batch_func, (samples, seg_sample_idxs, resolution, face_type, data_format) )] + self.generators = [ThisThreadGenerator ( self.batch_func, args )] else: - self.generators = [SubprocessGenerator ( self.batch_func, (samples, seg_sample_idxs, resolution, face_type, data_format), start_now=False ) \ - for i in range(self.generators_count) ] + self.generators = [SubprocessGenerator ( self.batch_func, args, start_now=False ) for i in range(self.generators_count) ] SubprocessGenerator.start_in_parallel( self.generators ) @@ -77,17 +77,19 @@ class SampleGeneratorFaceXSeg(SampleGeneratorBase): ty_range=[-0.05, 0.05] random_bilinear_resize_chance, random_bilinear_resize_max_size_per = 25,75 + sharpen_chance, sharpen_kernel_max_size = 25, 5 motion_blur_chance, motion_blur_mb_max_size = 25, 5 gaussian_blur_chance, gaussian_blur_kernel_max_size = 25, 5 + random_jpeg_compress_chance = 25 def gen_img_mask(sample): img = sample.load_bgr() h,w,c = img.shape - + if sample.seg_ie_polys.has_polys(): mask = np.zeros ((h,w,1), dtype=np.float32) sample.seg_ie_polys.overlay_mask(mask) - elif sample.has_xseg_mask(): + elif sample.has_xseg_mask(): mask = sample.get_xseg_mask() mask[mask < 0.5] = 0.0 mask[mask >= 0.5] = 1.0 @@ -121,7 +123,6 @@ class SampleGeneratorFaceXSeg(SampleGeneratorBase): img, mask = gen_img_mask(sample) if np.random.randint(2) == 0: - if len(bg_shuffle_idxs) == 0: bg_shuffle_idxs = seg_sample_idxs.copy() np.random.shuffle(bg_shuffle_idxs) @@ -130,29 +131,57 @@ class SampleGeneratorFaceXSeg(SampleGeneratorBase): bg_img, bg_mask = gen_img_mask(bg_sample) bg_wp = imagelib.gen_warp_params(resolution, True, rotation_range=[-180,180], scale_range=[-0.10, 0.10], tx_range=[-0.10, 0.10], ty_range=[-0.10, 0.10] ) - bg_img = imagelib.warp_by_params (bg_wp, bg_img, can_warp=False, can_transform=True, can_flip=True, border_replicate=False) + bg_img = imagelib.warp_by_params (bg_wp, bg_img, can_warp=False, can_transform=True, can_flip=True, border_replicate=True) bg_mask = imagelib.warp_by_params (bg_wp, bg_mask, can_warp=False, can_transform=True, can_flip=True, border_replicate=False) + bg_img = bg_img*(1-bg_mask) + if np.random.randint(2) == 0: + bg_img = imagelib.apply_random_hsv_shift(bg_img) + else: + bg_img = imagelib.apply_random_rgb_levels(bg_img) - c_mask = (1-bg_mask) * (1-mask) - img = img*(1-c_mask) + bg_img * c_mask + c_mask = 1.0 - (1-bg_mask) * (1-mask) + rnd = 0.15 + np.random.uniform()*0.85 + img = img*(c_mask) + img*(1-c_mask)*rnd + bg_img*(1-c_mask)*(1-rnd) warp_params = imagelib.gen_warp_params(resolution, random_flip, rotation_range=rotation_range, scale_range=scale_range, tx_range=tx_range, ty_range=ty_range ) - img = imagelib.warp_by_params (warp_params, img, can_warp=True, can_transform=True, can_flip=True, border_replicate=False) + img = imagelib.warp_by_params (warp_params, img, can_warp=True, can_transform=True, can_flip=True, border_replicate=True) mask = imagelib.warp_by_params (warp_params, mask, can_warp=True, can_transform=True, can_flip=True, border_replicate=False) img = np.clip(img.astype(np.float32), 0, 1) mask[mask < 0.5] = 0.0 mask[mask >= 0.5] = 1.0 mask = np.clip(mask, 0, 1) + + if np.random.randint(2) == 0: + # random face flare + krn = np.random.randint( resolution//4, resolution ) + krn = krn - krn % 2 + 1 + img = img + cv2.GaussianBlur(img*mask, (krn,krn), 0) + + if np.random.randint(2) == 0: + # random bg flare + krn = np.random.randint( resolution//4, resolution ) + krn = krn - krn % 2 + 1 + img = img + cv2.GaussianBlur(img*(1-mask), (krn,krn), 0) if np.random.randint(2) == 0: img = imagelib.apply_random_hsv_shift(img, mask=sd.random_circle_faded ([resolution,resolution])) else: img = imagelib.apply_random_rgb_levels(img, mask=sd.random_circle_faded ([resolution,resolution])) + + if np.random.randint(2) == 0: + img = imagelib.apply_random_sharpen( img, sharpen_chance, sharpen_kernel_max_size, mask=sd.random_circle_faded ([resolution,resolution])) + else: + img = imagelib.apply_random_motion_blur( img, motion_blur_chance, motion_blur_mb_max_size, mask=sd.random_circle_faded ([resolution,resolution])) + img = imagelib.apply_random_gaussian_blur( img, gaussian_blur_chance, gaussian_blur_kernel_max_size, mask=sd.random_circle_faded ([resolution,resolution])) + + if np.random.randint(2) == 0: + img = imagelib.apply_random_nearest_resize( img, random_bilinear_resize_chance, random_bilinear_resize_max_size_per, mask=sd.random_circle_faded ([resolution,resolution])) + else: + img = imagelib.apply_random_bilinear_resize( img, random_bilinear_resize_chance, random_bilinear_resize_max_size_per, mask=sd.random_circle_faded ([resolution,resolution])) + img = np.clip(img, 0, 1) - img = imagelib.apply_random_motion_blur( img, motion_blur_chance, motion_blur_mb_max_size, mask=sd.random_circle_faded ([resolution,resolution])) - img = imagelib.apply_random_gaussian_blur( img, gaussian_blur_chance, gaussian_blur_kernel_max_size, mask=sd.random_circle_faded ([resolution,resolution])) - img = imagelib.apply_random_bilinear_resize( img, random_bilinear_resize_chance, random_bilinear_resize_max_size_per, mask=sd.random_circle_faded ([resolution,resolution])) + img = imagelib.apply_random_jpeg_compress( img, random_jpeg_compress_chance, mask=sd.random_circle_faded ([resolution,resolution])) if data_format == "NCHW": img = np.transpose(img, (2,0,1) ) @@ -221,4 +250,48 @@ class SegmentedSampleFilterSubprocessor(Subprocessor): if self.count_xseg_mask: return idx, self.samples[idx].has_xseg_mask() else: - return idx, self.samples[idx].seg_ie_polys.get_pts_count() != 0 \ No newline at end of file + return idx, self.samples[idx].seg_ie_polys.get_pts_count() != 0 + +""" + bg_path = None + for path in paths: + bg_path = Path(path) / 'backgrounds' + if bg_path.exists(): + + break + if bg_path is None: + io.log_info(f'Random backgrounds will not be used. Place no face jpg images to aligned\backgrounds folder. ') + bg_pathes = None + else: + bg_pathes = pathex.get_image_paths(bg_path, image_extensions=['.jpg'], return_Path_class=True) + io.log_info(f'Using {len(bg_pathes)} random backgrounds from {bg_path}') + +if bg_pathes is not None: + bg_path = bg_pathes[ np.random.randint(len(bg_pathes)) ] + + bg_img = cv2_imread(bg_path) + if bg_img is not None: + bg_img = bg_img.astype(np.float32) / 255.0 + bg_img = imagelib.normalize_channels(bg_img, 3) + + bg_img = imagelib.random_crop(bg_img, resolution, resolution) + bg_img = cv2.resize(bg_img, (resolution, resolution), interpolation=cv2.INTER_LINEAR) + + if np.random.randint(2) == 0: + bg_img = imagelib.apply_random_hsv_shift(bg_img) + else: + bg_img = imagelib.apply_random_rgb_levels(bg_img) + + bg_wp = imagelib.gen_warp_params(resolution, True, rotation_range=[-180,180], scale_range=[0,0], tx_range=[0,0], ty_range=[0,0]) + bg_img = imagelib.warp_by_params (bg_wp, bg_img, can_warp=False, can_transform=True, can_flip=True, border_replicate=True) + + bg = img*(1-mask) + fg = img*mask + + c_mask = sd.random_circle_faded ([resolution,resolution]) + bg = ( bg_img*c_mask + bg*(1-c_mask) )*(1-mask) + + img = fg+bg + + else: +""" \ No newline at end of file diff --git a/samplelib/SampleLoader.py b/samplelib/SampleLoader.py index edb3775..2989354 100644 --- a/samplelib/SampleLoader.py +++ b/samplelib/SampleLoader.py @@ -23,7 +23,7 @@ class SampleLoader: try: samples = samplelib.PackedFaceset.load(samples_path) except: - io.log_err(f"Error occured while loading samplelib.PackedFaceset.load {str(samples_dat_path)}, {traceback.format_exc()}") + io.log_err(f"Error occured while loading samplelib.PackedFaceset.load {str(samples_path)}, {traceback.format_exc()}") if samples is None: raise ValueError("packed faceset not found.") diff --git a/samplelib/SampleProcessor.py b/samplelib/SampleProcessor.py index a2f1e0e..3b657d7 100644 --- a/samplelib/SampleProcessor.py +++ b/samplelib/SampleProcessor.py @@ -7,7 +7,8 @@ import numpy as np from core import imagelib from core.cv2ex import * -from core.imagelib import sd +from core.imagelib import sd, LinearMotionBlur +from core.imagelib.color_transfer import random_lab_rotation from facelib import FaceType, LandmarksProcessor @@ -26,6 +27,8 @@ class SampleProcessor(object): BGR = 1 #BGR G = 2 #Grayscale GGG = 3 #3xGrayscale + LAB_RAND_TRANSFORM = 4 # LAB random transform + class FaceMaskType(IntEnum): NONE = 0 @@ -56,18 +59,18 @@ class SampleProcessor(object): sample_landmarks = sample.landmarks ct_sample_bgr = None h,w,c = sample_bgr.shape - - def get_full_face_mask(): - xseg_mask = sample.get_xseg_mask() - if xseg_mask is not None: + + def get_full_face_mask(): + xseg_mask = sample.get_xseg_mask() + if xseg_mask is not None: if xseg_mask.shape[0] != h or xseg_mask.shape[1] != w: - xseg_mask = cv2.resize(xseg_mask, (w,h), interpolation=cv2.INTER_CUBIC) + xseg_mask = cv2.resize(xseg_mask, (w,h), interpolation=cv2.INTER_CUBIC) xseg_mask = imagelib.normalize_channels(xseg_mask, 1) return np.clip(xseg_mask, 0, 1) else: full_face_mask = LandmarksProcessor.get_image_hull_mask (sample_bgr.shape, sample_landmarks, eyebrows_expand_mod=sample.eyebrows_expand_mod ) return np.clip(full_face_mask, 0, 1) - + def get_eyes_mask(): eyes_mask = LandmarksProcessor.get_image_eye_mask (sample_bgr.shape, sample_landmarks) # set eye masks to 1-2 @@ -86,29 +89,33 @@ class SampleProcessor(object): if debug and is_face_sample: LandmarksProcessor.draw_landmarks (sample_bgr, sample_landmarks, (0, 1, 0)) - - params_per_resolution = {} - warp_rnd_state = np.random.RandomState (sample_rnd_seed-1) + + params_per_resolution = {} + warp_rnd_state = np.random.RandomState (sample_rnd_seed-1) for opts in output_sample_types: resolution = opts.get('resolution', None) if resolution is None: continue - params_per_resolution[resolution] = imagelib.gen_warp_params(resolution, - sample_process_options.random_flip, - rotation_range=sample_process_options.rotation_range, - scale_range=sample_process_options.scale_range, - tx_range=sample_process_options.tx_range, - ty_range=sample_process_options.ty_range, + params_per_resolution[resolution] = imagelib.gen_warp_params(resolution, + sample_process_options.random_flip, + rotation_range=sample_process_options.rotation_range, + scale_range=sample_process_options.scale_range, + tx_range=sample_process_options.tx_range, + ty_range=sample_process_options.ty_range, rnd_state=warp_rnd_state) outputs_sample = [] for opts in output_sample_types: sample_type = opts.get('sample_type', SPST.NONE) - channel_type = opts.get('channel_type', SPCT.NONE) + channel_type = opts.get('channel_type', SPCT.NONE) resolution = opts.get('resolution', 0) nearest_resize_to = opts.get('nearest_resize_to', None) warp = opts.get('warp', False) transform = opts.get('transform', False) + random_downsample = opts.get('random_downsample', False) + random_noise = opts.get('random_noise', False) + random_blur = opts.get('random_blur', False) + random_jpeg = opts.get('random_jpeg', False) motion_blur = opts.get('motion_blur', None) gaussian_blur = opts.get('gaussian_blur', None) random_bilinear_resize = opts.get('random_bilinear_resize', None) @@ -118,29 +125,29 @@ class SampleProcessor(object): normalize_tanh = opts.get('normalize_tanh', False) ct_mode = opts.get('ct_mode', None) data_format = opts.get('data_format', 'NHWC') - - if sample_type == SPST.FACE_MASK or sample_type == SPST.IMAGE: + + if sample_type == SPST.FACE_MASK or sample_type == SPST.IMAGE: border_replicate = False elif sample_type == SPST.FACE_IMAGE: border_replicate = True - - + + border_replicate = opts.get('border_replicate', border_replicate) borderMode = cv2.BORDER_REPLICATE if border_replicate else cv2.BORDER_CONSTANT - - + + if sample_type == SPST.FACE_IMAGE or sample_type == SPST.FACE_MASK: - if not is_face_sample: + if not is_face_sample: raise ValueError("face_samples should be provided for sample_type FACE_*") if sample_type == SPST.FACE_IMAGE or sample_type == SPST.FACE_MASK: face_type = opts.get('face_type', None) face_mask_type = opts.get('face_mask_type', SPFMT.NONE) - + if face_type is None: raise ValueError("face_type must be defined for face samples") - if sample_type == SPST.FACE_MASK: + if sample_type == SPST.FACE_MASK: if face_mask_type == SPFMT.FULL_FACE: img = get_full_face_mask() elif face_mask_type == SPFMT.EYES: @@ -149,42 +156,42 @@ class SampleProcessor(object): # sets both eyes and mouth mask parts img = get_full_face_mask() mask = img.copy() - mask[mask != 0.0] = 1.0 + mask[mask != 0.0] = 1.0 eye_mask = get_eyes_mask() * mask img = np.where(eye_mask > 1, eye_mask, img) mouth_mask = get_mouth_mask() * mask - img = np.where(mouth_mask > 2, mouth_mask, img) + img = np.where(mouth_mask > 2, mouth_mask, img) else: img = np.zeros ( sample_bgr.shape[0:2]+(1,), dtype=np.float32) if sample_face_type == FaceType.MARK_ONLY: mat = LandmarksProcessor.get_transform_mat (sample_landmarks, warp_resolution, face_type) img = cv2.warpAffine( img, mat, (warp_resolution, warp_resolution), flags=cv2.INTER_LINEAR ) - + img = imagelib.warp_by_params (params_per_resolution[resolution], img, warp, transform, can_flip=True, border_replicate=border_replicate, cv2_inter=cv2.INTER_LINEAR) img = cv2.resize( img, (resolution,resolution), interpolation=cv2.INTER_LINEAR ) else: if face_type != sample_face_type: - mat = LandmarksProcessor.get_transform_mat (sample_landmarks, resolution, face_type) + mat = LandmarksProcessor.get_transform_mat (sample_landmarks, resolution, face_type) img = cv2.warpAffine( img, mat, (resolution,resolution), borderMode=borderMode, flags=cv2.INTER_LINEAR ) else: if w != resolution: img = cv2.resize( img, (resolution, resolution), interpolation=cv2.INTER_LINEAR ) - + img = imagelib.warp_by_params (params_per_resolution[resolution], img, warp, transform, can_flip=True, border_replicate=border_replicate, cv2_inter=cv2.INTER_LINEAR) - + if len(img.shape) == 2: img = img[...,None] - + if channel_type == SPCT.G: out_sample = img.astype(np.float32) else: raise ValueError("only channel_type.G supported for the mask") elif sample_type == SPST.FACE_IMAGE: - img = sample_bgr - + img = sample_bgr + if random_rgb_levels: random_mask = sd.random_circle_faded ([w,w], rnd_state=np.random.RandomState (sample_rnd_seed) ) if random_circle_mask else None img = imagelib.apply_random_rgb_levels(img, mask=random_mask, rnd_state=np.random.RandomState (sample_rnd_seed) ) @@ -193,15 +200,15 @@ class SampleProcessor(object): random_mask = sd.random_circle_faded ([w,w], rnd_state=np.random.RandomState (sample_rnd_seed+1) ) if random_circle_mask else None img = imagelib.apply_random_hsv_shift(img, mask=random_mask, rnd_state=np.random.RandomState (sample_rnd_seed+1) ) - + if face_type != sample_face_type: mat = LandmarksProcessor.get_transform_mat (sample_landmarks, resolution, face_type) img = cv2.warpAffine( img, mat, (resolution,resolution), borderMode=borderMode, flags=cv2.INTER_CUBIC ) else: if w != resolution: img = cv2.resize( img, (resolution, resolution), interpolation=cv2.INTER_CUBIC ) - - # Apply random color transfer + + # Apply random color transfer if ct_mode is not None and ct_sample is not None or ct_mode == 'fs-aug': if ct_mode == 'fs-aug': img = imagelib.color_augmentation(img, sample_rnd_seed) @@ -210,27 +217,82 @@ class SampleProcessor(object): ct_sample_bgr = ct_sample.load_bgr() img = imagelib.color_transfer (ct_mode, img, cv2.resize( ct_sample_bgr, (resolution,resolution), interpolation=cv2.INTER_LINEAR ) ) - + + randomization_order = ['blur', 'noise', 'jpeg', 'down'] + np.random.shuffle(randomization_order) + for random_distortion in randomization_order: + # Apply random blur + if random_distortion == 'blur' and random_blur: + blur_type = np.random.choice(['motion', 'gaussian']) + + if blur_type == 'motion': + blur_k = np.random.randint(10, 20) + blur_angle = 360 * np.random.random() + img = LinearMotionBlur(img, blur_k, blur_angle) + elif blur_type == 'gaussian': + blur_sigma = 5 * np.random.random() + 3 + + if blur_sigma < 5.0: + kernel_size = 2.9 * blur_sigma # 97% of weight + else: + kernel_size = 2.6 * blur_sigma # 95% of weight + kernel_size = int(kernel_size) + kernel_size = kernel_size + 1 if kernel_size % 2 == 0 else kernel_size + + img = cv2.GaussianBlur(img, (kernel_size, kernel_size), blur_sigma) + + # Apply random noise + if random_distortion == 'noise' and random_noise: + noise_type = np.random.choice(['gaussian', 'laplace', 'poisson']) + noise_scale = (20 * np.random.random() + 20) + + if noise_type == 'gaussian': + noise = np.random.normal(scale=noise_scale, size=img.shape) + img += noise / 255.0 + elif noise_type == 'laplace': + noise = np.random.laplace(scale=noise_scale, size=img.shape) + img += noise / 255.0 + elif noise_type == 