mirror of
https://github.com/iperov/DeepFaceLab.git
synced 2025-08-21 14:03:19 -07:00
Merge pull request #126 from faceshiftlabs/feature/background-power
Feature/background power
This commit is contained in:
Jeremy Hummel
GitHub
commit
7f64e9dc3f
5 changed files with 62 additions and 1 deletions
|
|
@ -10,6 +10,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
|
|||
- [Freezeable layers (encoder/decoder/etc.)](https://github.com/faceshiftlabs/DeepFaceLab/tree/feature/freezable-weights)
|
||||
- [GAN stability improvements](https://github.com/faceshiftlabs/DeepFaceLab/tree/feature/gan-updates)
|
||||
|
||||
## [1.3.0] - 2020-03-20
|
||||
### Added
|
||||
- [Background Power training option](doc/features/background-power/README.md)
|
||||
|
||||
## [1.2.1] - 2020-03-20
|
||||
### Fixed
|
||||
- Fixes bug with `fs-aug` color mode.
|
||||
|
|
@ -49,7 +53,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
|
|||
- 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))
|
||||
|
||||
[Unreleased]: https://github.com/olivierlacan/keep-a-changelog/compare/v1.2.1...HEAD
|
||||
[Unreleased]: https://github.com/olivierlacan/keep-a-changelog/compare/v1.3.0...HEAD
|
||||
[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
|
||||
|
|
|
|||
32
doc/features/background-power/README.md
Normal file
32
doc/features/background-power/README.md
Normal file
|
|
@ -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`:*
|
||||
|
||||
|
||||
|
||||
## 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.
|
||||
BIN
doc/features/background-power/example.jpeg
Normal file
BIN
doc/features/background-power/example.jpeg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 129 KiB |
BIN
doc/features/background-power/example2.jpeg
Normal file
BIN
doc/features/background-power/example2.jpeg
Normal file
Binary file not shown.
|
After Width: | Height: | Size: 121 KiB |
|
|
@ -54,6 +54,7 @@ class SAEHDModel(ModelBase):
|
|||
default_lr_dropout = self.options['lr_dropout'] = lr_dropout
|
||||
|
||||
default_random_warp = self.options['random_warp'] = self.load_or_def_option('random_warp', True)
|
||||
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)
|
||||
|
|
@ -164,6 +165,8 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
|
|||
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 )
|
||||
|
||||
|
|
@ -413,12 +416,15 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
|
|||
gpu_target_dst_style_anti_masked = gpu_target_dst*(1.0 - gpu_target_dstm_style_blur)
|
||||
|
||||
gpu_target_src_anti_masked = gpu_target_src*(1.0-gpu_target_srcm_blur)
|
||||
gpu_target_dst_anti_masked = gpu_target_dst_style_anti_masked
|
||||
|
||||
gpu_target_src_masked_opt = gpu_target_src*gpu_target_srcm_blur if masked_training else gpu_target_src
|
||||
gpu_target_dst_masked_opt = gpu_target_dst_masked if masked_training else gpu_target_dst
|
||||
|
||||
gpu_pred_src_src_masked_opt = gpu_pred_src_src*gpu_target_srcm_blur if masked_training else gpu_pred_src_src
|
||||
gpu_pred_src_src_anti_masked = gpu_pred_src_src*(1.0-gpu_target_srcm_blur)
|
||||
gpu_pred_dst_dst_masked_opt = gpu_pred_dst_dst*gpu_target_dstm_blur if masked_training else gpu_pred_dst_dst
|
||||
gpu_pred_dst_dst_anti_masked = gpu_pred_dst_dst*(1.0-gpu_target_dstm_blur)
|
||||
|
||||
gpu_psd_target_dst_style_masked = gpu_pred_src_dst*gpu_target_dstm_style_blur
|
||||
gpu_psd_target_dst_style_anti_masked = gpu_pred_src_dst*(1.0 - gpu_target_dstm_style_blur)
|
||||
|
|
@ -442,6 +448,15 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
|
|||
|
||||
gpu_src_loss += tf.reduce_mean ( 10*tf.square( gpu_target_srcm - gpu_pred_src_srcm ),axis=[1,2,3] )
|
||||
|
||||
if self.options['background_power'] > 0:
|
||||
bg_factor = self.options['background_power']
|
||||
if resolution < 256:
|
||||
gpu_src_loss += bg_factor * tf.reduce_mean ( 10*nn.dssim(gpu_target_src_anti_masked, gpu_pred_src_src_anti_masked, 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_anti_masked, gpu_pred_src_src_anti_masked, 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_anti_masked, gpu_pred_src_src_anti_masked, 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_anti_masked - gpu_pred_src_src_anti_masked ), axis=[1,2,3])
|
||||
|
||||
face_style_power = self.options['face_style_power'] / 100.0
|
||||
if face_style_power != 0 and not self.pretrain:
|
||||
gpu_src_loss += nn.style_loss(gpu_psd_target_dst_style_masked, gpu_target_dst_style_masked, gaussian_blur_radius=resolution//16, loss_weight=10000*face_style_power)
|
||||
|
|
@ -469,6 +484,15 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
|
|||
|
||||
gpu_dst_loss += tf.reduce_mean ( 300*tf.abs ( gpu_target_dst*gpu_target_part_mask - gpu_pred_dst_dst*gpu_target_part_mask ), axis=[1,2,3])
|
||||
|
||||
if self.options['background_power'] > 0:
|
||||
bg_factor = self.options['background_power']
|
||||
if resolution < 256:
|
||||
gpu_dst_loss += bg_factor * tf.reduce_mean ( 10*nn.dssim(gpu_target_dst_anti_masked, gpu_pred_dst_dst_anti_masked, 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_anti_masked, gpu_pred_dst_dst_anti_masked, 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_anti_masked, gpu_pred_dst_dst_anti_masked, 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_anti_masked - gpu_pred_dst_dst_anti_masked ), axis=[1,2,3])
|
||||
|
||||
gpu_dst_loss += tf.reduce_mean ( 10*tf.square( gpu_target_dstm - gpu_pred_dst_dstm ),axis=[1,2,3] )
|
||||
|
||||
gpu_src_losses += [gpu_src_loss]
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue