diff --git a/.github/FUNDING.yml b/.github/FUNDING.yml new file mode 100644 index 0000000..39a0ccb --- /dev/null +++ b/.github/FUNDING.yml @@ -0,0 +1,12 @@ +# These are supported funding model platforms + +github: jmhummel +patreon: faceshiftlabs +open_collective: # Replace with a single Open Collective username +ko_fi: # Replace with a single Ko-fi username +tidelift: # Replace with a single Tidelift platform-name/package-name e.g., npm/babel +community_bridge: # Replace with a single Community Bridge project-name e.g., cloud-foundry +liberapay: # Replace with a single Liberapay username +issuehunt: # Replace with a single IssueHunt username +otechie: # Replace with a single Otechie username +custom: # Replace with up to 4 custom sponsorship URLs e.g., ['link1', 'link2'] diff --git a/.gitignore b/.gitignore index 16a6020..48276bc 100644 --- a/.gitignore +++ b/.gitignore @@ -5,4 +5,7 @@ !*.jpg !requirements* !Dockerfile* -!*.sh \ No newline at end of file +!*.sh +convert.py +randomColor.py +train.py 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 f232adc..5e62e13 100644 --- a/README.md +++ b/README.md @@ -1,4 +1,4 @@ - +
- +# CHANGELOG +### [View most recent changes](CHANGELOG.md) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - +Google Colab (github) +
@@ -53,133 +53,13 @@ DeepFaceLab is used by such popular youtube channels as
+
-# What can I do using DeepFaceLab? - -
- -## Replace the face - - - -
- -## De-age the face - -
- - - - - - - -
- -![](doc/youtube_icon.png) https://www.youtube.com/watch?v=Ddx5B-84ebo - -
- -## Replace the head - -
- - - - - - - -
- -![](doc/youtube_icon.png) https://www.youtube.com/watch?v=xr5FHd0AdlQ - -
- - - - - - - -
- -![](doc/youtube_icon.png) https://www.youtube.com/watch?v=RTjgkhMugVw - -
- - - - - - - -
- -![](doc/youtube_icon.png) https://www.youtube.com/watch?v=R9f7WD0gKPo - -
- -## Manipulate politicians lips -(voice replacement is not included!) -(also requires a skill in video editors such as *Adobe After Effects* or *Davinci Resolve*) - - - -![](doc/youtube_icon.png) https://www.youtube.com/watch?v=IvY-Abd2FfM - - - -![](doc/youtube_icon.png) https://www.youtube.com/watch?v=ERQlaJ_czHU - -
- -# Deepfake native resolution progress - -
- - - -
- - - -Unfortunately, there is no "make everything ok" button in DeepFaceLab. You should spend time studying the workflow and growing your skills. A skill in programs such as *AfterEffects* or *Davinci Resolve* is also desirable. - -
+
## Mini tutorial @@ -205,8 +85,8 @@ Unfortunately, there is no "make everything ok" button in DeepFaceLab. You shoul Contains new and prev releases.
-Google Colab (github) -by @chervonij . You can train fakes for free using Google Colab.
Personal fork from @chervonij repository. You can train fakes for free using Google Colab.
Linux (github) diff --git a/XSegEditor/QStringDB.py b/XSegEditor/QStringDB.py index b9100d2..5a660bb 100644 --- a/XSegEditor/QStringDB.py +++ b/XSegEditor/QStringDB.py @@ -85,6 +85,16 @@ 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.spinner_label = { 'en' : 'Step size', + 'ru' : 'Размер шага', + 'zh' : '台阶大小' + }[lang] + + QStringDB.spinner_label_tip = { 'en' : 'Minimum 5\nMaximum 500', + 'ru' : 'Минимум 5\nМаксимум 500', + 'zh' : '最少5个\n最多500' + }[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 ', diff --git a/XSegEditor/XSegEditor.py b/XSegEditor/XSegEditor.py index affc9f6..c2d4065 100644 --- a/XSegEditor/XSegEditor.py +++ b/XSegEditor/XSegEditor.py @@ -1173,6 +1173,8 @@ class MainWindow(QXMainWindow): self.cached_images = {} self.cached_has_ie_polys = {} + self.spin_box = QSpinBox() + self.initialize_ui() # Loader @@ -1297,7 +1299,7 @@ class MainWindow(QXMainWindow): def process_prev_image(self): key_mods = QApplication.keyboardModifiers() - step = 5 if key_mods == Qt.ShiftModifier else 1 + step = self.spin_box.value() if key_mods == Qt.ShiftModifier else 1 only_has_polys = key_mods == Qt.ControlModifier if self.canvas.op.is_initialized(): @@ -1323,7 +1325,7 @@ class MainWindow(QXMainWindow): def process_next_image(self, first_initialization=False): key_mods = QApplication.keyboardModifiers() - step = 0 if first_initialization else 5 if key_mods == Qt.ShiftModifier else 1 + step = 0 if first_initialization else self.spin_box.value() if key_mods == Qt.ShiftModifier else 1 only_has_polys = False if first_initialization else key_mods == Qt.ControlModifier if self.canvas.op.is_initialized(): @@ -1373,6 +1375,13 @@ class MainWindow(QXMainWindow): pad_image = QWidget() pad_image.setFixedSize(QUIConfig.preview_bar_icon_q_size) + + self.spin_box.setFocusPolicy(Qt.ClickFocus) + self.spin_box.setRange(5, 500) + self.spin_box.setSingleStep(1) + self.spin_box.installEventFilter(self) + self.spin_box.valueChanged.connect(self.on_spinbox_value_changed) + self.spin_box.setToolTip(QStringDB.spinner_label_tip) preview_image_bar_frame_l = QHBoxLayout() preview_image_bar_frame_l.setContentsMargins(0,0,0,0) @@ -1404,14 +1413,25 @@ class MainWindow(QXMainWindow): preview_image_bar.setLayout(preview_image_bar_l) label_font = QFont('Courier New') + self.filename_label = QLabel() self.filename_label.setFont(label_font) self.has_ie_polys_count_label = QLabel() + status_frame_1_2 = QHBoxLayout() + status_frame_1_2.setContentsMargins(0,0,0,0) + + step_string_label = QLabel() + step_string_label.setFont(label_font) + step_string_label.setText(QStringDB.spinner_label) + + status_frame_1_2.addWidget (step_string_label, alignment=Qt.AlignRight) + status_frame_1_2.addWidget (self.spin_box, alignment=Qt.AlignLeft) + status_frame_l = QHBoxLayout() status_frame_l.setContentsMargins(0,0,0,0) - status_frame_l.addWidget ( QLabel(), alignment=Qt.AlignCenter) + status_frame_l.addLayout (status_frame_1_2) status_frame_l.addWidget (self.filename_label, alignment=Qt.AlignCenter) status_frame_l.addWidget (self.has_ie_polys_count_label, alignment=Qt.AlignCenter) status_frame = QFrame() @@ -1438,6 +1458,21 @@ class MainWindow(QXMainWindow): else: self.move( QPoint(0,0)) + def eventFilter(self, obj, event): + if event.type() == QEvent.KeyPress and obj is