simcop2387/faceswap
1
0
Fork 0
mirror of https://github.com/deepfakes/faceswap synced 2025-06-07 10:43:27 -04:00
Code Issues Releases Wiki Activity

locales: Train settings

Browse source
This commit is contained in:
torzdf 2023-06-11 23:30:05 +01:00
parent a66214c58e
commit 02239d4183
5 changed files with 938 additions and 287 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

25
lib/config.py
View file

@ -3,6 +3,7 @@
Extends out :class:`configparser.ConfigParser` functionality by checking for default
configuration updates and returning data in it's correct format """
import gettext
import logging
import os
import sys
@ -16,6 +17,10 @@ from typing import Dict, List, Optional, Tuple, Union
from lib.utils import full_path_split
# LOCALES
_LANG = gettext.translation("lib.config", localedir="locales", fallback=True)
_ = _LANG.gettext
# Can't type OrderedDict fully on Python 3.8 or lower
if sys.version_info < (3, 9):
OrderedDictSectionType = OrderedDict
@ -282,7 +287,9 @@ class FaceswapConfig():
logger.error(err)
raise ValueError(err)
return configfile
dirname = os.path.dirname(sys.modules[self.__module__].__file__)
filepath = sys.modules[self.__module__].__file__
assert filepath is not None
dirname = os.path.dirname(filepath)
folder, fname = os.path.split(dirname)
retval = os.path.join(os.path.dirname(folder), "config", f"{fname}.ini")
logger.debug("Config File location: '%s'", retval)
@ -383,23 +390,23 @@ class FaceswapConfig():
""" Add extra helptext info from parameters """
helptext += "\n"
if not fixed:
helptext += "\nThis option can be updated for existing models.\n"
helptext += _("\nThis option can be updated for existing models.\n")
if datatype == list:
helptext += ("\nIf selecting multiple options then each option should be separated "
"by a space or a comma (e.g. item1, item2, item3)\n")
helptext += _("\nIf selecting multiple options then each option should be separated "
"by a space or a comma (e.g. item1, item2, item3)\n")
if choices and choices != "colorchooser":
helptext += f"\nChoose from: {choices}"
helptext += _("\nChoose from: {}").format(choices)
elif datatype == bool:
helptext += "\nChoose from: True, False"
helptext += _("\nChoose from: True, False")
elif datatype == int:
assert min_max is not None
cmin, cmax = min_max
helptext += f"\nSelect an integer between {cmin} and {cmax}"
helptext += _("\nSelect an integer between {} and {}").format(cmin, cmax)
elif datatype == float:
assert min_max is not None
cmin, cmax = min_max
helptext += f"\nSelect a decimal number between {cmin} and {cmax}"
helptext += f"\n[Default: {default}]"
helptext += _("\nSelect a decimal number between {} and {}").format(cmin, cmax)
helptext += _("\n[Default: {}]").format(default)
return helptext
def _check_exists(self) -> bool:

61
locales/lib.config.pot Normal file
View file

@ -0,0 +1,61 @@
# SOME DESCRIPTIVE TITLE.
# Copyright (C) YEAR THE PACKAGE'S COPYRIGHT HOLDER
# This file is distributed under the same license as the PACKAGE package.
# FIRST AUTHOR <EMAIL@ADDRESS>, YEAR.
#
#, fuzzy
msgid ""
msgstr ""
"Project-Id-Version: PACKAGE VERSION\n"
"Report-Msgid-Bugs-To: \n"
"POT-Creation-Date: 2023-06-11 23:28+0100\n"
"PO-Revision-Date: YEAR-MO-DA HO:MI+ZONE\n"
"Last-Translator: FULL NAME <EMAIL@ADDRESS>\n"
"Language-Team: LANGUAGE <LL@li.org>\n"
"Language: \n"
"MIME-Version: 1.0\n"
"Content-Type: text/plain; charset=CHARSET\n"
"Content-Transfer-Encoding: 8bit\n"
#: lib/config.py:393
msgid ""
"\n"
"This option can be updated for existing models.\n"
msgstr ""
#: lib/config.py:395
msgid ""
"\n"
"If selecting multiple options then each option should be separated by a "
"space or a comma (e.g. item1, item2, item3)\n"
msgstr ""
#: lib/config.py:398
msgid ""
"\n"
"Choose from: {}"
msgstr ""
#: lib/config.py:400
msgid ""
"\n"
"Choose from: True, False"
msgstr ""
#: lib/config.py:404
msgid ""
"\n"
"Select an integer between {} and {}"
msgstr ""
#: lib/config.py:408
msgid ""
"\n"
"Select a decimal number between {} and {}"
msgstr ""
#: lib/config.py:409
msgid ""
"\n"
"[Default: {}]"
msgstr ""

552
locales/plugins.train._config.pot Normal file
View file

@ -0,0 +1,552 @@
# SOME DESCRIPTIVE TITLE.
# Copyright (C) YEAR THE PACKAGE'S COPYRIGHT HOLDER
# This file is distributed under the same license as the PACKAGE package.
# FIRST AUTHOR <EMAIL@ADDRESS>, YEAR.
#
#, fuzzy
msgid ""
msgstr ""
"Project-Id-Version: PACKAGE VERSION\n"
"Report-Msgid-Bugs-To: \n"
"POT-Creation-Date: 2023-06-11 23:20+0100\n"
"PO-Revision-Date: YEAR-MO-DA HO:MI+ZONE\n"
"Last-Translator: FULL NAME <EMAIL@ADDRESS>\n"
"Language-Team: LANGUAGE <LL@li.org>\n"
"Language: \n"
"MIME-Version: 1.0\n"
"Content-Type: text/plain; charset=CHARSET\n"
"Content-Transfer-Encoding: 8bit\n"
#: plugins/train/_config.py:17
msgid ""
"\n"
"NB: Unless specifically stated, values changed here will only take effect "
"when creating a new model."
msgstr ""
#: plugins/train/_config.py:22
msgid ""
"Focal Frequency Loss. Analyzes the frequency spectrum of the images rather "
"than the images themselves. This loss function can be used on its own, but "
"the original paper found increased benefits when using it as a complementary "
"loss to another spacial loss function (e.g. MSE). Ref: Focal Frequency Loss "
"for Image Reconstruction and Synthesis https://arxiv.org/pdf/2012.12821.pdf "
"NB: This loss does not currently work on AMD cards."
msgstr ""
#: plugins/train/_config.py:29
msgid ""
"Nvidia FLIP. A perceptual loss measure that approximates the difference "
"perceived by humans as they alternate quickly (or flip) between two images. "
"Used on its own and this loss function creates a distinct grid on the "
"output. However it can be helpful when used as a complimentary loss "
"function. Ref: FLIP: A Difference Evaluator for Alternating Images: https://"
"research.nvidia.com/sites/default/files/node/3260/FLIP_Paper.pdf"
msgstr ""
#: plugins/train/_config.py:36
msgid ""
"Gradient Magnitude Similarity Deviation seeks to match the global standard "
"deviation of the pixel to pixel differences between two images. Similar in "
"approach to SSIM. Ref: Gradient Magnitude Similarity Deviation: An Highly "
"Efficient Perceptual Image Quality Index https://arxiv.org/ftp/arxiv/"
"papers/1308/1308.3052.pdf"
msgstr ""
#: plugins/train/_config.py:41
msgid ""
"The L_inf norm will reduce the largest individual pixel error in an image. "
"As each largest error is minimized sequentially, the overall error is "
"improved. This loss will be extremely focused on outliers."
msgstr ""
#: plugins/train/_config.py:45
msgid ""
"Laplacian Pyramid Loss. Attempts to improve results by focussing on edges "
"using Laplacian Pyramids. As this loss function gives priority to edges over "
"other low-frequency information, like color, it should not be used on its "
"own. The original implementation uses this loss as a complimentary function "
"to MSE. Ref: Optimizing the Latent Space of Generative Networks https://"
"arxiv.org/abs/1707.05776"
msgstr ""
#: plugins/train/_config.py:52
msgid ""
"LPIPS is a perceptual loss that uses the feature outputs of other pretrained "
"models as a loss metric. Be aware that this loss function will use more "
"VRAM. Used on its own and this loss will create a distinct moire pattern on "
"the output, however it can be helpful as a complimentary loss function. The "
"output of this function is strong, so depending on your chosen primary loss "
"function, you are unlikely going to want to set the weight above about 25%. "
"Ref: The Unreasonable Effectiveness of Deep Features as a Perceptual Metric "
"http://arxiv.org/abs/1801.03924\n"
"This variant uses the AlexNet backbone. A fairly light and old model which "
"performed best in the paper's original implementation.\n"
"NB: For AMD Users the final linear layer is not implemented."
msgstr ""
#: plugins/train/_config.py:62
msgid ""
"Same as lpips_alex, but using the SqueezeNet backbone. A more lightweight "
"version of AlexNet.\n"
"NB: For AMD Users the final linear layer is not implemented."
msgstr ""
#: plugins/train/_config.py:65
msgid ""
"Same as lpips_alex, but using the VGG16 backbone. A more heavyweight model.\n"
"NB: For AMD Users the final linear layer is not implemented."
msgstr ""
#: plugins/train/_config.py:68
msgid ""
"log(cosh(x)) acts similar to MSE for small errors and to MAE for large "
"errors. Like MSE, it is very stable and prevents overshoots when errors are "
"near zero. Like MAE, it is robust to outliers."
msgstr ""
#: plugins/train/_config.py:72
msgid ""
"Mean absolute error will guide reconstructions of each pixel towards its "
"median value in the training dataset. Robust to outliers but as a median, it "
"can potentially ignore some infrequent image types in the dataset."
msgstr ""
#: plugins/train/_config.py:76
msgid ""
"Mean squared error will guide reconstructions of each pixel towards its "
"average value in the training dataset. As an avg, it will be susceptible to "
"outliers and typically produces slightly blurrier results. Ref: Multi-Scale "
"Structural Similarity for Image Quality Assessment https://www.cns.nyu.edu/"
"pub/eero/wang03b.pdf"
msgstr ""
#: plugins/train/_config.py:81
msgid ""
"Multiscale Structural Similarity Index Metric is similar to SSIM except that "
"it performs the calculations along multiple scales of the input image."
msgstr ""
#: plugins/train/_config.py:84
msgid ""
"Smooth_L1 is a modification of the MAE loss to correct two of its "
"disadvantages. This loss has improved stability and guidance for small "
"errors. Ref: A General and Adaptive Robust Loss Function https://arxiv.org/"
"pdf/1701.03077.pdf"
msgstr ""
#: plugins/train/_config.py:88
msgid ""
"Structural Similarity Index Metric is a perception-based loss that considers "
"changes in texture, luminance, contrast, and local spatial statistics of an "
"image. Potentially delivers more realistic looking images. Ref: Image "
"Quality Assessment: From Error Visibility to Structural Similarity http://"
"www.cns.nyu.edu/pub/eero/wang03-reprint.pdf"
msgstr ""
#: plugins/train/_config.py:93
msgid ""
"Instead of minimizing the difference between the absolute value of each "
"pixel in two reference images, compute the pixel to pixel spatial difference "
"in each image and then minimize that difference between two images. Allows "
"for large color shifts, but maintains the structure of the image."
msgstr ""
#: plugins/train/_config.py:97
msgid "Do not use an additional loss function."
msgstr ""
#: plugins/train/_config.py:117
msgid "Options that apply to all models"
msgstr ""
#: plugins/train/_config.py:126 plugins/train/_config.py:150
msgid "face"
msgstr ""
#: plugins/train/_config.py:128
msgid ""
"How to center the training image. The extracted images are centered on the "
"middle of the skull based on the face's estimated pose. A subsection of "
"these images are used for training. The centering used dictates how this "
"subsection will be cropped from the aligned images.\n"
"\tface: Centers the training image on the center of the face, adjusting for "
"pitch and yaw.\n"
"\thead: Centers the training image on the center of the head, adjusting for "
"pitch and yaw. NB: You should only select head centering if you intend to "
"include the full head (including hair) in the final swap. This may give "
"mixed results. Additionally, it is only worth choosing head centering if you "
"are training with a mask that includes the hair (e.g. BiSeNet-FP-Head).\n"
"\tlegacy: The 'original' extraction technique. Centers the training image "
"near the tip of the nose with no adjustment. Can result in the edges of the "
"face appearing outside of the training area."
msgstr ""
#: plugins/train/_config.py:152
msgid ""
"How much of the extracted image to train on. A lower coverage will limit the "
"model's scope to a zoomed-in central area while higher amounts can include "
"the entire face. A trade-off exists between lower amounts given more detail "
"versus higher amounts avoiding noticeable swap transitions. For 'Face' "
"centering you will want to leave this above 75%. For Head centering you will "
"most likely want to set this to 100%. Sensible values for 'Legacy' centering "
"are:\n"
"\t62.5% spans from eyebrow to eyebrow.\n"
"\t75.0% spans from temple to temple.\n"
"\t87.5% spans from ear to ear.\n"
"\t100.0% is a mugshot."
msgstr ""
#: plugins/train/_config.py:168 plugins/train/_config.py:179
msgid "initialization"
msgstr ""
#: plugins/train/_config.py:170
msgid ""
"Use ICNR to tile the default initializer in a repeating pattern. This "
"strategy is designed for pairing with sub-pixel / pixel shuffler to reduce "
"the 'checkerboard effect' in image reconstruction. \n"
"\t https://arxiv.org/ftp/arxiv/papers/1707/1707.02937.pdf"
msgstr ""
#: plugins/train/_config.py:181
msgid ""
"Use Convolution Aware Initialization for convolutional layers. This can help "
"eradicate the vanishing and exploding gradient problem as well as lead to "
"higher accuracy, lower loss and faster convergence.\n"
"NB:\n"
"\t This can use more VRAM when creating a new model so you may want to lower "
"the batch size for the first run. The batch size can be raised again when "
"reloading the model. \n"
"\t Multi-GPU is not supported for this option, so you should start the model "
"on a single GPU. Once training has started, you can stop training, enable "
"multi-GPU and resume.\n"
"\t Building the model will likely take several minutes as the calculations "
"for this initialization technique are expensive. This will only impact "
"starting a new model."
msgstr ""
#: plugins/train/_config.py:198 plugins/train/_config.py:223
#: plugins/train/_config.py:238 plugins/train/_config.py:265
msgid "optimizer"
msgstr ""
#: plugins/train/_config.py:202
msgid ""
"The optimizer to use.\n"
"\t adabelief - Adapting Stepsizes by the Belief in Observed Gradients. An "
"optimizer with the aim to converge faster, generalize better and remain more "
"stable. (https://arxiv.org/abs/2010.07468). NB: Epsilon for AdaBelief needs "
"to be set to a smaller value than other Optimizers. Generally setting the "
"'Epsilon Exponent' to around '-16' should work.\n"
"\t adam - Adaptive Moment Optimization. A stochastic gradient descent method "
"that is based on adaptive estimation of first-order and second-order "
"moments.\n"
"\t nadam - Adaptive Moment Optimization with Nesterov Momentum. Much like "
"Adam but uses a different formula for calculating momentum.\n"
"\t rms-prop - Root Mean Square Propagation. Maintains a moving (discounted) "
"average of the square of the gradients. Divides the gradient by the root of "
"this average."
msgstr ""
#: plugins/train/_config.py:225
msgid ""
"Learning rate - how fast your network will learn (how large are the "
"modifications to the model weights after one batch of training). Values that "
"are too large might result in model crashes and the inability of the model "
"to find the best solution. Values that are too small might be unable to "
"escape from dead-ends and find the best global minimum."
msgstr ""
#: plugins/train/_config.py:240
msgid ""
"The epsilon adds a small constant to weight updates to attempt to avoid "
"'divide by zero' errors. Unless you are using the AdaBelief Optimizer, then "
"Generally this option should be left at default value, For AdaBelief, "
"setting this to around '-16' should work.\n"
"In all instances if you are getting 'NaN' loss values, and have been unable "
"to resolve the issue any other way (for example, increasing batch size, or "
"lowering learning rate), then raising the epsilon can lead to a more stable "
"model. It may, however, come at the cost of slower training and a less "
"accurate final result.\n"
"NB: The value given here is the 'exponent' to the epsilon. For example, "
"choosing '-7' will set the epsilon to 1e-7. Choosing '-3' will set the "
"epsilon to 0.001 (1e-3)."
