mirror of
https://github.com/iperov/DeepFaceLab.git
synced 2025-07-06 13:02:15 -07:00
638 lines
34 KiB
Python
638 lines
34 KiB
Python
from functools import partial
|
|
|
|
import numpy as np
|
|
|
|
import mathlib
|
|
from facelib import FaceType
|
|
from interact import interact as io
|
|
from models import ModelBase
|
|
from nnlib import nnlib
|
|
from samplelib import *
|
|
|
|
|
|
#SAE - Styled AutoEncoder
|
|
class SAEv2Model(ModelBase):
|
|
|
|
#override
|
|
def onInitializeOptions(self, is_first_run, ask_override):
|
|
yn_str = {True:'y',False:'n'}
|
|
|
|
default_resolution = 128
|
|
default_archi = 'df'
|
|
default_face_type = 'f'
|
|
default_learn_mask = True
|
|
|
|
if is_first_run:
|
|
resolution = io.input_int("Resolution ( 64-256 ?:help skip:128) : ", default_resolution, help_message="More resolution requires more VRAM and time to train. Value will be adjusted to multiple of 16.")
|
|
resolution = np.clip (resolution, 64, 256)
|
|
while np.modf(resolution / 16)[0] != 0.0:
|
|
resolution -= 1
|
|
self.options['resolution'] = resolution
|
|
|
|
self.options['face_type'] = io.input_str ("Half, mid full, or full face? (h/mf/f, ?:help skip:f) : ", default_face_type, ['h','mf','f'], help_message="Half face has better resolution, but covers less area of cheeks. Mid face is 30% wider than half face.").lower()
|
|
self.options['learn_mask'] = io.input_bool ( f"Learn mask? (y/n, ?:help skip:{yn_str[default_learn_mask]} ) : " , default_learn_mask, help_message="Learning mask can help model to recognize face directions. Learn without mask can reduce model size, in this case converter forced to use 'not predicted mask' that is not smooth as predicted.")
|
|
else:
|
|
self.options['resolution'] = self.options.get('resolution', default_resolution)
|
|
self.options['face_type'] = self.options.get('face_type', default_face_type)
|
|
self.options['learn_mask'] = self.options.get('learn_mask', default_learn_mask)
|
|
|
|
if (is_first_run or ask_override) and 'tensorflow' in self.device_config.backend:
|
|
def_optimizer_mode = self.options.get('optimizer_mode', 1)
|
|
self.options['optimizer_mode'] = io.input_int ("Optimizer mode? ( 1,2,3 ?:help skip:%d) : " % (def_optimizer_mode), def_optimizer_mode, help_message="1 - no changes. 2 - allows you to train x2 bigger network consuming RAM. 3 - allows you to train x3 bigger network consuming huge amount of RAM and slower, depends on CPU power.")
|
|
else:
|
|
self.options['optimizer_mode'] = self.options.get('optimizer_mode', 1)
|
|
|
|
if is_first_run:
|
|
self.options['archi'] = io.input_str ("AE architecture (df, liae ?:help skip:%s) : " % (default_archi) , default_archi, ['df','liae'], help_message="'df' keeps faces more natural. 'liae' can fix overly different face shapes.").lower() #-s version is slower, but has decreased change to collapse.
|
|
else:
|
|
self.options['archi'] = self.options.get('archi', default_archi)
|
|
|
|
default_ae_dims = 256 if 'liae' in self.options['archi'] else 512
|
|
default_ed_ch_dims = 21
|
|
|
|
if is_first_run:
|
|
self.options['ae_dims'] = np.clip ( io.input_int("AutoEncoder dims (32-1024 ?:help skip:%d) : " % (default_ae_dims) , default_ae_dims, help_message="All face information will packed to AE dims. If amount of AE dims are not enough, then for example closed eyes will not be recognized. More dims are better, but require more VRAM. You can fine-tune model size to fit your GPU." ), 32, 1024 )
|
|