self.spin_box: + if event.key() == Qt.Key_Return or event.key() == Qt.Key_Enter and self.spin_box.hasFocus(): + self.spin_box.clearFocus() + + if event.type() == QEvent.MouseButtonPress and obj is self.spin_box: + if event.button() == Qt.LeftButton and self.spin_box.hasFocus(): + self.spin_box.clearFocus() + + return super().eventFilter(obj, event) + + def on_spinbox_value_changed(self, value): + if value == self.spin_box.maximum() or value == self.spin_box.minimum(): + self.spin_box.clearFocus() + def get_has_ie_polys_count(self): return self.has_ie_polys_count diff --git a/core/imagelib/__init__.py b/core/imagelib/__init__.py index 11234a5..1ed95dc 100644 --- a/core/imagelib/__init__.py +++ b/core/imagelib/__init__.py @@ -12,7 +12,7 @@ from .warp import gen_warp_params, warp_by_params 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 +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 random_crop, normalize_channels, cut_odd_image, overlay_alpha_image diff --git a/core/imagelib/color_transfer.py b/core/imagelib/color_transfer.py index d269de2..6a19d30 100644 --- a/core/imagelib/color_transfer.py +++ b/core/imagelib/color_transfer.py @@ -1,6 +1,9 @@ import cv2 import numexpr as ne import numpy as np +from numpy import linalg as npla +import random +from scipy.stats import special_ortho_group import scipy as sp from numpy import linalg as npla @@ -9,14 +12,12 @@ def color_transfer_sot(src,trg, steps=10, batch_size=5, reg_sigmaXY=16.0, reg_si """ Color Transform via Sliced Optimal Transfer ported by @iperov from https://github.com/dcoeurjo/OTColorTransfer - src - any float range any channel image dst - any float range any channel image, same shape as src steps - number of solver steps batch_size - solver batch size reg_sigmaXY - apply regularization and sigmaXY of filter, otherwise set to 0.0 reg_sigmaV - sigmaV of filter - return value - clip it manually """ if not np.issubdtype(src.dtype, np.floating): @@ -334,3 +335,72 @@ def color_transfer(ct_mode, img_src, img_trg): else: raise ValueError(f"unknown ct_mode {ct_mode}") return out + + +# 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) + face = random_lab(face, seed) + img[:, :, :3] = face + 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) + amount_l = 30 / 100 + amount_ab = 8 / 100 + randoms = [(random.random() * amount_l * 2) - amount_l, # L adjust + (random.random() * amount_ab * 2) - amount_ab, # A adjust + (random.random() * amount_ab * 2) - amount_ab] # B adjust + image = cv2.cvtColor( # pylint:disable=no-member + image, cv2.COLOR_BGR2LAB).astype("float32") / 255.0 # pylint:disable=no-member + + for idx, adjustment in enumerate(randoms): + if adjustment >= 0: + image[:, :, idx] = ((1 - image[:, :, idx]) * adjustment) + image[:, :, idx] + else: + image[:, :, idx] = image[:, :, idx] * (1 + adjustment) + image = cv2.cvtColor((image * 255.0).astype("uint8"), # pylint:disable=no-member + cv2.COLOR_LAB2BGR) # pylint:disable=no-member + return image + +def random_clahe(image, seed=None): + """ Randomly perform Contrast Limited Adaptive Histogram Equalization """ + random.seed(seed) + contrast_random = random.random() + if contrast_random > 50 / 100: + return image + + # base_contrast = image.shape[0] // 128 + base_contrast = 1 # testing because it breaks on small sizes + grid_base = random.random() * 4 + contrast_adjustment = int(grid_base * (base_contrast / 2)) + grid_size = base_contrast + contrast_adjustment + + clahe = cv2.createCLAHE(clipLimit=2.0, # pylint: disable=no-member + tileGridSize=(grid_size, grid_size)) + for chan in range(3): + image[:, :, chan] = clahe.apply(image[:, :, chan]) + return image diff --git a/core/leras/__init__.py b/core/leras/__init__.py index 7d9fb2b..735a81e 100644 --- a/core/leras/__init__.py +++ b/core/leras/__init__.py @@ -1 +1 @@ -from .nn import nn \ No newline at end of file +from .nn import nn 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 4accaf6..c9c3077 100644 --- a/core/leras/layers/__init__.py +++ b/core/leras/layers/__init__.py @@ -15,4 +15,6 @@ from .TLU import * from .ScaleAdd import * from .DenseNorm import * from .AdaIN import * -from .TanhPolar import * \ No newline at end of file +from .TanhPolar import * +from .MsSsim import * +from .TanhPolar import * diff --git a/core/leras/models/PatchDiscriminator.py b/core/leras/models/PatchDiscriminator.py index 9b94e9f..5bad9e4 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) @@ -133,7 +133,6 @@ class UNetPatchDiscriminator(nn.ModelBase): def on_build(self, patch_size, in_ch, base_ch = 16, use_fp16 = False): self.use_fp16 = use_fp16 conv_dtype = tf.float16 if use_fp16 else tf.float32 - class ResidualBlock(nn.ModelBase): def on_build(self, ch, kernel_size=3 ): self.conv1 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME', dtype=conv_dtype) @@ -150,7 +149,7 @@ class UNetPatchDiscriminator(nn.ModelBase): self.convs = [] self.upconvs = [] 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', dtype=conv_dtype) @@ -158,8 +157,14 @@ class UNetPatchDiscriminator(nn.ModelBase): 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', dtype=conv_dtype) ) + 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', dtype=conv_dtype) ) + 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', dtype=conv_dtype) self.center_out = nn.Conv2D( level_chs[len(layers)-1], 1, kernel_size=1, padding='VALID', dtype=conv_dtype) @@ -169,14 +174,15 @@ class UNetPatchDiscriminator(nn.ModelBase): def forward(self, x): if self.use_fp16: x = tf.cast(x, tf.float16) - x = tf.nn.leaky_relu( self.in_conv(x), 0.2 ) encs = [] for conv in self.convs: encs.insert(0, x) 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) in enumerate(zip(self.upconvs, encs)): @@ -192,3 +198,129 @@ class UNetPatchDiscriminator(nn.ModelBase): return center_out, 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, use_fp16 = False): + self.use_fp16 = use_fp16 + conv_dtype = tf.float16 if use_fp16 else tf.float32 + + class ResidualBlock(nn.ModelBase): + def on_build(self, ch, kernel_size=3 ): + self.conv1 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME', dtype=conv_dtype) + self.conv2 