msgstr ""
#: plugins/train/_config.py:258
msgid ""
"[Not PlaidML] Apply AutoClipping to the gradients. AutoClip analyzes the "
"gradient weights and adjusts the normalization value dynamically to fit the "
"data. Can help prevent NaNs and improve model optimization at the expense of "
"VRAM. Ref: AutoClip: Adaptive Gradient Clipping for Source Separation "
"Networks https://arxiv.org/abs/2007.14469"
msgstr ""
#: plugins/train/_config.py:271 plugins/train/_config.py:283
#: plugins/train/_config.py:297 plugins/train/_config.py:314
msgid "network"
msgstr ""
#: plugins/train/_config.py:273
msgid ""
"Use reflection padding rather than zero padding with convolutions. Each "
"convolution must pad the image boundaries to maintain the proper sizing. "
"More complex padding schemes can reduce artifacts at the border of the "
"image.\n"
"\t http://www-cs.engr.ccny.cuny.edu/~wolberg/cs470/hw/hw2_pad.txt"
msgstr ""
#: plugins/train/_config.py:286
msgid ""
"[Nvidia Only]. Enable the Tensorflow GPU 'allow_growth' configuration "
"option. This option prevents Tensorflow from allocating all of the GPU VRAM "
"at launch but can lead to higher VRAM fragmentation and slower performance. "
"Should only be enabled if you are receiving errors regarding 'cuDNN fails to "
"initialize' when commencing training."
msgstr ""
#: plugins/train/_config.py:299
msgid ""
"[Not PlaidML], NVIDIA GPUs can run operations in float16 faster than in "
"float32. Mixed precision allows you to use a mix of float16 with float32, to "
"get the performance benefits from float16 and the numeric stability benefits "
"from float32.\n"
"\n"
"This is untested on DirectML backend, but will run on most Nvidia models. it "
"will only speed up training on more recent GPUs. Those with compute "
"capability 7.0 or higher will see the greatest performance benefit from "
"mixed precision because they have Tensor Cores. Older GPUs offer no math "
"performance benefit for using mixed precision, however memory and bandwidth "
"savings can enable some speedups. Generally RTX GPUs and later will offer "
"the most benefit."
msgstr ""
#: plugins/train/_config.py:316
msgid ""
"If a 'NaN' is generated in the model, this means that the model has "
"corrupted and the model is likely to start deteriorating from this point on. "
"Enabling NaN protection will stop training immediately in the event of a "
"NaN. The last save will not contain the NaN, so you may still be able to "
"rescue your model."
msgstr ""
#: plugins/train/_config.py:329
msgid "convert"
msgstr ""
#: plugins/train/_config.py:331
msgid ""
"[GPU Only]. The number of faces to feed through the model at once when "
"running the Convert process.\n"
"\n"
"NB: Increasing this figure is unlikely to improve convert speed, however, if "
"you are getting Out of Memory errors, then you may want to reduce the batch "
"size."
msgstr ""
#: plugins/train/_config.py:350
msgid ""
"Loss configuration options\n"
"Loss is the mechanism by which a Neural Network judges how well it thinks "
"that it is recreating a face."
msgstr ""
#: plugins/train/_config.py:357 plugins/train/_config.py:369
#: plugins/train/_config.py:382 plugins/train/_config.py:402
#: plugins/train/_config.py:414 plugins/train/_config.py:434
#: plugins/train/_config.py:446 plugins/train/_config.py:466
#: plugins/train/_config.py:482 plugins/train/_config.py:498
#: plugins/train/_config.py:515
msgid "loss"
msgstr ""
#: plugins/train/_config.py:361
msgid "The loss function to use."
msgstr ""
#: plugins/train/_config.py:373
msgid ""
"The second loss function to use. If using a structural based loss (such as "
"SSIM, MS-SSIM or GMSD) it is common to add an L1 regularization(MAE) or L2 "
"regularization (MSE) function. You can adjust the weighting of this loss "
"function with the loss_weight_2 option."
msgstr ""
#: plugins/train/_config.py:388
msgid ""
"The amount of weight to apply to the second loss function.\n"
"\n"
"\n"
"\n"
"The value given here is as a percentage denoting how much the selected "
"function should contribute to the overall loss cost of the model. For "
"example:\n"
"\t 100 - The loss calculated for the second loss function will be applied at "
"its full amount towards the overall loss score. \n"
"\t 25 - The loss calculated for the second loss function will be reduced by "
"a quarter prior to adding to the overall loss score. \n"
"\t 400 - The loss calculated for the second loss function will be mulitplied "
"4 times prior to adding to the overall loss score. \n"
"\t 0 - Disables the second loss function altogether."
msgstr ""
#: plugins/train/_config.py:406
msgid ""
"The third loss function to use. You can adjust the weighting of this loss "
"function with the loss_weight_3 option."
msgstr ""
#: plugins/train/_config.py:420
msgid ""
"The amount of weight to apply to the third loss function.\n"
"\n"
"\n"
"\n"
"The value given here is as a percentage denoting how much the selected "
"function should contribute to the overall loss cost of the model. For "
"example:\n"
"\t 100 - The loss calculated for the third loss function will be applied at "
"its full amount towards the overall loss score. \n"
"\t 25 - The loss calculated for the third loss function will be reduced by a "
"quarter prior to adding to the overall loss score. \n"
"\t 400 - The loss calculated for the third loss function will be mulitplied "
"4 times prior to adding to the overall loss score. \n"
"\t 0 - Disables the third loss function altogether."
msgstr ""
#: plugins/train/_config.py:438
msgid ""
"The fourth loss function to use. You can adjust the weighting of this loss "
"function with the loss_weight_3 option."
msgstr ""
#: plugins/train/_config.py:452
msgid ""
"The amount of weight to apply to the fourth loss function.\n"
"\n"
"\n"
"\n"
"The value given here is as a percentage denoting how much the selected "
"function should contribute to the overall loss cost of the model. For "
"example:\n"
"\t 100 - The loss calculated for the fourth loss function will be applied at "
"its full amount towards the overall loss score. \n"
"\t 25 - The loss calculated for the fourth loss function will be reduced by "
"a quarter prior to adding to the overall loss score. \n"
"\t 400 - The loss calculated for the fourth loss function will be mulitplied "
"4 times prior to adding to the overall loss score. \n"
"\t 0 - Disables the fourth loss function altogether."
msgstr ""
#: plugins/train/_config.py:471
msgid ""
"The loss function to use when learning a mask.\n"
"\t MAE - Mean absolute error will guide reconstructions of each pixel "
"towards its median value in the training dataset. Robust to outliers but as "
"a median, it can potentially ignore some infrequent image types in the "
"dataset.\n"
"\t MSE - Mean squared error will guide reconstructions of each pixel towards "
"its average value in the training dataset. As an average, it will be "
"susceptible to outliers and typically produces slightly blurrier results."
msgstr ""
#: plugins/train/_config.py:488
msgid ""
"The amount of priority to give to the eyes.\n"
"\n"
"The value given here is as a multiplier of the main loss score. For "
"example:\n"
"\t 1 - The eyes will receive the same priority as the rest of the face. \n"
"\t 10 - The eyes will be given a score 10 times higher than the rest of the "
"face.\n"
"\n"
"NB: Penalized Mask Loss must be enable to use this option."
msgstr ""
#: plugins/train/_config.py:504
msgid ""
"The amount of priority to give to the mouth.\n"
"\n"
"The value given here is as a multiplier of the main loss score. For "
"Example:\n"
"\t 1 - The mouth will receive the same priority as the rest of the face. \n"
"\t 10 - The mouth will be given a score 10 times higher than the rest of the "
"face.\n"
"\n"
"NB: Penalized Mask Loss must be enable to use this option."
msgstr ""
#: plugins/train/_config.py:517
msgid ""
"Image loss function is weighted by mask presence. For areas of the image "
"without the facial mask, reconstruction errors will be ignored while the "
"masked face area is prioritized. May increase overall quality by focusing "
"attention on the core face area."