self.options['ed_ch_dims'] = np.clip ( io.input_int("Encoder/Decoder dims per channel (10-85 ?:help skip:%d) : " % (default_ed_ch_dims) , default_ed_ch_dims, help_message="More dims help to recognize more facial features and achieve sharper result, but require more VRAM. You can fine-tune model size to fit your GPU." ), 10, 85 )
|
|
else:
|
|
self.options['ae_dims'] = self.options.get('ae_dims', default_ae_dims)
|
|
self.options['ed_ch_dims'] = self.options.get('ed_ch_dims', default_ed_ch_dims)
|
|
|
|
default_face_style_power = 0.0
|
|
default_bg_style_power = 0.0
|
|
if is_first_run or ask_override:
|
|
default_face_style_power = default_face_style_power if is_first_run else self.options.get('face_style_power', default_face_style_power)
|
|
self.options['face_style_power'] = np.clip ( io.input_number("Face style power ( 0.0 .. 100.0 ?:help skip:%.2f) : " % (default_face_style_power), default_face_style_power,
|
|
help_message="Learn to transfer face style details such as light and color conditions. Warning: Enable it only after 10k iters, when predicted face is clear enough to start learn style. Start from 0.1 value and check history changes. Enabling this option increases the chance of model collapse."), 0.0, 100.0 )
|
|
|
|
default_bg_style_power = default_bg_style_power if is_first_run else self.options.get('bg_style_power', default_bg_style_power)
|
|
self.options['bg_style_power'] = np.clip ( io.input_number("Background style power ( 0.0 .. 100.0 ?:help skip:%.2f) : " % (default_bg_style_power), default_bg_style_power,
|
|
help_message="Learn to transfer image around face. This can make face more like dst. Enabling this option increases the chance of model collapse."), 0.0, 100.0 )
|
|
|
|
default_apply_random_ct = False if is_first_run else self.options.get('apply_random_ct', False)
|
|
self.options['apply_random_ct'] = io.input_bool (f"Apply random color transfer to src faceset? (y/n, ?:help skip:{yn_str[default_apply_random_ct]}) : ", default_apply_random_ct, help_message="Increase variativity of src samples by apply LCT color transfer from random dst samples. It is like 'face_style' learning, but more precise color transfer and without risk of model collapse, also it does not require additional GPU resources, but the training time may be longer, due to the src faceset is becoming more diverse.")
|
|
|
|
default_true_face_training = False if is_first_run else self.options.get('true_face_training', False)
|
|
self.options['true_face_training'] = io.input_bool (f"Enable 'true face' training? (y/n, ?:help skip:{yn_str[default_true_face_training]}) : ", default_true_face_training, help_message="The result face will be more like src and will get extra sharpness. Enable it for last 10-20k iterations before conversion.")
|
|
|
|
if nnlib.device.backend != 'plaidML': # todo https://github.com/plaidml/plaidml/issues/301
|
|
default_clipgrad = False if is_first_run else self.options.get('clipgrad', False)
|
|
self.options['clipgrad'] = io.input_bool (f"Enable gradient clipping? (y/n, ?:help skip:{yn_str[default_clipgrad]}) : ", default_clipgrad, help_message="Gradient clipping reduces chance of model collapse, sacrificing speed of training.")
|
|
else:
|
|
self.options['clipgrad'] = False
|
|
|
|
else:
|
|
self.options['face_style_power'] = self.options.get('face_style_power', default_face_style_power)
|
|
self.options['bg_style_power'] = self.options.get('bg_style_power', default_bg_style_power)
|
|
self.options['apply_random_ct'] = self.options.get('apply_random_ct', False)
|
|
self.options['clipgrad'] = self.options.get('clipgrad', False)
|
|
|
|
if is_first_run:
|
|