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME', dtype=conv_dtype) + + 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', dtype=conv_dtype) + + 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', dtype=conv_dtype) ) + + 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', dtype=conv_dtype) ) + + 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', dtype=conv_dtype) + + self.center_out = nn.Conv2D( level_chs[len(layers)-1], 1, kernel_size=1, padding='VALID', dtype=conv_dtype) + self.center_conv = nn.Conv2D( level_chs[len(layers)-1], level_chs[len(layers)-1], kernel_size=1, padding='VALID', dtype=conv_dtype) + + + def forward(self, x): + if self.use_fp16: + x = tf.cast(x, tf.float16) + + 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) + + x = self.out_conv(x) + + if self.use_fp16: + center_out = tf.cast(center_out, tf.float32) + x = tf.cast(x, tf.float32) + + return center_out, x + +nn.UNetPatchDiscriminatorV2 = UNetPatchDiscriminatorV2 diff --git a/core/leras/nn.py b/core/leras/nn.py index f392aaf..7c28874 100644 --- a/core/leras/nn.py +++ b/core/leras/nn.py @@ -40,7 +40,7 @@ class nn(): conv2d_spatial_axes = None floatx = None - + @staticmethod def initialize(device_config=None, floatx="float32", data_format="NHWC"): @@ -87,7 +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 @@ -99,7 +99,7 @@ class nn(): import core.leras.optimizers import core.leras.models import core.leras.archis - + # Configure tensorflow session-config if len(device_config.devices) == 0: config = tf.ConfigProto(device_count={'GPU': 0}) @@ -107,13 +107,13 @@ class nn(): else: nn.tf_default_device_name = f'/{device_config.devices[0].tf_dev_type}:0' - config = tf.ConfigProto() + 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) @@ -260,7 +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 bd690da..aeb077f 100644 --- a/core/leras/ops/__init__.py +++ b/core/leras/ops/__init__.py @@ -244,6 +244,19 @@ 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 ] @@ -475,4 +488,3 @@ def bilinear_sampler(img, x, y): return out nn.bilinear_sampler = bilinear_sampler - 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/facelib/FaceType.py b/facelib/FaceType.py index 745cff3..edb40ff 100644 --- a/facelib/FaceType.py +++ b/facelib/FaceType.py @@ -7,6 +7,7 @@ class FaceType(IntEnum): FULL = 2 FULL_NO_ALIGN = 3 WHOLE_FACE = 4 + CUSTOM = 5 HEAD = 10 HEAD_NO_ALIGN = 20 @@ -30,7 +31,8 @@ to_string_dict = { FaceType.HALF : 'half_face', FaceType.WHOLE_FACE : 'whole_face', FaceType.HEAD : 'head', FaceType.HEAD_NO_ALIGN : 'head_no_align', - + FaceType.CUSTOM : 'mve_custom', + FaceType.MARK_ONLY :'mark_only', } diff --git a/facelib/LandmarksProcessor.py b/facelib/LandmarksProcessor.py index 8e5d51b..3db7419 100644 --- a/facelib/LandmarksProcessor.py +++ b/facelib/LandmarksProcessor.py @@ -382,11 +382,9 @@ def expand_eyebrows(lmrks, eyebrows_expand_mod=1.0): # Adjust eyebrow arrays lmrks[17:22] = top_l + eyebrows_expand_mod * 0.5 * (top_l - bot_l) lmrks[22:27] = top_r + eyebrows_expand_mod * 0.5 * (top_r - bot_r) + return lmrks - - - def get_image_hull_mask (image_shape, image_landmarks, eyebrows_expand_mod=1.0 ): hull_mask = np.zeros(image_shape[0:2]+(1,),dtype=np.float32) @@ -441,7 +439,7 @@ def get_image_mouth_mask (image_shape, image_landmarks): image_landmarks = image_landmarks.astype(np.int) - cv2.fillConvexPoly( hull_mask, cv2.convexHull( image_landmarks[60:]), (1,) ) + cv2.fillConvexPoly( hull_mask, cv2.convexHull( image_landmarks[48:60]), (1,) ) dilate = h // 32 hull_mask = cv2.dilate(hull_mask, cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(dilate,dilate)), iterations = 1 ) 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 493c7ed..82e53b1 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.") @@ -127,6 +127,8 @@ if __name__ == "__main__": 'silent_start' : arguments.silent_start, 'execute_programs' : [ [int(x[0]), x[1] ] for x in arguments.execute_program ], 'debug' : arguments.debug, + 'dump_ckpt' : arguments.dump_ckpt, + 'flask_preview' : arguments.flask_preview, } from mainscripts import Trainer Trainer.main(**kwargs) @@ -144,7 +146,10 @@ if __name__ == "__main__": p.add_argument('--cpu-only', action="store_true", dest="cpu_only", default=False, help="Train on CPU.") p.add_argument('--force-gpu-idxs', dest="force_gpu_idxs", default=None, help="Force to choose GPU indexes separated by comma.") p.add_argument('--silent-start', action="store_true", dest="silent_start", default=False, help="Silent start. Automatically chooses Best GPU and last used model.") - + 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) @@ -158,6 +163,16 @@ if __name__ == "__main__": p.add_argument('--model', required=True, dest="model_name", choices=pathex.get_all_dir_names_startswith ( Path(__file__).parent / 'models' , 'Model_'), help="Model class name.") p.set_defaults (func=process_exportdfm) + def process_exportdfm(arguments): + osex.set_process_lowest_prio() + from mainscripts import ExportDFM + ExportDFM.main(model_class_name = arguments.model_name, saved_models_path = Path(arguments.model_dir)) + + p = subparsers.add_parser( "exportdfm", help="Export model to use in DeepFaceLive.") + p.add_argument('--model-dir', required=True, action=fixPathAction, dest="model_dir", help="Saved models dir.") + p.add_argument('--model', required=True, dest="model_name", choices=pathex.get_all_dir_names_startswith ( Path(__file__).parent / 'models' , 'Model_'), help="Model class name.") + p.set_defaults (func=process_exportdfm) + def process_merge(arguments): osex.set_process_lowest_prio() from mainscripts import Merger @@ -280,10 +295,10 @@ 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): @@ -291,11 +306,11 @@ 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): @@ -305,8 +320,8 @@ 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() @@ -314,8 +329,8 @@ 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() @@ -323,8 +338,8 @@ 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): @@ -333,7 +348,7 @@ 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) @@ -344,9 +359,9 @@ if __name__ == "__main__": if exit_code == 0: print ("Done.") - + exit(exit_code) - + ''' import code code.interact(local=dict(globals(), **locals())) diff --git a/mainscripts/Trainer.py b/mainscripts/Trainer.py index df74ca3..3e34fe1 100644 --- a/mainscripts/Trainer.py +++ b/mainscripts/Trainer.py @@ -1,9 +1,11 @@ -import os +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 @@ -14,6 +16,7 @@ import models from core.interact import interact as io def trainerThread (s2c, c2s, e, + socketio=None, model_class_name = None, saved_models_path = None, training_data_src_path = None, @@ -59,51 +62,54 @@ def trainerThread (s2c, c2s, e, is_reached_goal = model.is_reached_iter_goal() - 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() + 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() - 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 @@ -111,18 +117,20 @@ 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("!!!") @@ -131,19 +139,19 @@ 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: @@ -151,19 +159,22 @@ 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) if model.get_iter() == 1: 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.') + need_save = False while time.time() - last_save_time >= save_interval_min*60: last_save_time += save_interval_min*60 @@ -173,7 +184,7 @@ 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() @@ -182,8 +193,8 @@ 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': @@ -200,43 +211,227 @@ 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'} ) + 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: if not c2s.empty(): - input = c2s.get() - op = input.get('op','') + item = c2s.get() + op = item.get('op', '') if op == 'close': break 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) @@ -252,33 +447,33 @@ 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 == 'close': @@ -289,22 +484,22 @@ 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 @@ -315,27 +510,28 @@ 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 @@ -355,6 +551,6 @@ 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/models/ModelBase.py b/models/ModelBase.py index f446efa..634e797 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 @@ -182,13 +183,15 @@ 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.retraining_samples = self.options.get('retraining_samples', False) self.on_initialize() self.options['batch_size'] = self.batch_size @@ -280,13 +283,21 @@ 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(): @@ -320,6 +331,10 @@ class ModelBase(object): self.options['batch_size'] = self.batch_size = batch_size + def ask_retraining_samples(self, default_value=False): + default_retraining_samples = self.load_or_def_option('retraining_samples', default_value) + self.options['retraining_samples'] = io.input_bool("Retrain high loss samples", default_retraining_samples, help_message="Periodically retrains last 16 \"high-loss\" sample") + #overridable def on_initialize_options(self): @@ -382,6 +397,9 @@ class ModelBase(object): def get_history_previews(self): return self.onGetPreview (self.sample_for_preview, for_history=True) + def get_history_previews(self): + return self.onGetPreview (self.sample_for_preview, for_history=True) + def get_preview_history_writer(self): if self.preview_history_writer is None: self.preview_history_writer = PreviewHistoryWriter() @@ -417,33 +435,32 @@ 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 = [] @@ -547,7 +564,7 @@ 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" @@ -586,10 +603,9 @@ 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 index 82b0dc5..23050e9 100644 --- a/models/Model_AMP/Model.py +++ b/models/Model_AMP/Model.py @@ -4,7 +4,6 @@ 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 @@ -16,6 +15,8 @@ class AMPModel(ModelBase): #override def on_initialize_options(self): + default_retraining_samples = self.options['retraining_samples'] = self.load_or_def_option('retraining_samples', False) + # default_usefp16 = self.options['use_fp16'] = self.load_or_def_option('use_fp16', False) 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) @@ -27,11 +28,28 @@ class AMPModel(ModelBase): 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', 0.5) + default_eyes_mouth_prio = self.options['eyes_mouth_prio'] = self.load_or_def_option('eyes_mouth_prio', False) default_uniform_yaw = self.options['uniform_yaw'] = self.load_or_def_option('uniform_yaw', False) + + # Uncomment it just if you want to impelement other loss functions + #default_loss_function = self.options['loss_function'] = self.load_or_def_option('loss_function', 'SSIM') + default_blur_out_mask = self.options['blur_out_mask'] = self.load_or_def_option('blur_out_mask', False) + + default_adabelief = self.options['adabelief'] = self.load_or_def_option('adabelief', True) + default_lr_dropout = self.options['lr_dropout'] = self.load_or_def_option('lr_dropout', 'n') + 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) + + # Uncomment it just if you want to impelement other loss functions + #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) ask_override = self.ask_override() @@ -39,9 +57,12 @@ class AMPModel(ModelBase): self.ask_autobackup_hour() self.ask_write_preview_history() self.ask_target_iter() + self.ask_retraining_samples() self.ask_random_src_flip() self.ask_random_dst_flip() self.ask_batch_size(8) + # self.options['use_fp16'] = io.input_bool ("Use fp16", default_usefp16, help_message='Increases training/inference speed, reduces model size. Model may crash. Enable it after 1-5k iters.') + 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 .") @@ -73,8 +94,11 @@ class AMPModel(ModelBase): self.options['morph_factor'] = morph_factor if self.is_first_run() or ask_override: + 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.') + self.options['blur_out_mask'] = io.input_bool ("Blur out mask", default_blur_out_mask, help_message='Blurs nearby area outside of applied face mask of training samples. The result is the background near the face is smoothed and less noticeable on swapped face. The exact xseg mask in src and dst faceset is required.') + 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.") default_gan_power = self.options['gan_power'] = self.load_or_def_option('gan_power', 0.0) @@ -84,7 +108,13 @@ class AMPModel(ModelBase): 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['adabelief'] = io.input_bool ("Use AdaBelief optimizer?", default_adabelief, help_message="Use AdaBelief optimizer. It requires more VRAM, but the accuracy and the generalization of the model is higher.") + 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 .. 5.0", help_message="Forces the neural network to learn small details of the face. Enable it only when the face is trained enough with 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, 5.0 ) @@ -95,7 +125,11 @@ class AMPModel(ModelBase): gan_dims = np.clip ( io.input_int("GAN dimensions", default_gan_dims, add_info="4-512", 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, 512 ) self.options['gan_dims'] = gan_dims - self.options['ct_mode'] = io.input_str (f"Color transfer for src faceset", default_ct_mode, ['none','rct','lct','mkl','idt','sot'], help_message="Change color distribution of src samples close to dst samples. If src faceset is deverse enough, then lct mode is fine in most cases.") + #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.gan_model_changed = (default_gan_patch_size != self.options['gan_patch_size']) or (default_gan_dims != self.options['gan_dims']) @@ -122,6 +156,8 @@ class AMPModel(ModelBase): morph_factor = self.options['morph_factor'] gan_power = self.gan_power = self.options['gan_power'] random_warp = self.options['random_warp'] + + eyes_mouth_prio = self.options['eyes_mouth_prio'] blur_out_mask = self.options['blur_out_mask'] @@ -129,7 +165,9 @@ class AMPModel(ModelBase): if ct_mode == 'none': ct_mode = None - use_fp16 = False + adabelief = self.options['adabelief'] + + # use_fp16 = self.options['use_fp16'] if self.is_exporting: use_fp16 = io.input_bool ("Export quantized?", False, help_message='Makes the exported model faster. If you have problems, disable this option.') @@ -300,13 +338,15 @@ class AMPModel(ModelBase): lr_dropout = 1.0 self.G_weights = self.encoder.get_weights() + self.decoder.get_weights() - self.src_dst_opt = nn.AdaBelief(lr=5e-5, lr_dropout=lr_dropout, lr_cos=lr_cos, clipnorm=clipnorm, name='src_dst_opt') + OptimizerClass = nn.AdaBelief if adabelief else nn.RMSprop + + self.src_dst_opt = OptimizerClass(lr=5e-5, lr_dropout=lr_dropout, lr_cos=lr_cos, clipnorm=clipnorm, name='src_dst_opt') self.src_dst_opt.initialize_variables (self.G_weights, vars_on_cpu=optimizer_vars_on_cpu) self.model_filename_list += [ (self.src_dst_opt, 'src_dst_opt.npy') ] 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.GAN_opt = nn.AdaBelief(lr=5e-5, lr_dropout=lr_dropout, lr_cos=lr_cos, clipnorm=clipnorm, name='GAN_opt') + self.GAN = nn.UNetPatchDiscriminator(patch_size=self.options['gan_patch_size'], in_ch=input_ch, base_ch=self.options['gan_dims'], use_fp16=use_fp16, name="GAN") + self.GAN_opt = OptimizerClass(lr=5e-5, lr_dropout=lr_dropout, lr_cos=lr_cos, clipnorm=clipnorm, name='GAN_opt') self.GAN_opt.initialize_variables ( self.GAN.get_weights(), vars_on_cpu=optimizer_vars_on_cpu) self.model_filename_list += [ [self.GAN, 'GAN.npy'], [self.GAN_opt, 'GAN_opt.npy'] ] @@ -397,7 +437,7 @@ class AMPModel(ModelBase): y = 1-nn.gaussian_blur(gpu_target_srcm, sigma) y = tf.where(tf.equal(y, 0), tf.ones_like(y), y) gpu_target_src = gpu_target_src*gpu_target_srcm + (x/y)*gpu_target_srcm_anti - + x = nn.gaussian_blur(gpu_target_dst*gpu_target_dstm_anti, sigma) y = 1-nn.gaussian_blur(gpu_target_dstm, sigma) y = tf.where(tf.equal(y, 0), tf.ones_like(y), y) @@ -424,8 +464,9 @@ class AMPModel(ModelBase): gpu_dst_loss += tf.reduce_mean (10*tf.square(gpu_target_dst_masked-gpu_pred_dst_dst_masked), axis=[1,2,3]) # Eyes+mouth prio loss - 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_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]) + 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_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]) # Mask loss gpu_src_loss += tf.reduce_mean ( 10*tf.square( gpu_target_srcm - gpu_pred_src_srcm ),axis=[1,2,3] ) @@ -558,30 +599,52 @@ class AMPModel(ModelBase): 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(scale_range=[-0.15, 0.15], random_flip=self.random_src_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':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':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':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.EYES_MOUTH, 'face_type':face_type, 'data_format':nn.data_format, 'resolution': resolution}, + sample_process_options=SampleProcessor.Options(scale_range=[-0.125, 0.125], random_flip=self.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, + 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(scale_range=[-0.15, 0.15], random_flip=self.random_dst_flip), - output_sample_types = [ {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':random_warp, 'transform':True, 'channel_type' : SampleProcessor.ChannelType.BGR, 'face_type':face_type, 'data_format':nn.data_format, 'resolution': resolution}, - {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':False , 'transform':True, 'channel_type' : SampleProcessor.ChannelType.BGR, 'face_type':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':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.EYES_MOUTH, 'face_type':face_type, 'data_format':nn.data_format, 'resolution': resolution}, + sample_process_options=SampleProcessor.Options(scale_range=[-0.125, 0.125], random_flip=self.