msgstr ""
#: plugins/train/_config.py:528 plugins/train/_config.py:570
#: plugins/train/_config.py:584 plugins/train/_config.py:593
msgid "mask"
msgstr ""
#: plugins/train/_config.py:531
msgid ""
"The mask to be used for training. If you have selected 'Learn Mask' or "
"'Penalized Mask Loss' you must select a value other than 'none'. The "
"required mask should have been selected as part of the Extract process. If "
"it does not exist in the alignments file then it will be generated prior to "
"training commencing.\n"
"\tnone: Don't use a mask.\n"
"\tbisenet-fp_face: Relatively lightweight NN based mask that provides more "
"refined control over the area to be masked (configurable in mask settings). "
"Use this version of bisenet-fp if your model is trained with 'face' or "
"'legacy' centering.\n"
"\tbisenet-fp_head: Relatively lightweight NN based mask that provides more "
"refined control over the area to be masked (configurable in mask settings). "
"Use this version of bisenet-fp if your model is trained with 'head' "
"centering.\n"
"\tcomponents: Mask designed to provide facial segmentation based on the "
"positioning of landmark locations. A convex hull is constructed around the "
"exterior of the landmarks to create a mask.\n"
"\tcustom_face: Custom user created, face centered mask.\n"
"\tcustom_head: Custom user created, head centered mask.\n"
"\textended: Mask designed to provide facial segmentation based on the "
"positioning of landmark locations. A convex hull is constructed around the "
"exterior of the landmarks and the mask is extended upwards onto the "
"forehead.\n"
"\tvgg-clear: Mask designed to provide smart segmentation of mostly frontal "
"faces clear of obstructions. Profile faces and obstructions may result in "
"sub-par performance.\n"
"\tvgg-obstructed: Mask designed to provide smart segmentation of mostly "
"frontal faces. The mask model has been specifically trained to recognize "
"some facial obstructions (hands and eyeglasses). Profile faces may result in "
"sub-par performance.\n"
"\tunet-dfl: Mask designed to provide smart segmentation of mostly frontal "
"faces. The mask model has been trained by community members and will need "
"testing for further description. Profile faces may result in sub-par "
"performance."
msgstr ""
#: plugins/train/_config.py:572
msgid ""
"Apply gaussian blur to the mask input. This has the effect of smoothing the "
"edges of the mask, which can help with poorly calculated masks and give less "
"of a hard edge to the predicted mask. The size is in pixels (calculated from "
"a 128px mask). Set to 0 to not apply gaussian blur. This value should be "
"odd, if an even number is passed in then it will be rounded to the next odd "
"number."
msgstr ""
#: plugins/train/_config.py:586
msgid ""
"Sets pixels that are near white to white and near black to black. Set to 0 "
"for off."
msgstr ""
#: plugins/train/_config.py:595
msgid ""
"Dedicate a portion of the model to learning how to duplicate the input mask. "
"Increases VRAM usage in exchange for learning a quick ability to try to "
"replicate more complex mask models."
msgstr ""

2
plugins/extract/_config.py
View file

@ -11,7 +11,7 @@ from lib.config import FaceswapConfig
_LANG = gettext.translation("plugins.extract._config", localedir="locales", fallback=True)
_ = _LANG.gettext
logger = logging.getLogger(__name__) # pylint: disable=invalid-name
logger = logging.getLogger(__name__)
class Config(FaceswapConfig):

585
plugins/train/_config.py
View file

@ -1,46 +1,54 @@
#!/usr/bin/env python3
""" Default configurations for models """
import gettext
import logging
import os
from lib.config import FaceswapConfig
from plugins.plugin_loader import PluginLoader
logger = logging.getLogger(__name__) # pylint: disable=invalid-name
# LOCALES
_LANG = gettext.translation("plugins.train._config", localedir="locales", fallback=True)
_ = _LANG.gettext
ADDITIONAL_INFO = ("\nNB: Unless specifically stated, values changed here will only take effect "
"when creating a new model.")
logger = logging.getLogger(__name__)
_LOSS_HELP = dict(
ffl="Focal Frequency Loss. Analyzes the frequency spectrum of the images rather than the "
ADDITIONAL_INFO = _("\nNB: Unless specifically stated, values changed here will only take effect "
"when creating a new model.")
_LOSS_HELP = {
"ffl": _(
"Focal Frequency Loss. Analyzes the frequency spectrum of the images rather than the "
"images themselves. This loss function can be used on its own, but the original paper "
"found increased benefits when using it as a complementary loss to another spacial loss "
"function (e.g. MSE). Ref: Focal Frequency Loss for Image Reconstruction and Synthesis "
"https://arxiv.org/pdf/2012.12821.pdf NB: This loss does not currently work on AMD cards.",
flip="Nvidia FLIP. A perceptual loss measure that approximates the difference perceived by "
"humans as they alternate quickly (or flip) between two images. Used on its own and this "
"loss function creates a distinct grid on the output. However it can be helpful when "
"used as a complimentary loss function. Ref: FLIP: A Difference Evaluator for "
"Alternating Images: "
"https://research.nvidia.com/sites/default/files/node/3260/FLIP_Paper.pdf",
gmsd=(
"https://arxiv.org/pdf/2012.12821.pdf NB: This loss does not currently work on AMD "
"cards."),
"flip": _(
"Nvidia FLIP. A perceptual loss measure that approximates the difference perceived by "
"humans as they alternate quickly (or flip) between two images. Used on its own and this "
"loss function creates a distinct grid on the output. However it can be helpful when "
"used as a complimentary loss function. Ref: FLIP: A Difference Evaluator for "
"Alternating Images: "
"https://research.nvidia.com/sites/default/files/node/3260/FLIP_Paper.pdf"),
"gmsd": _(
"Gradient Magnitude Similarity Deviation seeks to match the global standard deviation of "
"the pixel to pixel differences between two images. Similar in approach to SSIM. Ref: "
"Gradient Magnitude Similarity Deviation: An Highly Efficient Perceptual Image Quality "
"Index https://arxiv.org/ftp/arxiv/papers/1308/1308.3052.pdf"),
l_inf_norm=(
"l_inf_norm": _(
"The L_inf norm will reduce the largest individual pixel error in an image. As "
"each largest error is minimized sequentially, the overall error is improved. This loss "
"will be extremely focused on outliers."),
laploss=(
"laploss": _(
"Laplacian Pyramid Loss. Attempts to improve results by focussing on edges using "
"Laplacian Pyramids. As this loss function gives priority to edges over other low-"
"frequency information, like color, it should not be used on its own. The original "
"implementation uses this loss as a complimentary function to MSE. "
"Ref: Optimizing the Latent Space of Generative Networks "
"https://arxiv.org/abs/1707.05776"),
lpips_alex=(
"lpips_alex": _(
"LPIPS is a perceptual loss that uses the feature outputs of other pretrained models as a "
"loss metric. Be aware that this loss function will use more VRAM. Used on its own and "
"this loss will create a distinct moire pattern on the output, however it can be helpful "
@ -50,42 +58,43 @@ _LOSS_HELP = dict(
"Metric http://arxiv.org/abs/1801.03924\nThis variant uses the AlexNet backbone. A fairly "
"light and old model which performed best in the paper's original implementation.\nNB: "
"For AMD Users the final linear layer is not implemented."),
lpips_squeeze=(
"lpips_squeeze": _(
"Same as lpips_alex, but using the SqueezeNet backbone. A more lightweight "
"version of AlexNet.\nNB: For AMD Users the final linear layer is not implemented."),
lpips_vgg16="Same as lpips_alex, but using the VGG16 backbone. A more heavyweight model.\n"
"NB: For AMD Users the final linear layer is not implemented.",
logcosh=(
"lpips_vgg16": _(
"Same as lpips_alex, but using the VGG16 backbone. A more heavyweight model.\n"
"NB: For AMD Users the final linear layer is not implemented."),
"logcosh": _(
"log(cosh(x)) acts similar to MSE for small errors and to MAE for large errors. Like "
"MSE, it is very stable and prevents overshoots when errors are near zero. Like MAE, it "
"is robust to outliers."),
mae=(
"mae": _(
"Mean absolute error will guide reconstructions of each pixel towards its median value in "
"the training dataset. Robust to outliers but as a median, it can potentially ignore some "
"infrequent image types in the dataset."),
mse=(
"mse": _(
"Mean squared error will guide reconstructions of each pixel towards its average value in "
"the training dataset. As an avg, it will be susceptible to outliers and typically "
"produces slightly blurrier results. Ref: Multi-Scale Structural Similarity for Image "
"Quality Assessment https://www.cns.nyu.edu/pub/eero/wang03b.pdf"),
ms_ssim=(
"ms_ssim": _(
"Multiscale Structural Similarity Index Metric is similar to SSIM except that it "
"performs the calculations along multiple scales of the input image."),
smooth_loss=(
"smooth_loss": _(
"Smooth_L1 is a modification of the MAE loss to correct two of its disadvantages. "
"This loss has improved stability and guidance for small errors. Ref: A General and "
"Adaptive Robust Loss Function https://arxiv.org/pdf/1701.03077.pdf"),
ssim=(
"ssim": _(
"Structural Similarity Index Metric is a perception-based loss that considers changes in "
"texture, luminance, contrast, and local spatial statistics of an image. Potentially "
"delivers more realistic looking images. Ref: Image Quality Assessment: From Error "
"Visibility to Structural Similarity http://www.cns.nyu.edu/pub/eero/wang03-reprint.pdf"),
pixel_gradient_diff=(
"pixel_gradient_diff": _(
"Instead of minimizing the difference between the absolute value of each "
"pixel in two reference images, compute the pixel to pixel spatial difference in each "
"image and then minimize that difference between two images. Allows for large color "
"shifts, but maintains the structure of the image."),
none="Do not use an additional loss function.")
"none": _("Do not use an additional loss function.")}
_NON_PRIMARY_LOSS = ["flip", "lpips_alex", "lpips_squeeze", "lpips_vgg16", "none"]
@ -105,7 +114,7 @@ class Config(FaceswapConfig):
logger.debug("Setting global config")
section = "global"
self.add_section(section,
"Options that apply to all models" + ADDITIONAL_INFO)
_("Options that apply to all models") + ADDITIONAL_INFO)
self.add_item(
section=section,
title="centering",
@ -114,21 +123,22 @@ class Config(FaceswapConfig):
default="face",
choices=["face", "head", "legacy"],
fixed=True,
group="face",
info="How to center the training image. The extracted images are centered on the "
"middle of the skull based on the face's estimated pose. A subsection of these "
"images are used for training. The centering used dictates how this subsection "
"will be cropped from the aligned images."
"\n\tface: Centers the training image on the center of the face, adjusting for "
"pitch and yaw."
"\n\thead: Centers the training image on the center of the head, adjusting for "
"pitch and yaw. NB: You should only select head centering if you intend to "
"include the full head (including hair) in the final swap. This may give mixed "
"results. Additionally, it is only worth choosing head centering if you are "
"training with a mask that includes the hair (e.g. BiSeNet-FP-Head)."
"\n\tlegacy: The 'original' extraction technique. Centers the training image "
"near the tip of the nose with no adjustment. Can result in the edges of the "
"face appearing outside of the training area.")
group=_("face"),
info=_(
"How to center the training image. The extracted images are centered on the "
"middle of the skull based on the face's estimated pose. A subsection of these "
"images are used for training. The centering used dictates how this subsection "
"will be cropped from the aligned images."
"\n\tface: Centers the training image on the center of the face, adjusting for "
"pitch and yaw."
"\n\thead: Centers the training image on the center of the head, adjusting for "
"pitch and yaw. NB: You should only select head centering if you intend to "
"include the full head (including hair) in the final swap. This may give mixed "
"results. Additionally, it is only worth choosing head centering if you are "
"training with a mask that includes the hair (e.g. BiSeNet-FP-Head)."
"\n\tlegacy: The 'original' extraction technique. Centers the training image "
"near the tip of the nose with no adjustment. Can result in the edges of the "
"face appearing outside of the training area."))
self.add_item(
section=section,
title="coverage",
@ -137,68 +147,71 @@ class Config(FaceswapConfig):
min_max=(62.5, 100.0),
rounding=2,
fixed=True,
group="face",
info="How much of the extracted image to train on. A lower coverage will limit the "
"model's scope to a zoomed-in central area while higher amounts can include the "
"entire face. A trade-off exists between lower amounts given more detail "
"versus higher amounts avoiding noticeable swap transitions. For 'Face' "
"centering you will want to leave this above 75%. For Head centering you will "
"most likely want to set this to 100%. Sensible values for 'Legacy' "
"centering are:"
"\n\t62.5% spans from eyebrow to eyebrow."
"\n\t75.0% spans from temple to temple."
"\n\t87.5% spans from ear to ear."
"\n\t100.0% is a mugshot.")
group=_("face"),
info=_(
"How much of the extracted image to train on. A lower coverage will limit the "
"model's scope to a zoomed-in central area while higher amounts can include the "
"entire face. A trade-off exists between lower amounts given more detail "
"versus higher amounts avoiding noticeable swap transitions. For 'Face' "
"centering you will want to leave this above 75%. For Head centering you will "
"most likely want to set this to 100%. Sensible values for 'Legacy' "
"centering are:"
"\n\t62.5% spans from eyebrow to eyebrow."
"\n\t75.0% spans from temple to temple."
"\n\t87.5% spans from ear to ear."
"\n\t100.0% is a mugshot."))
self.add_item(
section=section,
title="icnr_init",
datatype=bool,
default=False,
group="initialization",
info="Use ICNR to tile the default initializer in a repeating pattern. "
"This strategy is designed for pairing with sub-pixel / pixel shuffler "
"to reduce the 'checkerboard effect' in image reconstruction. "
"\n\t https://arxiv.org/ftp/arxiv/papers/1707/1707.02937.pdf")
group=_("initialization"),
info=_(
"Use ICNR to tile the default initializer in a repeating pattern. "
"This strategy is designed for pairing with sub-pixel / pixel shuffler "
"to reduce the 'checkerboard effect' in image reconstruction. "
"\n\t https://arxiv.org/ftp/arxiv/papers/1707/1707.02937.pdf"))
self.add_item(
section=section,
title="conv_aware_init",
datatype=bool,
default=False,
group="initialization",
info="Use Convolution Aware Initialization for convolutional layers. "
"This can help eradicate the vanishing and exploding gradient problem "
"as well as lead to higher accuracy, lower loss and faster convergence.\nNB:"
"\n\t This can use more VRAM when creating a new model so you may want to "
"lower the batch size for the first run. The batch size can be raised "
"again when reloading the model. "
"\n\t Multi-GPU is not supported for this option, so you should start the model "
"on a single GPU. Once training has started, you can stop training, enable "
"multi-GPU and resume."
"\n\t Building the model will likely take several minutes as the calculations "
"for this initialization technique are expensive. This will only impact starting "
"a new model.")
group=_("initialization"),
info=_(
"Use Convolution Aware Initialization for convolutional layers. "
"This can help eradicate the vanishing and exploding gradient problem "
"as well as lead to higher accuracy, lower loss and faster convergence.\nNB:"
"\n\t This can use more VRAM when creating a new model so you may want to "
"lower the batch size for the first run. The batch size can be raised "
"again when reloading the model. "
"\n\t Multi-GPU is not supported for this option, so you should start the model "
"on a single GPU. Once training has started, you can stop training, enable "
"multi-GPU and resume."
"\n\t Building the model will likely take several minutes as the calculations "
"for this initialization technique are expensive. This will only impact starting "
"a new model."))
self.add_item(
section=section,
title="optimizer",
datatype=str,
gui_radio=True,
group="optimizer",
group=_("optimizer"),
default="adam",
choices=["adabelief", "adam", "nadam", "rms-prop"],
info="The optimizer to use."
"\n\t adabelief - Adapting Stepsizes by the Belief in Observed Gradients. An "
"optimizer with the aim to converge faster, generalize better and remain more "
"stable. (https://arxiv.org/abs/2010.07468). NB: Epsilon for AdaBelief needs to "
"be set to a smaller value than other Optimizers. Generally setting the 'Epsilon "
"Exponent' to around '-16' should work."
"\n\t adam - Adaptive Moment Optimization. A stochastic gradient descent method "
"that is based on adaptive estimation of first-order and second-order moments."
"\n\t nadam - Adaptive Moment Optimization with Nesterov Momentum. Much like "
"Adam but uses a different formula for calculating momentum."
"\n\t rms-prop - Root Mean Square Propagation. Maintains a moving (discounted) "
"average of the square of the gradients. Divides the gradient by the root of "
"this average.")
info=_(
"The optimizer to use."
"\n\t adabelief - Adapting Stepsizes by the Belief in Observed Gradients. An "
"optimizer with the aim to converge faster, generalize better and remain more "
"stable. (https://arxiv.org/abs/2010.07468). NB: Epsilon for AdaBelief needs to "
"be set to a smaller value than other Optimizers. Generally setting the 'Epsilon "
"Exponent' to around '-16' should work."
"\n\t adam - Adaptive Moment Optimization. A stochastic gradient descent method "
"that is based on adaptive estimation of first-order and second-order moments."
"\n\t nadam - Adaptive Moment Optimization with Nesterov Momentum. Much like "
"Adam but uses a different formula for calculating momentum."
"\n\t rms-prop - Root Mean Square Propagation. Maintains a moving (discounted) "
"average of the square of the gradients. Divides the gradient by the root of "
"this average."))
self.add_item(
section=section,
title="learning_rate",
@ -207,12 +220,13 @@ class Config(FaceswapConfig):
min_max=(1e-6, 1e-4),
rounding=6,
fixed=False,
group="optimizer",
info="Learning rate - how fast your network will learn (how large are the "
"modifications to the model weights after one batch of training). Values that "
"are too large might result in model crashes and the inability of the model to "
"find the best solution. Values that are too small might be unable to escape "
"from dead-ends and find the best global minimum.")
group=_("optimizer"),
info=_(
"Learning rate - how fast your network will learn (how large are the "
"modifications to the model weights after one batch of training). Values that "
"are too large might result in model crashes and the inability of the model to "
"find the best solution. Values that are too small might be unable to escape "
"from dead-ends and find the best global minimum."))
self.add_item(
section=section,
title="epsilon_exponent",
@ -221,82 +235,88 @@ class Config(FaceswapConfig):
min_max=(-20, 0),
rounding=1,
fixed=False,
group="optimizer",
info="The epsilon adds a small constant to weight updates to attempt to avoid 'divide "
"by zero' errors. Unless you are using the AdaBelief Optimizer, then Generally "
"this option should be left at default value, For AdaBelief, setting this to "
"around '-16' should work.\n"
"In all instances if you are getting 'NaN' loss values, and have been unable to "
"resolve the issue any other way (for example, increasing batch size, or "
"lowering learning rate), then raising the epsilon can lead to a more stable "
"model. It may, however, come at the cost of slower training and a less accurate "
"final result.\n"
"NB: The value given here is the 'exponent' to the epsilon. For example, "
"choosing '-7' will set the epsilon to 1e-7. Choosing '-3' will set the epsilon "
"to 0.001 (1e-3).")
group=_("optimizer"),
info=_(
"The epsilon adds a small constant to weight updates to attempt to avoid 'divide "
"by zero' errors. Unless you are using the AdaBelief Optimizer, then Generally "
"this option should be left at default value, For AdaBelief, setting this to "
"around '-16' should work.\n"
"In all instances if you are getting 'NaN' loss values, and have been unable to "
"resolve the issue any other way (for example, increasing batch size, or "
"lowering learning rate), then raising the epsilon can lead to a more stable "
"model. It may, however, come at the cost of slower training and a less accurate "
"final result.\n"
"NB: The value given here is the 'exponent' to the epsilon. For example, "
"choosing '-7' will set the epsilon to 1e-7. Choosing '-3' will set the epsilon "
"to 0.001 (1e-3)."))
self.add_item(
section=section,
title="autoclip",
datatype=bool,
default=False,
info="[Not PlaidML] Apply AutoClipping to the gradients. AutoClip analyzes the "
"gradient weights and adjusts the normalization value dynamically to fit the "
"data. Can help prevent NaNs and improve model optimization at the expense of "
"VRAM. Ref: AutoClip: Adaptive Gradient Clipping for Source Separation Networks "
"https://arxiv.org/abs/2007.14469",
info=_(
"[Not PlaidML] Apply AutoClipping to the gradients. AutoClip analyzes the "
"gradient weights and adjusts the normalization value dynamically to fit the "
"data. Can help prevent NaNs and improve model optimization at the expense of "
"VRAM. Ref: AutoClip: Adaptive Gradient Clipping for Source Separation Networks "
"https://arxiv.org/abs/2007.14469"),
fixed=False,
gui_radio=True,
group="optimizer")
group=_("optimizer"))
self.add_item(
section=section,
title="reflect_padding",
datatype=bool,
default=False,
group="network",
info="Use reflection padding rather than zero padding with convolutions. "
"Each convolution must pad the image boundaries to maintain the proper "
"sizing. More complex padding schemes can reduce artifacts at the "
"border of the image."
"\n\t http://www-cs.engr.ccny.cuny.edu/~wolberg/cs470/hw/hw2_pad.txt")
group=_("network"),
info=_(
"Use reflection padding rather than zero padding with convolutions. "
"Each convolution must pad the image boundaries to maintain the proper "
"sizing. More complex padding schemes can reduce artifacts at the "
"border of the image."
"\n\t http://www-cs.engr.ccny.cuny.edu/~wolberg/cs470/hw/hw2_pad.txt"))
self.add_item(
section=section,
title="allow_growth",
datatype=bool,
default=False,
group="network",
group=_("network"),
fixed=False,
info="[Nvidia Only]. Enable the Tensorflow GPU 'allow_growth' configuration option. "
"This option prevents Tensorflow from allocating all of the GPU VRAM at launch "
"but can lead to higher VRAM fragmentation and slower performance. Should only "
"be enabled if you are receiving errors regarding 'cuDNN fails to initialize' "
"when commencing training.")
info=_(
"[Nvidia Only]. Enable the Tensorflow GPU 'allow_growth' configuration option. "
"This option prevents Tensorflow from allocating all of the GPU VRAM at launch "
"but can lead to higher VRAM fragmentation and slower performance. Should only "
"be enabled if you are receiving errors regarding 'cuDNN fails to initialize' "
"when commencing training."))
self.add_item(
section=section,
title="mixed_precision",
datatype=bool,
default=False,
fixed=False,
group="network",
info="[Not PlaidML], NVIDIA GPUs can run operations in float16 faster than in "
"float32. Mixed precision allows you to use a mix of float16 with float32, to "
"get the performance benefits from float16 and the numeric stability benefits "
"from float32.\n\nThis is untested on DirectML backend, but will run on most "
"Nvidia models. it will only speed up training on more recent GPUs. Those with "
"compute capability 7.0 or higher will see the greatest performance benefit from "
"mixed precision because they have Tensor Cores. Older GPUs offer no math "
"performance benefit for using mixed precision, however memory and bandwidth "
"savings can enable some speedups. Generally RTX GPUs and later will offer the "
"most benefit.")
group=_("network"),
info=_(
"[Not PlaidML], NVIDIA GPUs can run operations in float16 faster than in "
"float32. Mixed precision allows you to use a mix of float16 with float32, to "
"get the performance benefits from float16 and the numeric stability benefits "
"from float32.\n\nThis is untested on DirectML backend, but will run on most "
"Nvidia models. it will only speed up training on more recent GPUs. Those with "
"compute capability 7.0 or higher will see the greatest performance benefit from "
"mixed precision because they have Tensor Cores. Older GPUs offer no math "
"performance benefit for using mixed precision, however memory and bandwidth "
"savings can enable some speedups. Generally RTX GPUs and later will offer the "
"most benefit."))
self.add_item(
section=section,
title="nan_protection",
datatype=bool,
default=True,
group="network",
info="If a 'NaN' is generated in the model, this means that the model has corrupted "
"and the model is likely to start deteriorating from this point on. Enabling NaN "
"protection will stop training immediately in the event of a NaN. The last save "
"will not contain the NaN, so you may still be able to rescue your model.",
group=_("network"),
info=_(
"If a 'NaN' is generated in the model, this means that the model has corrupted "
"and the model is likely to start deteriorating from this point on. Enabling NaN "
"protection will stop training immediately in the event of a NaN. The last save "
"will not contain the NaN, so you may still be able to rescue your model."),
fixed=False)
self.add_item(
section=section,
@ -306,11 +326,12 @@ class Config(FaceswapConfig):
min_max=(1, 32),
rounding=1,
fixed=False,
group="convert",
info="[GPU Only]. The number of faces to feed through the model at once when running "
"the Convert process.\n\nNB: Increasing this figure is unlikely to improve "
"convert speed, however, if you are getting Out of Memory errors, then you may "
"want to reduce the batch size.")
group=_("convert"),
info=_(
"[GPU Only]. The number of faces to feed through the model at once when running "
"the Convert process.\n\nNB: Increasing this figure is unlikely to improve "
"convert speed, however, if you are getting Out of Memory errors, then you may "
"want to reduce the batch size."))