self.options['pretrain'] = io.input_bool ("Pretrain the model? (y/n, ?:help skip:n) : ", False, help_message="Pretrain the model with large amount of various faces. This technique may help to train the fake with overly different face shapes and light conditions of src/dst data. Face will be look more like a morphed. To reduce the morph effect, some model files will be initialized but not be updated after pretrain: LIAE: inter_AB.h5 DF: encoder.h5. The longer you pretrain the model the more morphed face will look. After that, save and run the training again.")
|
|
else:
|
|
self.options['pretrain'] = False
|
|
|
|
#override
|
|
def onInitialize(self):
|
|
exec(nnlib.import_all(), locals(), globals())
|
|
self.set_vram_batch_requirements({1.5:4,4:8})
|
|
|
|
resolution = self.options['resolution']
|
|
learn_mask = self.options['learn_mask']
|
|
|
|
ae_dims = self.options['ae_dims']
|
|
ed_ch_dims = self.options['ed_ch_dims']
|
|
self.pretrain = self.options['pretrain'] = self.options.get('pretrain', False)
|
|
if not self.pretrain:
|
|
self.options.pop('pretrain')
|
|
|
|
bgr_shape = (resolution, resolution, 3)
|
|
mask_shape = (resolution, resolution, 1)
|
|
|
|
apply_random_ct = self.options.get('apply_random_ct', False)
|
|
self.true_face_training = self.options.get('true_face_training', False)
|
|
masked_training = True
|
|
|
|
class CommonModel(object):
|
|
def downscale (self, dim, kernel_size=5, dilation_rate=1, use_activator=True):
|
|
def func(x):
|
|
if not use_activator:
|
|
return SubpixelDownscaler()(Conv2D(dim // 4, kernel_size=kernel_size, strides=1, dilation_rate=dilation_rate, padding='same')(x))
|
|
else:
|
|
return SubpixelDownscaler()(LeakyReLU(0.1)(Conv2D(dim // 4, kernel_size=kernel_size, strides=1, dilation_rate=dilation_rate, padding='same')(x)))
|
|
return func
|
|
|
|
def upscale (self, dim, size=(2,2)):
|
|
def func(x):
|
|
return SubpixelUpscaler(size=size)(LeakyReLU(0.1)(Conv2D(dim * np.prod(size) , kernel_size=3, strides=1, padding='same')(x)))
|
|
return func
|
|
|
|
def ResidualBlock(self, dim):
|
|
def func(inp):
|
|
x = Conv2D(dim, kernel_size=3, padding='same')(inp)
|
|
x = LeakyReLU(0.2)(x)
|
|
x = Conv2D(dim, kernel_size=3, padding='same')(x)
|
|
x = Add()([x, inp])
|
|
x = LeakyReLU(0.2)(x)
|
|
return x
|
|
return func
|
|
|
|
class SAEDFModel(CommonModel):
|
|
def __init__(self, resolution, ae_dims, e_ch_dims, d_ch_dims, learn_mask):
|
|
super().__init__()
|
|
self.learn_mask = learn_mask
|
|
|
|
output_nc = 3
|
|
bgr_shape = (resolution, resolution, output_nc)
|
|
mask_shape = (resolution, resolution, 1)
|
|
lowest_dense_res = resolution // 16
|
|
e_dims = output_nc*e_ch_dims
|
|
|
|
|
|
|
|
def enc_flow(e_ch_dims, ae_dims, lowest_dense_res):
|
|
dims = output_nc * e_ch_dims
|
|
if dims % 2 != 0:
|
|
dims += 1
|
|
|
|
def func(inp):
|
|
x = self.downscale(dims , 3, 1 )(inp)
|
|
x = self.downscale(dims*2, 3, 1 )(x)
|
|
x = self.downscale(dims*4, 3, 1 )(x)
|
|
x0 = self.downscale(dims*8, 3, 1 )(x)
|
|
|
|
x = self.downscale(dims , 5, 1 )(inp)
|
|
x = self.downscale(dims*2, 5, 1 )(x)
|
|
x = self.downscale(dims*4, 5, 1 )(x)
|
|
x1 = self.downscale(dims*8, 5, 1 )(x)
|
|
|
|
x = self.downscale(dims , 5, 2 )(inp)
|
|
x = self.downscale(dims*2, 5, 2 )(x)
|
|
x = self.downscale(dims*4, 5, 2 )(x)
|
|
x2 = self.downscale(dims*8, 5, 2 )(x)
|
|
|
|
x = self.downscale(dims , 7, 2 )(inp)
|
|
x = self.downscale(dims*2, 7, 2 )(x)
|
|
x = self.downscale(dims*4, 7, 2 )(x)
|
|
x3 = self.downscale(dims*8, 7, 2 )(x)
|
|
|
|
x = Concatenate()([x0,x1,x2,x3])
|
|
|
|
x = Dense(ae_dims)(Flatten()(x))
|
|
x = Dense(lowest_dense_res * lowest_dense_res * ae_dims)(x)
|
|
x = Reshape((lowest_dense_res, lowest_dense_res, ae_dims))(x)
|
|
x = self.upscale(ae_dims)(x)