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, + uniform_yaw_distribution=self.options['uniform_yaw'], #or self.pretrain, generators_count=dst_generators_count ) ]) + if self.options['retraining_samples']: + self.last_src_samples_loss = [] + self.last_dst_samples_loss = [] + def export_dfm (self): output_path=self.get_strpath_storage_for_file('model.dfm') @@ -651,6 +714,27 @@ class AMPModel(ModelBase): src_loss, dst_loss = self.train (warped_src, target_src, target_srcm, target_srcm_em, warped_dst, target_dst, target_dstm, target_dstm_em) + if self.options['retraining_samples']: + for i in range(bs): + self.last_src_samples_loss.append ( (src_loss[i], target_src[i], target_srcm[i], target_srcm_em[i]) ) + self.last_dst_samples_loss.append ( (dst_loss[i], target_dst[i], target_dstm[i], target_dstm_em[i]) ) + + if len(self.last_src_samples_loss) >= bs*16: + src_samples_loss = sorted(self.last_src_samples_loss, key=operator.itemgetter(0), reverse=True) + dst_samples_loss = sorted(self.last_dst_samples_loss, key=operator.itemgetter(0), reverse=True) + + target_src = np.stack( [ x[1] for x in src_samples_loss[:bs] ] ) + target_srcm = np.stack( [ x[2] for x in src_samples_loss[:bs] ] ) + target_srcm_em = np.stack( [ x[3] for x in src_samples_loss[:bs] ] ) + + target_dst = np.stack( [ x[1] for x in dst_samples_loss[:bs] ] ) + target_dstm = np.stack( [ x[2] for x in dst_samples_loss[:bs] ] ) + target_dstm_em = np.stack( [ x[3] for x in dst_samples_loss[:bs] ] ) + + src_loss, dst_loss = self.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.GAN_train (warped_src, target_src, target_srcm, target_srcm_em, warped_dst, target_dst, target_dstm, target_dstm_em) diff --git a/models/Model_SAEHD/Model.py b/models/Model_SAEHD/Model.py index ecfaa73..948b7a1 100644 --- a/models/Model_SAEHD/Model.py +++ b/models/Model_SAEHD/Model.py @@ -1,6 +1,5 @@ import multiprocessing import operator -from functools import partial import numpy as np @@ -26,7 +25,6 @@ class SAEHDModel(ModelBase): else: suggest_batch_size = 4 - yn_str = {True:'y',False:'n'} min_res = 64 max_res = 640 @@ -42,7 +40,8 @@ class SAEHDModel(ModelBase): 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_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', False) + default_eyes_prio = self.options['eyes_prio'] = self.load_or_def_option('eyes_prio', False) + default_mouth_prio = self.options['mouth_prio'] = self.load_or_def_option('mouth_prio', False) default_uniform_yaw = self.options['uniform_yaw'] = self.load_or_def_option('uniform_yaw', False) default_blur_out_mask = self.options['blur_out_mask'] = self.load_or_def_option('blur_out_mask', False) @@ -52,20 +51,32 @@ 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_hsv_power = self.options['random_hsv_power'] = self.load_or_def_option('random_hsv_power', 0.0) + 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_retraining_samples() self.ask_random_src_flip() self.ask_random_dst_flip() self.ask_batch_size(suggest_batch_size) @@ -75,10 +86,7 @@ class SAEHDModel(ModelBase): 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 16 and 32 for -d archi.") 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'], help_message="Half / mid face / full face / whole face / head. 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() + 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=\ @@ -133,16 +141,21 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... self.options['d_mask_dims'] = d_mask_dims + d_mask_dims % 2 if self.is_first_run() or ask_override: - if self.options['face_type'] == 'wf' or self.options['face_type'] == 'head': + if self.options['face_type'] == 'wf' or self.options['face_type'] == 'head' or self.options['face_type'] == 'custom': 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['eyes_prio'] = io.input_bool ("Eyes priority", default_eyes_prio, help_message='Helps to fix eye problems during training like "alien eyes" and wrong eyes direction ( especially on HD architectures ) by forcing the neural network to train eyes with higher priority. before/after https://i.imgur.com/YQHOuSR.jpg ') + 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.') self.options['blur_out_mask'] = io.input_bool ("Blur out mask", default_blur_out_mask, help_message='Blurs nearby area outside of applied face mask of training samples. The result is the background near the face is smoothed and less noticeable on swapped face. The exact xseg mask in src and dst faceset is required.') + 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.") @@ -151,28 +164,48 @@ 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['random_hsv_power'] = np.clip ( io.input_number ("Random hue/saturation/light intensity", default_random_hsv_power, add_info="0.0 .. 0.3", help_message="Random hue/saturation/light intensity applied to the src face set only at the input of the neural network. Stabilizes color perturbations during face swapping. Reduces the quality of the color transfer by selecting the closest one in the src faceset. Thus the src faceset must be diverse enough. Typical fine value is 0.05"), 0.0, 0.3 ) - self.options['gan_power'] = np.clip ( io.input_number ("GAN power", default_gan_power, add_info="0.0 .. 5.0", 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, 5.0 ) + 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: - 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 + 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="4-512", 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, 512 ) - self.options['gan_dims'] = gan_dims + 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'], help_message="Change color distribution of src samples close to dst samples. Try all modes to find the best.") + 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. Forces random_warp=N, random_flips=Y, gan_power=0.0, lr_dropout=N, styles=0.0, uniform_yaw=Y") @@ -197,12 +230,11 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... 