def _set_loss(self) -> None:
# pylint:disable=line-too-long
@ -326,171 +347,177 @@ class Config(FaceswapConfig):
logger.debug("Setting Loss config")
section = "global.loss"
self.add_section(section,
"Loss configuration options\n"
"Loss is the mechanism by which a Neural Network judges how well it "
"thinks that it is recreating a face." + ADDITIONAL_INFO)
_("Loss configuration options\n"
"Loss is the mechanism by which a Neural Network judges how well it "
"thinks that it is recreating a face.") + ADDITIONAL_INFO)
self.add_item(
section=section,
title="loss_function",
datatype=str,
group="loss",
group=_("loss"),
default="ssim",
fixed=False,
choices=[x for x in sorted(_LOSS_HELP) if x not in _NON_PRIMARY_LOSS],
info="The loss function to use.\n\n\t" +
"\n\n\t".join(f"{k}: {v}"
for k, v in sorted(_LOSS_HELP.items())
if k not in _NON_PRIMARY_LOSS))
info=(_("The loss function to use.") +
"\n\n\t" + "\n\n\t".join(f"{k}: {v}"
for k, v in sorted(_LOSS_HELP.items())
if k not in _NON_PRIMARY_LOSS)))
self.add_item(
section=section,
title="loss_function_2",
datatype=str,
group="loss",
group=_("loss"),
default="mse",
fixed=False,
choices=list(sorted(_LOSS_HELP)),
info="The second loss function to use. If using a structural based loss (such as "
"SSIM, MS-SSIM or GMSD) it is common to add an L1 regularization(MAE) or L2 "
"regularization (MSE) function. You can adjust the weighting of this loss "
"function with the loss_weight_2 option.\n\n\t" +
"\n\n\t".join(f"{k}: {v}"
for k, v in sorted(_LOSS_HELP.items())))
info=(_("The second loss function to use. If using a structural based loss (such as "
"SSIM, MS-SSIM or GMSD) it is common to add an L1 regularization(MAE) or L2 "
"regularization (MSE) function. You can adjust the weighting of this loss "
"function with the loss_weight_2 option.") +
"\n\n\t" + "\n\n\t".join(f"{k}: {v}" for k, v in sorted(_LOSS_HELP.items()))))
self.add_item(
section=section,
title="loss_weight_2",
datatype=int,
group="loss",
group=_("loss"),
min_max=(0, 400),
rounding=1,
default=100,
fixed=False,
info="The amount of weight to apply to the second loss function.\n\n"
"\n\nThe value given here is as a percentage denoting how much the selected "
"function should contribute to the overall loss cost of the model. For example:"
"\n\t 100 - The loss calculated for the second loss function will be applied at "
"its full amount towards the overall loss score. "
"\n\t 25 - The loss calculated for the second loss function will be reduced by a "
"quarter prior to adding to the overall loss score. "
"\n\t 400 - The loss calculated for the second loss function will be mulitplied "
"4 times prior to adding to the overall loss score. "
"\n\t 0 - Disables the second loss function altogether.")
info=_(
"The amount of weight to apply to the second loss function.\n\n"
"\n\nThe value given here is as a percentage denoting how much the selected "
"function should contribute to the overall loss cost of the model. For example:"
"\n\t 100 - The loss calculated for the second loss function will be applied at "
"its full amount towards the overall loss score. "
"\n\t 25 - The loss calculated for the second loss function will be reduced by a "
"quarter prior to adding to the overall loss score. "
"\n\t 400 - The loss calculated for the second loss function will be mulitplied "
"4 times prior to adding to the overall loss score. "
"\n\t 0 - Disables the second loss function altogether."))
self.add_item(
section=section,
title="loss_function_3",
datatype=str,
group="loss",
group=_("loss"),
default="none",
fixed=False,
choices=list(sorted(_LOSS_HELP)),
info="The third loss function to use. You can adjust the weighting of this loss "
"function with the loss_weight_3 option.\n\n\t" +
"\n\n\t".join(f"{k}: {v}"
for k, v in sorted(_LOSS_HELP.items())))
info=(_("The third loss function to use. You can adjust the weighting of this loss "
"function with the loss_weight_3 option.") +
"\n\n\t" +
"\n\n\t".join(f"{k}: {v}" for k, v in sorted(_LOSS_HELP.items()))))
self.add_item(
section=section,
title="loss_weight_3",
datatype=int,
group="loss",
group=_("loss"),
min_max=(0, 400),
rounding=1,
default=0,
fixed=False,
info="The amount of weight to apply to the third loss function.\n\n"
"\n\nThe value given here is as a percentage denoting how much the selected "
"function should contribute to the overall loss cost of the model. For example:"
"\n\t 100 - The loss calculated for the third loss function will be applied at "
"its full amount towards the overall loss score. "
"\n\t 25 - The loss calculated for the third loss function will be reduced by a "
"quarter prior to adding to the overall loss score. "
"\n\t 400 - The loss calculated for the third loss function will be mulitplied 4 "
"times prior to adding to the overall loss score. "
"\n\t 0 - Disables the third loss function altogether.")
info=_(
"The amount of weight to apply to the third loss function.\n\n"
"\n\nThe value given here is as a percentage denoting how much the selected "
"function should contribute to the overall loss cost of the model. For example:"
"\n\t 100 - The loss calculated for the third loss function will be applied at "
"its full amount towards the overall loss score. "
"\n\t 25 - The loss calculated for the third loss function will be reduced by a "
"quarter prior to adding to the overall loss score. "
"\n\t 400 - The loss calculated for the third loss function will be mulitplied 4 "
"times prior to adding to the overall loss score. "
"\n\t 0 - Disables the third loss function altogether."))
self.add_item(
section=section,
title="loss_function_4",
datatype=str,
group="loss",
group=_("loss"),
default="none",
fixed=False,
choices=list(sorted(_LOSS_HELP)),
info="The fourth loss function to use. You can adjust the weighting of this loss "
"function with the loss_weight_3 option.\n\n\t" +
"\n\n\t".join(f"{k}: {v}"
for k, v in sorted(_LOSS_HELP.items())))
info=(_("The fourth loss function to use. You can adjust the weighting of this loss "
"function with the loss_weight_3 option.") +
"\n\n\t" +
"\n\n\t".join(f"{k}: {v}" for k, v in sorted(_LOSS_HELP.items()))))
self.add_item(
section=section,
title="loss_weight_4",
datatype=int,
group="loss",
group=_("loss"),
min_max=(0, 400),
rounding=1,
default=0,
fixed=False,
info="The amount of weight to apply to the fourth loss function.\n\n"
"\n\nThe value given here is as a percentage denoting how much the selected "
"function should contribute to the overall loss cost of the model. For example:"
"\n\t 100 - The loss calculated for the fourth loss function will be applied at "
"its full amount towards the overall loss score. "
"\n\t 25 - The loss calculated for the fourth loss function will be reduced by a "
"quarter prior to adding to the overall loss score. "
"\n\t 400 - The loss calculated for the fourth loss function will be mulitplied "
"4 times prior to adding to the overall loss score. "
"\n\t 0 - Disables the fourth loss function altogether.")
info=_(
"The amount of weight to apply to the fourth loss function.\n\n"
"\n\nThe value given here is as a percentage denoting how much the selected "
"function should contribute to the overall loss cost of the model. For example:"
"\n\t 100 - The loss calculated for the fourth loss function will be applied at "
"its full amount towards the overall loss score. "
"\n\t 25 - The loss calculated for the fourth loss function will be reduced by a "
"quarter prior to adding to the overall loss score. "
"\n\t 400 - The loss calculated for the fourth loss function will be mulitplied "
"4 times prior to adding to the overall loss score. "
"\n\t 0 - Disables the fourth loss function altogether."))
self.add_item(
section=section,
title="mask_loss_function",
datatype=str,
group="loss",
group=_("loss"),
default="mse",
fixed=False,
choices=["mae", "mse"],
info="The loss function to use when learning a mask."
"\n\t MAE - Mean absolute error will guide reconstructions of each pixel "
"towards its median value in the training dataset. Robust to outliers but as "
"a median, it can potentially ignore some infrequent image types in the dataset."
"\n\t MSE - Mean squared error will guide reconstructions of each pixel "
"towards its average value in the training dataset. As an average, it will be "
"susceptible to outliers and typically produces slightly blurrier results.")
info=_(
"The loss function to use when learning a mask."
"\n\t MAE - Mean absolute error will guide reconstructions of each pixel "
"towards its median value in the training dataset. Robust to outliers but as "
"a median, it can potentially ignore some infrequent image types in the dataset."