|
|
return x
|
|
return func
|
|
|
|
def dec_flow(output_nc, d_ch_dims, is_mask=False):
|
|
dims = output_nc * d_ch_dims
|
|
if dims % 2 != 0:
|
|
dims += 1
|
|
|
|
|
|
|
|
def func(x):
|
|
|
|
for i in [8,4,2]:
|
|
x = self.upscale(dims*i)(x)
|
|
|
|
if not is_mask:
|
|
x0 = x
|
|
x = self.upscale( (dims*i)//2 )(x)
|
|
x = self.ResidualBlock( (dims*i)//2 )(x)
|
|
x = self.downscale( dims*i, use_activator=False ) (x)
|
|
x = Add()([x, x0])
|
|
x = LeakyReLU(0.2)(x)
|
|
|
|
return Conv2D(output_nc, kernel_size=5, padding='same', activation='sigmoid')(x)
|
|
|
|
return func
|
|
|
|
self.encoder = modelify(enc_flow(e_ch_dims, ae_dims, lowest_dense_res)) ( Input(bgr_shape) )
|
|
|
|
sh = K.int_shape( self.encoder.outputs[0] )[1:]
|
|
self.decoder_src = modelify(dec_flow(output_nc, d_ch_dims)) ( Input(sh) )
|
|
self.decoder_dst = modelify(dec_flow(output_nc, d_ch_dims)) ( Input(sh) )
|
|
|
|
if learn_mask:
|
|
self.decoder_srcm = modelify(dec_flow(1, d_ch_dims, is_mask=True)) ( Input(sh) )
|
|
self.decoder_dstm = modelify(dec_flow(1, d_ch_dims, is_mask=True)) ( Input(sh) )
|
|
|
|
self.src_dst_trainable_weights = self.encoder.trainable_weights + self.decoder_src.trainable_weights + self.decoder_dst.trainable_weights
|
|
|
|
if learn_mask:
|
|
self.src_dst_mask_trainable_weights = self.encoder.trainable_weights + self.decoder_srcm.trainable_weights + self.decoder_dstm.trainable_weights
|
|
|
|
self.warped_src, self.warped_dst = Input(bgr_shape), Input(bgr_shape)
|
|
self.target_src, self.target_dst = Input(bgr_shape), Input(bgr_shape)
|
|
self.target_srcm, self.target_dstm = Input(mask_shape), Input(mask_shape)
|
|
self.src_code, self.dst_code = self.encoder(self.warped_src), self.encoder(self.warped_dst)
|
|
|
|
self.pred_src_src = self.decoder_src(self.src_code)
|
|
self.pred_dst_dst = self.decoder_dst(self.dst_code)
|
|
self.pred_src_dst = self.decoder_src(self.dst_code)
|
|
|
|
if learn_mask:
|
|
self.pred_src_srcm = self.decoder_srcm(self.src_code)
|
|
self.pred_dst_dstm = self.decoder_dstm(self.dst_code)
|
|
self.pred_src_dstm = self.decoder_srcm(self.dst_code)
|
|
|
|
def get_model_filename_list(self, exclude_for_pretrain=False):
|
|
ar = []
|
|
if not exclude_for_pretrain:
|
|
ar += [ [self.encoder, 'encoder.h5'] ]
|
|
ar += [ [self.decoder_src, 'decoder_src.h5'],
|
|
[self.decoder_dst, 'decoder_dst.h5'] ]
|
|
if self.learn_mask:
|
|
ar += [ [self.decoder_srcm, 'decoder_srcm.h5'],
|
|
[self.decoder_dstm, 'decoder_dstm.h5'] ]
|
|
return ar
|
|
|
|
class SAELIAEModel(CommonModel):
|
|
def __init__(self, resolution, ae_dims, e_ch_dims, d_ch_dims, learn_mask):
|
|
super().__init__()
|
|
self.learn_mask = learn_mask
|
|
|
|
output_nc = 3
|
|
bgr_shape = (resolution, resolution, output_nc)
|
|
mask_shape = (resolution, resolution, 1)
|
|
|
|
lowest_dense_res = resolution // 16
|
|
|
|
def enc_flow(e_ch_dims):
|
|
dims = output_nc*e_ch_dims
|
|
if dims % 2 != 0:
|
|
dims += 1
|
|
|
|
def func(inp):
|
|
x = self.downscale(dims , 3, 1 )(inp)
|
|
x = self.downscale(dims*2, 3, 1 )(x)
|
|
x = self.downscale(dims*4, 3, 1 )(x)
|
|
x0 = self.downscale(dims*8, 3, 1 )(x)
|
|
|
|
x = self.downscale(dims , 5, 1 )(inp)
|
|
x = self.downscale(dims*2, 5, 1 )(x)
|
|
x = self.downscale(dims*4, 5, 1 )(x)
|
|
x1 = self.downscale(dims*8, 5, 1 )(x)
|
|
|
|
x = self.downscale(dims , 5, 2 )(inp)
|
|
x = self.downscale(dims*2, 5, 2 )(x)
|
|
x = self.downscale(dims*4, 5, 2 )(x)
|
|
x2 = self.downscale(dims*8, 5, 2 )(x)
|
|
|
|
x = self.downscale(dims , 7, 2 )(inp)
|
|
x = self.downscale(dims*2, 7, 2 )(x)
|
|
x = self.downscale(dims*4, 7, 2 )(x)
|
|