'mf' : FaceType.MID_FULL, 'f' : FaceType.FULL, 'wf' : FaceType.WHOLE_FACE, + 'custom' : FaceType.CUSTOM, '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'] + eyes_prio = self.options['eyes_prio'] + mouth_prio = self.options['mouth_prio'] archi_split = self.options['archi'].split('-') @@ -222,6 +254,10 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... self.set_iter(0) adabelief = self.options['adabelief'] + + use_fp16 = self.options['use_fp16'] + if self.is_exporting: + use_fp16 = io.input_bool ("Export quantized?", False, help_message='Makes the exported model faster. If you have problems, disable this option.') use_fp16 = False if self.is_exporting: @@ -233,7 +269,7 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... random_dst_flip = self.random_dst_flip if not self.pretrain else True random_hsv_power = self.options['random_hsv_power'] if not self.pretrain else 0.0 blur_out_mask = self.options['blur_out_mask'] - + if self.pretrain: self.options_show_override['lr_dropout'] = 'n' self.options_show_override['random_warp'] = False @@ -313,8 +349,12 @@ 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", use_fp16=self.options['use_fp16']) + 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'], use_fp16=self.options['use_fp16'], name="D_src") + self.model_filename_list += [ [self.D_src, 'GAN.npy'] ] # Initialize optimizers lr=5e-5 @@ -347,9 +387,14 @@ 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, lr_cos=lr_cos, 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, lr_cos=lr_cos, 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, lr_cos=lr_cos, 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 @@ -380,10 +425,26 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... 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,:,:,:] + gpu_target_srcm_all = self.target_srcm[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_srcm_anti = 1-gpu_target_srcm_all + gpu_target_dstm_anti = 1-gpu_target_dstm_all + + if blur_out_mask: + sigma = resolution / 128 + + x = nn.gaussian_blur(gpu_target_src*gpu_target_srcm_anti, sigma) + y = 1-nn.gaussian_blur(gpu_target_srcm_all, sigma) + y = tf.where(tf.equal(y, 0), tf.ones_like(y), y) + gpu_target_src = gpu_target_src*gpu_target_srcm_all + (x/y)*gpu_target_srcm_anti + + x = nn.gaussian_blur(gpu_target_dst*gpu_target_dstm_anti, sigma) + y = 1-nn.gaussian_blur(gpu_target_dstm_all, sigma) + y = tf.where(tf.equal(y, 0), tf.ones_like(y), y) + gpu_target_dst = gpu_target_dst*gpu_target_dstm_all + (x/y)*gpu_target_dstm_anti gpu_target_srcm_anti = 1-gpu_target_srcm gpu_target_dstm_anti = 1-gpu_target_dstm @@ -434,6 +495,16 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... gpu_pred_dst_dstm_list.append(gpu_pred_dst_dstm) gpu_pred_src_dstm_list.append(gpu_pred_src_dstm) + # unpack masks from one combined mask + gpu_target_srcm = tf.clip_by_value (gpu_target_srcm_all, 0, 1) + gpu_target_dstm = tf.clip_by_value (gpu_target_dstm_all, 0, 1) + gpu_target_srcm_eye_mouth = tf.clip_by_value (gpu_target_srcm_em-1, 0, 1) + gpu_target_dstm_eye_mouth = tf.clip_by_value (gpu_target_dstm_em-1, 0, 1) + gpu_target_srcm_mouth = tf.clip_by_value (gpu_target_srcm_em-2, 0, 1) + gpu_target_dstm_mouth = tf.clip_by_value (gpu_target_dstm_em-2, 0, 1) + gpu_target_srcm_eyes = tf.clip_by_value (gpu_target_srcm_eye_mouth-gpu_target_srcm_mouth, 0, 1) + gpu_target_dstm_eyes = tf.clip_by_value (gpu_target_dstm_eye_mouth-gpu_target_dstm_mouth, 0, 1) + 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_srcm_anti_blur = 1.0-gpu_target_srcm_blur @@ -455,18 +526,47 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... gpu_pred_src_src_masked_opt = gpu_pred_src_src*gpu_target_srcm_blur if masked_training else gpu_pred_src_src gpu_pred_dst_dst_masked_opt = gpu_pred_dst_dst*gpu_target_dstm_blur if masked_training else gpu_pred_dst_dst - 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_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]) + if eyes_prio or mouth_prio: + if eyes_prio and mouth_prio: + gpu_target_part_mask = gpu_target_srcm_eye_mouth + elif eyes_prio: + gpu_target_part_mask = gpu_target_srcm_eyes + elif mouth_prio: + gpu_target_part_mask = gpu_target_srcm_mouth + + gpu_src_loss += tf.reduce_mean ( 300*tf.abs ( gpu_target_src*gpu_target_part_mask - gpu_pred_src_src*gpu_target_part_mask ), 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]) + 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_pred_src_dst_no_code_grad*tf.stop_gradient(gpu_pred_src_dstm), tf.stop_gradient(gpu_pred_dst_dst*gpu_pred_dst_dstm), gaussian_blur_radius=resolution//8, loss_weight=10000*face_style_power) @@ -479,15 +579,44 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... gpu_src_loss += tf.reduce_mean( (10*bg_style_power)*nn.dssim( gpu_psd_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_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_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]) + if eyes_prio or mouth_prio: + if eyes_prio and mouth_prio: + gpu_target_part_mask = gpu_target_dstm_eye_mouth + elif eyes_prio: + gpu_target_part_mask = gpu_target_dstm_eyes + elif mouth_prio: + gpu_target_part_mask = gpu_target_dstm_mouth + + 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] ) @@ -517,22 +646,37 @@ 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() @@ -544,9 +688,7 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... 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 )] - - + gpu_G_loss_gvs += [ nn.gradients ( gpu_G_loss, self.src_dst_trainable_weights ) ] # Average losses and gradients, and create optimizer update ops @@ -606,7 +748,7 @@ 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 @@ -672,28 +814,50 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... 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(scale_range=[-0.15, 0.15], random_flip=random_src_flip), - output_sample_types = [ {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':random_warp, 'transform':True, 'channel_type' : SampleProcessor.ChannelType.BGR, 'ct_mode': ct_mode, 'random_hsv_shift_amount' : random_hsv_power, '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}, + 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.EYES_MOUTH, '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(scale_range=[-0.15, 0.15], random_flip=random_dst_flip), - output_sample_types = [ {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':random_warp, 'transform':True, 'channel_type' : SampleProcessor.ChannelType.BGR, '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, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution}, + 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.EYES_MOUTH, '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 ) ]) + if self.options['retraining_samples']: + self.last_src_samples_loss = [] + self.last_dst_samples_loss = [] + if self.pretrain_just_disabled: self.update_sample_for_preview(force_new=True) @@ -773,6 +937,29 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... 