'poisson': + noise_lam = (15 * np.random.random() + 15) + noise = np.random.poisson(lam=noise_lam, size=img.shape) + img += noise / 255.0 + + # Apply random jpeg compression + if random_distortion == 'jpeg' and random_jpeg: + img = np.clip(img*255, 0, 255).astype(np.uint8) + jpeg_compression_level = np.random.randint(50, 85) + encode_param = [int(cv2.IMWRITE_JPEG_QUALITY), jpeg_compression_level] + _, enc_img = cv2.imencode('.jpg', img, encode_param) + img = cv2.imdecode(enc_img, cv2.IMREAD_UNCHANGED).astype(np.float32) / 255.0 + + # Apply random downsampling + if random_distortion == 'down' and random_downsample: + down_res = np.random.randint(int(0.125*resolution), int(0.25*resolution)) + img = cv2.resize(img, (down_res, down_res), interpolation=cv2.INTER_CUBIC) + img = cv2.resize(img, (resolution, resolution), interpolation=cv2.INTER_CUBIC) + img = imagelib.warp_by_params (params_per_resolution[resolution], img, warp, transform, can_flip=True, border_replicate=border_replicate) - img = np.clip(img.astype(np.float32), 0, 1) - - if motion_blur is not None: + img = np.clip(img.astype(np.float32), 0, 1) + + if motion_blur is not None: random_mask = sd.random_circle_faded ([resolution,resolution], rnd_state=np.random.RandomState (sample_rnd_seed+2)) if random_circle_mask else None img = imagelib.apply_random_motion_blur(img, *motion_blur, mask=random_mask,rnd_state=np.random.RandomState (sample_rnd_seed+2) ) if gaussian_blur is not None: random_mask = sd.random_circle_faded ([resolution,resolution], rnd_state=np.random.RandomState (sample_rnd_seed+3)) if random_circle_mask else None img = imagelib.apply_random_gaussian_blur(img, *gaussian_blur, mask=random_mask,rnd_state=np.random.RandomState (sample_rnd_seed+3) ) - + if random_bilinear_resize is not None: random_mask = sd.random_circle_faded ([resolution,resolution], rnd_state=np.random.RandomState (sample_rnd_seed+4)) if random_circle_mask else None img = imagelib.apply_random_bilinear_resize(img, *random_bilinear_resize, mask=random_mask,rnd_state=np.random.RandomState (sample_rnd_seed+4) ) - - - + + + # Transform from BGR to desired channel_type if channel_type == SPCT.BGR: out_sample = img + elif channel_type == SPCT.LAB_RAND_TRANSFORM: + out_sample = random_lab_rotation(img, sample_rnd_seed) elif channel_type == SPCT.G: out_sample = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)[...,None] elif channel_type == SPCT.GGG: @@ -239,22 +301,22 @@ class SampleProcessor(object): # Final transformations if nearest_resize_to is not None: out_sample = cv2_resize(out_sample, (nearest_resize_to,nearest_resize_to), interpolation=cv2.INTER_NEAREST) - + if not debug: if normalize_tanh: out_sample = np.clip (out_sample * 2.0 - 1.0, -1.0, 1.0) if data_format == "NCHW": out_sample = np.transpose(out_sample, (2,0,1) ) elif sample_type == SPST.IMAGE: - img = sample_bgr + img = sample_bgr img = imagelib.warp_by_params (params_per_resolution[resolution], img, warp, transform, can_flip=True, border_replicate=True) img = cv2.resize( img, (resolution, resolution), interpolation=cv2.INTER_CUBIC ) out_sample = img - + if data_format == "NCHW": out_sample = np.transpose(out_sample, (2,0,1) ) - - + + elif sample_type == SPST.LANDMARKS_ARRAY: l = sample_landmarks l = np.concatenate ( [ np.expand_dims(l[:,0] / w,-1), np.expand_dims(l[:,1] / h,-1) ], -1 )