x3 = self.downscale(dims*8, 7, 2 )(x)
|
|
|
|
x = Concatenate()([x0,x1,x2,x3])
|
|
|
|
x = Flatten()(x)
|
|
return x
|
|
return func
|
|
|
|
def inter_flow(lowest_dense_res, ae_dims):
|
|
def func(x):
|
|
x = Dense(ae_dims)(x)
|
|
x = Dense(lowest_dense_res * lowest_dense_res * ae_dims*2)(x)
|
|
x = Reshape((lowest_dense_res, lowest_dense_res, ae_dims*2))(x)
|
|
x = self.upscale(ae_dims*2)(x)
|
|
return x
|
|
return func
|
|
|
|
def dec_flow(output_nc, d_ch_dims, is_mask=False):
|
|
dims = output_nc * d_ch_dims
|
|
if dims % 2 != 0:
|
|
dims += 1
|
|
|
|
def func(x):
|
|
|
|
for i in [8,4,2]:
|
|
x = self.upscale(dims*i)(x)
|
|
|
|
if not is_mask:
|
|
x0 = x
|
|
x = self.upscale( (dims*i)//2 )(x)
|
|
x = self.ResidualBlock( (dims*i)//2 )(x)
|
|
x = self.downscale( dims*i, use_activator=False ) (x)
|
|
x = Add()([x, x0])
|
|
x = LeakyReLU(0.2)(x)
|
|
|
|
return Conv2D(output_nc, kernel_size=5, padding='same', activation='sigmoid')(x)
|
|
|
|
return func
|
|
|
|
self.encoder = modelify(enc_flow(e_ch_dims)) ( Input(bgr_shape) )
|
|
|
|
sh = K.int_shape( self.encoder.outputs[0] )[1:]
|
|
self.inter_B = modelify(inter_flow(lowest_dense_res, ae_dims)) ( Input(sh) )
|
|
self.inter_AB = modelify(inter_flow(lowest_dense_res, ae_dims)) ( Input(sh) )
|
|
|
|
sh = np.array(K.int_shape( self.inter_B.outputs[0] )[1:])*(1,1,2)
|
|
self.decoder = modelify(dec_flow(output_nc, d_ch_dims)) ( Input(sh) )
|
|
|
|
if learn_mask:
|
|
self.decoderm = modelify(dec_flow(1, d_ch_dims, is_mask=True)) ( Input(sh) )
|
|
|
|
self.src_dst_trainable_weights = self.encoder.trainable_weights + self.inter_B.trainable_weights + self.inter_AB.trainable_weights + self.decoder.trainable_weights
|
|
|
|
if learn_mask:
|
|
self.src_dst_mask_trainable_weights = self.encoder.trainable_weights + self.inter_B.trainable_weights + self.inter_AB.trainable_weights + self.decoderm.trainable_weights
|
|
|
|
self.warped_src, self.warped_dst = Input(bgr_shape), Input(bgr_shape)
|
|
self.target_src, self.target_dst = Input(bgr_shape), Input(bgr_shape)
|
|
self.target_srcm, self.target_dstm = Input(mask_shape), Input(mask_shape)
|
|
|
|
warped_src_code = self.encoder (self.warped_src)
|
|
warped_src_inter_AB_code = self.inter_AB (warped_src_code)
|
|
self.src_code = Concatenate()([warped_src_inter_AB_code,warped_src_inter_AB_code])
|
|
|
|
warped_dst_code = self.encoder (self.warped_dst)
|
|
warped_dst_inter_B_code = self.inter_B (warped_dst_code)
|
|
warped_dst_inter_AB_code = self.inter_AB (warped_dst_code)
|
|
self.dst_code = Concatenate()([warped_dst_inter_B_code,warped_dst_inter_AB_code])
|
|
|
|
src_dst_code = Concatenate()([warped_dst_inter_AB_code,warped_dst_inter_AB_code])
|
|
|
|
self.pred_src_src = self.decoder(self.src_code)
|
|
self.pred_dst_dst = self.decoder(self.dst_code)
|
|
self.pred_src_dst = self.decoder(src_dst_code)
|
|
|
|
if learn_mask:
|
|
self.pred_src_srcm = self.decoderm(self.src_code)
|
|
self.pred_dst_dstm = self.decoderm(self.dst_code)
|
|
self.pred_src_dstm = self.decoderm(src_dst_code)
|
|
|
|
def get_model_filename_list(self, exclude_for_pretrain=False):
|
|
ar = [ [self.encoder, 'encoder.h5'],
|
|
[self.inter_B, 'inter_B.h5'] ]
|
|
|
|
if not exclude_for_pretrain:
|
|
ar += [ [self.inter_AB, 'inter_AB.h5'] ]
|
|
|
|
ar += [ [self.decoder, 'decoder.h5'] ]
|
|
|
|
if self.learn_mask:
|
|
ar += [ [self.decoderm, 'decoderm.h5'] ]
|
|
|
|
return ar
|
|
|
|
if 'df' in self.options['archi']:
|
|
self.model = SAEDFModel (resolution, ae_dims, ed_ch_dims, ed_ch_dims, learn_mask)
|
|
elif 'liae' in self.options['archi']:
|
|
self.model = SAELIAEModel (resolution, ae_dims, ed_ch_dims, ed_ch_dims, learn_mask)