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) + if self.options['retraining_samples']: + bs = self.get_batch_size() + + 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.options['true_face_power'] != 0 and not self.pretrain: self.D_train (warped_src, warped_dst) @@ -780,14 +967,14 @@ Examples: df, liae, df-d, df-ud, liae-ud, ... 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, for_history=False): ( (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] )] @@ -802,12 +989,17 @@ 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 )), ] @@ -832,10 +1024,27 @@ 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/requirements-colab.txt b/requirements-colab.txt index 33546b2..1d0f166 100644 --- a/requirements-colab.txt +++ b/requirements-colab.txt @@ -8,4 +8,4 @@ scikit-image==0.14.2 scipy==1.4.1 colorama tensorflow-gpu==2.4.0 -tf2onnx==1.9.3 \ No newline at end of file +tf2onnx==1.9.3 diff --git a/requirements-cuda.txt b/requirements-cuda.txt index b70520d..41bcc29 100644 --- a/requirements-cuda.txt +++ b/requirements-cuda.txt @@ -9,4 +9,6 @@ scipy==1.4.1 colorama tensorflow-gpu==2.4.0 pyqt5 -tf2onnx==1.9.3 \ No newline at end of file +tf2onnx==1.9.3 +Flask==1.1.1 +flask-socketio==4.2.1 diff --git a/samplelib/SampleProcessor.py b/samplelib/SampleProcessor.py index 7432e75..db8bb0a 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 @@ -26,15 +27,17 @@ class SampleProcessor(object): BGR = 1 #BGR G = 2 #Grayscale GGG = 3 #3xGrayscale + LAB_RAND_TRANSFORM = 4 # LAB random transform + class FaceMaskType(IntEnum): - NONE = 0 + NONE = 0 FULL_FACE = 1 # mask all hull as grayscale EYES = 2 # mask eyes hull as grayscale - EYES_MOUTH = 3 # eyes and mouse + FULL_FACE_EYES = 3 # eyes and mouse class Options(object): - def __init__(self, random_flip = True, rotation_range=[-10,10], scale_range=[-0.05, 0.05], tx_range=[-0.05, 0.05], ty_range=[-0.05, 0.05] ): + def __init__(self, random_flip = True, rotation_range=[-2,2], scale_range=[-0.05, 0.05], tx_range=[-0.05, 0.05], ty_range=[-0.05, 0.05] ): self.random_flip = random_flip self.rotation_range = rotation_range self.scale_range = scale_range @@ -57,28 +60,32 @@ 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) - return np.clip(eyes_mask, 0, 1) - - def get_eyes_mouth_mask(): - eyes_mask = LandmarksProcessor.get_image_eye_mask (sample_bgr.shape, sample_landmarks) + # set eye masks to 1-2 + clip = np.clip(eyes_mask, 0, 1) + clip[clip > 0.1] += 1 + return clip + + def get_mouth_mask(): mouth_mask = LandmarksProcessor.get_image_mouth_mask (sample_bgr.shape, sample_landmarks) - mask = eyes_mask + mouth_mask - return np.clip(mask, 0, 1) - + # set eye masks to 2-3 + clip = np.clip(mouth_mask, 0, 1) + clip[clip > 0.1] += 2 + return clip + is_face_sample = sample_landmarks is not None if debug and is_face_sample: @@ -93,16 +100,20 @@ class SampleProcessor(object): warp = opts.get('warp', False) transform = opts.get('transform', False) random_hsv_shift_amount = opts.get('random_hsv_shift_amount', 0) + 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) normalize_tanh = opts.get('normalize_tanh', False) ct_mode = opts.get('ct_mode', None) data_format = opts.get('data_format', 'NHWC') - + rnd_seed_shift = opts.get('rnd_seed_shift', 0) warp_rnd_seed_shift = opts.get('warp_rnd_seed_shift', rnd_seed_shift) - + rnd_state = np.random.RandomState (sample_rnd_seed+rnd_seed_shift) warp_rnd_state = np.random.RandomState (sample_rnd_seed+warp_rnd_seed_shift) - + warp_params = imagelib.gen_warp_params(resolution, sample_process_options.random_flip, rotation_range=sample_process_options.rotation_range, @@ -112,37 +123,43 @@ class SampleProcessor(object): rnd_state=rnd_state, warp_rnd_state=warp_rnd_state, ) - - 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: img = get_eyes_mask() - elif face_mask_type == SPFMT.EYES_MOUTH: - mask = get_full_face_mask().copy() - mask[mask != 0.0] = 1.0 - img = get_eyes_mouth_mask()*mask + elif face_mask_type == SPFMT.FULL_FACE_EYES: + # sets both eyes and mouth mask parts + img = get_full_face_mask() + mask = img.copy() + 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) else: img = np.zeros ( sample_bgr.shape[0:2]+(1,), dtype=np.float32) @@ -150,48 +167,97 @@ class SampleProcessor(object): raise NotImplementedError() 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 (warp_params, 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 (warp_params, img, warp, transform, can_flip=True, border_replicate=border_replicate, cv2_inter=cv2.INTER_LINEAR) - if face_mask_type == SPFMT.EYES_MOUTH: - div = img.max() - if div != 0.0: - img = img / div # normalize to 1.0 after warp - 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 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 - if ct_mode is not None and ct_sample is not None: - if ct_sample_bgr is None: - 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 ) ) - + + # 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) + else: + if ct_sample_bgr is None: + 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) + if random_hsv_shift_amount != 0: a = random_hsv_shift_amount h_amount = max(1, int(360*a*0.5)) @@ -202,12 +268,13 @@ class SampleProcessor(object): img = np.clip( cv2.cvtColor(cv2.merge([img_h, img_s, img_v]), cv2.COLOR_HSV2BGR) , 0, 1 ) img = imagelib.warp_by_params (warp_params, img, warp, transform, can_flip=True, border_replicate=border_replicate) - img = np.clip(img.astype(np.float32), 0, 1) # 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: @@ -216,22 +283,22 @@ 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 (warp_params, 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 ) @@ -255,4 +322,3 @@ class SampleProcessor(object): outputs += [outputs_sample] return outputs -