"\n\t MSE - Mean squared error will guide reconstructions of each pixel "
"towards its average value in the training dataset. As an average, it will be "
"susceptible to outliers and typically produces slightly blurrier results."))
self.add_item(
section=section,
title="eye_multiplier",
datatype=int,
group="loss",
group=_("loss"),
min_max=(1, 40),
rounding=1,
default=3,
fixed=False,
info="The amount of priority to give to the eyes.\n\nThe value given here is as a "
"multiplier of the main loss score. For example:"
"\n\t 1 - The eyes will receive the same priority as the rest of the face. "
"\n\t 10 - The eyes will be given a score 10 times higher than the rest of the "
"face."
"\n\nNB: Penalized Mask Loss must be enable to use this option.")
info=_(
"The amount of priority to give to the eyes.\n\nThe value given here is as a "
"multiplier of the main loss score. For example:"
"\n\t 1 - The eyes will receive the same priority as the rest of the face. "
"\n\t 10 - The eyes will be given a score 10 times higher than the rest of the "
"face."
"\n\nNB: Penalized Mask Loss must be enable to use this option."))
self.add_item(
section=section,
title="mouth_multiplier",
datatype=int,
group="loss",
group=_("loss"),
min_max=(1, 40),
rounding=1,
default=2,
fixed=False,
info="The amount of priority to give to the mouth.\n\nThe value given here is as a "
"multiplier of the main loss score. For Example:"
"\n\t 1 - The mouth will receive the same priority as the rest of the face. "
"\n\t 10 - The mouth will be given a score 10 times higher than the rest of the "
"face."
"\n\nNB: Penalized Mask Loss must be enable to use this option.")
info=_(
"The amount of priority to give to the mouth.\n\nThe value given here is as a "
"multiplier of the main loss score. For Example:"
"\n\t 1 - The mouth will receive the same priority as the rest of the face. "
"\n\t 10 - The mouth will be given a score 10 times higher than the rest of the "
"face."
"\n\nNB: Penalized Mask Loss must be enable to use this option."))
self.add_item(
section=section,
title="penalized_mask_loss",
datatype=bool,
default=True,
group="loss",
info="Image loss function is weighted by mask presence. For areas of "
"the image without the facial mask, reconstruction errors will be "
"ignored while the masked face area is prioritized. May increase "
"overall quality by focusing attention on the core face area.")
group=_("loss"),
info=_(
"Image loss function is weighted by mask presence. For areas of "
"the image without the facial mask, reconstruction errors will be "
"ignored while the masked face area is prioritized. May increase "
"overall quality by focusing attention on the core face area."))
self.add_item(
section=section,
title="mask_type",
@ -498,40 +525,41 @@ class Config(FaceswapConfig):
default="extended",
choices=PluginLoader.get_available_extractors("mask",
add_none=True, extend_plugin=True),
group="mask",
group=_("mask"),
gui_radio=True,
info="The mask to be used for training. If you have selected 'Learn Mask' or "
"'Penalized Mask Loss' you must select a value other than 'none'. The required "
"mask should have been selected as part of the Extract process. If it does not "
"exist in the alignments file then it will be generated prior to training "
"commencing."
"\n\tnone: Don't use a mask."
"\n\tbisenet-fp_face: Relatively lightweight NN based mask that provides more "
"refined control over the area to be masked (configurable in mask settings). "
"Use this version of bisenet-fp if your model is trained with 'face' or "
"'legacy' centering."
"\n\tbisenet-fp_head: Relatively lightweight NN based mask that provides more "
"refined control over the area to be masked (configurable in mask settings). "
"Use this version of bisenet-fp if your model is trained with 'head' centering."
"\n\tcomponents: Mask designed to provide facial segmentation based on the "
"positioning of landmark locations. A convex hull is constructed around the "
"exterior of the landmarks to create a mask."
"\n\tcustom_face: Custom user created, face centered mask."
"\n\tcustom_head: Custom user created, head centered mask."
"\n\textended: Mask designed to provide facial segmentation based on the "
"positioning of landmark locations. A convex hull is constructed around the "
"exterior of the landmarks and the mask is extended upwards onto the forehead."
"\n\tvgg-clear: Mask designed to provide smart segmentation of mostly frontal "
"faces clear of obstructions. Profile faces and obstructions may result in "
"sub-par performance."
"\n\tvgg-obstructed: Mask designed to provide smart segmentation of mostly "
"frontal faces. The mask model has been specifically trained to recognize "
"some facial obstructions (hands and eyeglasses). Profile faces may result in "
"sub-par performance."
"\n\tunet-dfl: Mask designed to provide smart segmentation of mostly frontal "
"faces. The mask model has been trained by community members and will need "
"testing for further description. Profile faces may result in sub-par "
"performance.")
info=_(
"The mask to be used for training. If you have selected 'Learn Mask' or "
"'Penalized Mask Loss' you must select a value other than 'none'. The required "
"mask should have been selected as part of the Extract process. If it does not "
"exist in the alignments file then it will be generated prior to training "
"commencing."
"\n\tnone: Don't use a mask."
"\n\tbisenet-fp_face: Relatively lightweight NN based mask that provides more "
"refined control over the area to be masked (configurable in mask settings). "
"Use this version of bisenet-fp if your model is trained with 'face' or "
"'legacy' centering."
"\n\tbisenet-fp_head: Relatively lightweight NN based mask that provides more "
"refined control over the area to be masked (configurable in mask settings). "
"Use this version of bisenet-fp if your model is trained with 'head' centering."
"\n\tcomponents: Mask designed to provide facial segmentation based on the "
"positioning of landmark locations. A convex hull is constructed around the "
"exterior of the landmarks to create a mask."
"\n\tcustom_face: Custom user created, face centered mask."
"\n\tcustom_head: Custom user created, head centered mask."
"\n\textended: Mask designed to provide facial segmentation based on the "
"positioning of landmark locations. A convex hull is constructed around the "
"exterior of the landmarks and the mask is extended upwards onto the forehead."
"\n\tvgg-clear: Mask designed to provide smart segmentation of mostly frontal "
"faces clear of obstructions. Profile faces and obstructions may result in "
"sub-par performance."
"\n\tvgg-obstructed: Mask designed to provide smart segmentation of mostly "
"frontal faces. The mask model has been specifically trained to recognize "
"some facial obstructions (hands and eyeglasses). Profile faces may result in "
"sub-par performance."
"\n\tunet-dfl: Mask designed to provide smart segmentation of mostly frontal "
"faces. The mask model has been trained by community members and will need "
"testing for further description. Profile faces may result in sub-par "
"performance."))
self.add_item(
section=section,
title="mask_blur_kernel",
@ -539,12 +567,13 @@ class Config(FaceswapConfig):
min_max=(0, 9),
rounding=1,
default=3,
group="mask",
info="Apply gaussian blur to the mask input. This has the effect of smoothing the "
"edges of the mask, which can help with poorly calculated masks and give less "
"of a hard edge to the predicted mask. The size is in pixels (calculated from "
"a 128px mask). Set to 0 to not apply gaussian blur. This value should be odd, "
"if an even number is passed in then it will be rounded to the next odd number.")
group=_("mask"),
info=_(
"Apply gaussian blur to the mask input. This has the effect of smoothing the "
"edges of the mask, which can help with poorly calculated masks and give less "
"of a hard edge to the predicted mask. The size is in pixels (calculated from "
"a 128px mask). Set to 0 to not apply gaussian blur. This value should be odd, "
"if an even number is passed in then it will be rounded to the next odd number."))
self.add_item(
section=section,
title="mask_threshold",
@ -552,15 +581,17 @@ class Config(FaceswapConfig):
default=4,
min_max=(0, 50),
rounding=1,
group="mask",
info="Sets pixels that are near white to white and near black to black. Set to 0 for "
"off.")
group=_("mask"),
info=_(
"Sets pixels that are near white to white and near black to black. Set to 0 for "
"off."))
self.add_item(
section=section,
title="learn_mask",
datatype=bool,
default=False,
group="mask",
info="Dedicate a portion of the model to learning how to duplicate the input "
"mask. Increases VRAM usage in exchange for learning a quick ability to try "
"to replicate more complex mask models.")
group=_("mask"),
info=_(
"Dedicate a portion of the model to learning how to duplicate the input "
"mask. Increases VRAM usage in exchange for learning a quick ability to try "
"to replicate more complex mask models."))