|
|
|
|
self.opt_dis_model = []
|
|
|
|
if self.true_face_training:
|
|
def dis_flow(ndf=256):
|
|
def func(x):
|
|
x, = x
|
|
|
|
code_res = K.int_shape(x)[1]
|
|
|
|
x = Conv2D( ndf, 4, strides=2, padding='valid')( ZeroPadding2D(1)(x) )
|
|
x = LeakyReLU(0.1)(x)
|
|
|
|
x = Conv2D( ndf*2, 3, strides=2, padding='valid')( ZeroPadding2D(1)(x) )
|
|
x = LeakyReLU(0.1)(x)
|
|
|
|
if code_res > 8:
|
|
x = Conv2D( ndf*4, 3, strides=2, padding='valid')( ZeroPadding2D(1)(x) )
|
|
x = LeakyReLU(0.1)(x)
|
|
|
|
if code_res > 16:
|
|
x = Conv2D( ndf*8, 3, strides=2, padding='valid')( ZeroPadding2D(1)(x) )
|
|
x = LeakyReLU(0.1)(x)
|
|
|
|
if code_res > 32:
|
|
x = Conv2D( ndf*8, 3, strides=2, padding='valid')( ZeroPadding2D(1)(x) )
|
|
x = LeakyReLU(0.1)(x)
|
|
|
|
return Conv2D( 1, 1, strides=1, padding='valid', activation='sigmoid')(x)
|
|
return func
|
|
|
|
sh = [ Input( K.int_shape(self.model.src_code)[1:] ) ]
|
|
self.dis = modelify(dis_flow()) (sh)
|
|
|
|
self.opt_dis_model = [ (self.dis, 'dis.h5') ]
|
|
|
|
loaded, not_loaded = [], self.model.get_model_filename_list()+self.opt_dis_model
|
|
if not self.is_first_run():
|
|
loaded, not_loaded = self.load_weights_safe(not_loaded)
|
|
|
|
CA_models = [ model for model, _ in not_loaded ]
|
|
|
|
self.CA_conv_weights_list = []
|
|
for model in CA_models:
|
|
for layer in model.layers:
|
|
if type(layer) == keras.layers.Conv2D:
|
|
self.CA_conv_weights_list += [layer.weights[0]] #- is Conv2D kernel_weights
|
|
|
|
target_srcm = gaussian_blur( max(1, resolution // 32) )(self.model.target_srcm)
|
|
target_dstm = gaussian_blur( max(1, resolution // 32) )(self.model.target_dstm)
|
|
|
|
target_src_masked = self.model.target_src*target_srcm
|
|
target_dst_masked = self.model.target_dst*target_dstm
|
|
target_dst_anti_masked = self.model.target_dst*(1.0 - target_dstm)
|
|
|
|
target_src_masked_opt = target_src_masked if masked_training else self.model.target_src
|
|
target_dst_masked_opt = target_dst_masked if masked_training else self.model.target_dst
|
|
|
|
pred_src_src_masked_opt = self.model.pred_src_src*target_srcm if masked_training else self.model.pred_src_src
|
|
pred_dst_dst_masked_opt = self.model.pred_dst_dst*target_dstm if masked_training else self.model.pred_dst_dst
|
|
|
|
psd_target_dst_masked = self.model.pred_src_dst*target_dstm
|
|
psd_target_dst_anti_masked = self.model.pred_src_dst*(1.0 - target_dstm)
|
|
|
|
if self.is_training_mode:
|
|
self.src_dst_opt = RMSprop(lr=5e-5, clipnorm=1.0 if self.options['clipgrad'] else 0.0, tf_cpu_mode=self.options['optimizer_mode']-1)
|
|
self.src_dst_mask_opt = RMSprop(lr=5e-5, clipnorm=1.0 if self.options['clipgrad'] else 0.0, tf_cpu_mode=self.options['optimizer_mode']-1)
|
|
self.D_opt = RMSprop(lr=5e-5, clipnorm=1.0 if self.options['clipgrad'] else 0.0, tf_cpu_mode=self.options['optimizer_mode']-1)
|
|
|
|
src_loss = K.mean ( 10*dssim(kernel_size=int(resolution/11.6),max_value=1.0)( target_src_masked_opt, pred_src_src_masked_opt) )
|
|
src_loss += K.mean ( 10*K.square( target_src_masked_opt - pred_src_src_masked_opt ) )
|
|
|
|
face_style_power = self.options['face_style_power'] / 100.0
|
|
if face_style_power != 0:
|
|
src_loss += style_loss(gaussian_blur_radius=resolution//16, loss_weight=face_style_power, wnd_size=0)( psd_target_dst_masked, target_dst_masked )
|
|
|
|
bg_style_power = self.options['bg_style_power'] / 100.0
|
|
if bg_style_power != 0:
|
|
src_loss += K.mean( (10*bg_style_power)*dssim(kernel_size=int(resolution/11.6),max_value=1.0)( psd_target_dst_anti_masked, target_dst_anti_masked ))
|
|
src_loss += K.mean( (10*bg_style_power)*K.square( psd_target_dst_anti_masked - target_dst_anti_masked ))
|
|
|
|
dst_loss = K.mean( 10*dssim(kernel_size=int(resolution/11.6),max_value=1.0)(target_dst_masked_opt, pred_dst_dst_masked_opt) )
|
|
dst_loss += K.mean( 10*K.square( target_dst_masked_opt - pred_dst_dst_masked_opt ) )
|
|
|
|
G_loss = src_loss+dst_loss
|
|
|
|
if self.true_face_training:
|
|
def DLoss(labels,logits):
|
|
return K.mean(K.binary_crossentropy(labels,logits))
|
|
|
|
src_code_d = self.dis( self.model.src_code )
|
|
src_code_d_ones = K.ones_like(src_code_d)
|
|
src_code_d_zeros = K.zeros_like(src_code_d)
|
|
dst_code_d = self.dis( self.model.dst_code )
|
|
dst_code_d_ones = K.ones_like(dst_code_d)
|
|
G_loss += 0.01*DLoss(src_code_d_ones, src_code_d)
|
|
|
|
loss_D = (DLoss(dst_code_d_ones , dst_code_d) + \
|
|
DLoss(src_code_d_zeros, src_code_d) ) * 0.5
|
|
|
|
self.D_train = K.function ([self.model.warped_src, self.model.warped_dst],[loss_D], self.D_opt.get_updates(loss_D, self.dis.trainable_weights) )
|
|
|
|
self.src_dst_train = K.function ([self.model.warped_src, self.model.warped_dst, self.model.target_src, self.model.target_srcm, self.model.target_dst, self.model.target_dstm],
|
|
[src_loss,dst_loss],
|
|
self.src_dst_opt.get_updates( G_loss, self.model.src_dst_trainable_weights)
|
|
)
|
|
|
|
if self.options['learn_mask']:
|
|
src_mask_loss = K.mean(K.square(self.model.target_srcm-self.model.pred_src_srcm))
|
|
dst_mask_loss = K.mean(K.square(self.model.target_dstm-self.model.pred_dst_dstm))
|
|
self.src_dst_mask_train = K.function ([self.model.warped_src, self.model.warped_dst, self.model.target_srcm, self.model.target_dstm],[src_mask_loss, dst_mask_loss], self.src_dst_mask_opt.get_updates(src_mask_loss+dst_mask_loss, self.model.src_dst_mask_trainable_weights ) )
|
|
|
|
if self.options['learn_mask']:
|
|
self.AE_view = K.function ([self.model.warped_src, self.model.warped_dst], [self.model.pred_src_src, self.model.pred_dst_dst, self.model.pred_dst_dstm, self.model.pred_src_dst, self.model.pred_src_dstm])
|
|
else:
|
|
self.AE_view = K.function ([self.model.warped_src, self.model.warped_dst], [self.model.pred_src_src, self.model.pred_dst_dst, self.model.pred_src_dst ])
|
|
|
|
else:
|
|
if self.options['learn_mask']:
|
|
self.AE_convert = K.function ([self.model.warped_dst],[ self.model.pred_src_dst, self.model.pred_dst_dstm, self.model.pred_src_dstm ])
|
|
else:
|
|
self.AE_convert = K.function ([self.model.warped_dst],[ self.model.pred_src_dst ])
|
|
|
|
|
|
if self.is_training_mode:
|
|
t = SampleProcessor.Types
|
|
|
|
if self.options['face_type'] == 'h':
|
|
face_type = t.FACE_TYPE_HALF
|
|
elif self.options['face_type'] == 'mf':
|
|
face_type = t.FACE_TYPE_MID_FULL
|
|
elif self.options['face_type'] == 'f':
|
|
face_type = t.FACE_TYPE_FULL
|
|
|
|
t_mode_bgr = t.MODE_BGR if not self.pretrain else t.MODE_BGR_SHUFFLE
|
|
|
|
training_data_src_path = self.training_data_src_path
|
|
training_data_dst_path = self.training_data_dst_path
|
|
sort_by_yaw = self.sort_by_yaw
|
|
|
|
if self.pretrain and self.pretraining_data_path is not None:
|
|
training_data_src_path = self.pretraining_data_path
|
|
training_data_dst_path = self.pretraining_data_path
|
|
sort_by_yaw = False
|
|
|
|
self.set_training_data_generators ([
|
|
SampleGeneratorFace(training_data_src_path, sort_by_yaw_target_samples_path=training_data_dst_path if sort_by_yaw else None,
|
|
random_ct_samples_path=training_data_dst_path if apply_random_ct else None,
|
|
debug=self.is_debug(), batch_size=self.batch_size,
|
|
sample_process_options=SampleProcessor.Options(random_flip=self.random_flip, scale_range=np.array([-0.05, 0.05])+self.src_scale_mod / 100.0 ),
|
|
output_sample_types = [ {'types' : (t.IMG_WARPED_TRANSFORMED, face_type, t_mode_bgr), 'resolution':resolution, 'apply_ct': apply_random_ct},
|
|
{'types' : (t.IMG_TRANSFORMED, face_type, t_mode_bgr), 'resolution': resolution, 'apply_ct': apply_random_ct },
|
|
{'types' : (t.IMG_TRANSFORMED, face_type, t.MODE_M), 'resolution': resolution } ]
|
|
),
|
|
|
|
SampleGeneratorFace(training_data_dst_path, debug=self.is_debug(), batch_size=self.batch_size,
|
|
sample_process_options=SampleProcessor.Options(random_flip=self.random_flip, ),
|
|
output_sample_types = [ {'types' : (t.IMG_WARPED_TRANSFORMED, face_type, t_mode_bgr), 'resolution':resolution},
|
|
{'types' : (t.IMG_TRANSFORMED, face_type, t_mode_bgr), 'resolution': resolution},
|
|
{'types' : (t.IMG_TRANSFORMED, face_type, t.MODE_M), 'resolution': resolution} ])
|
|
])
|
|
|
|
#override
|
|
def get_model_filename_list(self):
|
|
return self.model.get_model_filename_list ( exclude_for_pretrain=(self.pretrain and self.iter != 0) ) +self.opt_dis_model
|
|
|
|
#override
|
|
def onSave(self):
|
|
self.save_weights_safe( self.get_model_filename_list()+self.opt_dis_model )
|
|
|
|
#override
|
|
def on_success_train_one_iter(self):
|
|
if len(self.CA_conv_weights_list) != 0:
|
|
exec(nnlib.import_all(), locals(), globals())
|
|
CAInitializerMP ( self.CA_conv_weights_list )
|
|
self.CA_conv_weights_list = []
|
|
|
|
#override
|
|
def onTrainOneIter(self, generators_samples, generators_list):
|
|
warped_src, target_src, target_srcm = generators_samples[0]
|
|
warped_dst, target_dst, target_dstm = generators_samples[1]
|
|
|
|
feed = [warped_src, warped_dst, target_src, target_srcm, target_dst, target_dstm]
|
|
|
|
src_loss, dst_loss, = self.src_dst_train (feed)
|
|
|
|
if self.true_face_training:
|
|
self.D_train([warped_src, warped_dst])
|
|
|
|
if self.options['learn_mask']:
|
|
feed = [ warped_src, warped_dst, target_srcm, target_dstm ]
|
|
src_mask_loss, dst_mask_loss, = self.src_dst_mask_train (feed)
|
|
|
|
return ( ('src_loss', src_loss), ('dst_loss', dst_loss), )
|
|
|
|
#override
|
|
def onGetPreview(self, sample):
|
|
test_S = sample[0][1][0:4] #first 4 samples
|
|
test_S_m = sample[0][2][0:4] #first 4 samples
|
|
test_D = sample[1][1][0:4]
|
|
test_D_m = sample[1][2][0:4]
|
|
|
|
if self.options['learn_mask']:
|
|
S, D, SS, DD, DDM, SD, SDM = [ np.clip(x, 0.0, 1.0) for x in ([test_S,test_D] + self.AE_view ([test_S, test_D]) ) ]
|
|
DDM, SDM, = [ np.repeat (x, (3,), -1) for x in [DDM, SDM] ]
|
|
else:
|
|
S, D, SS, DD, SD, = [ np.clip(x, 0.0, 1.0) for x in ([test_S,test_D] + self.AE_view ([test_S, test_D]) ) ]
|
|
|
|
result = []
|
|
st = []
|
|
for i in range(len(test_S)):
|
|
ar = S[i], SS[i], D[i], DD[i], SD[i]
|
|
|
|
st.append ( np.concatenate ( ar, axis=1) )
|
|
|
|
result += [ ('SAEHD', np.concatenate (st, axis=0 )), ]
|
|
|
|
if self.options['learn_mask']:
|
|
st_m = []
|
|
for i in range(len(test_S)):
|
|
ar = S[i]*test_S_m[i], SS[i], D[i]*test_D_m[i], DD[i]*DDM[i], SD[i]*(DDM[i]*SDM[i])
|
|
st_m.append ( np.concatenate ( ar, axis=1) )
|
|
|
|
result += [ ('SAEHD masked', np.concatenate (st_m, axis=0 )), ]
|
|
|
|
return result
|
|
|
|
def predictor_func (self, face=None, dummy_predict=False):
|
|
if dummy_predict:
|
|
self.AE_convert ([ np.zeros ( (1, self.options['resolution'], self.options['resolution'], 3), dtype=np.float32 ) ])
|
|
else:
|
|
if self.options['learn_mask']:
|
|
bgr, mask_dst_dstm, mask_src_dstm = self.AE_convert ([face[np.newaxis,...]])
|
|
mask = mask_dst_dstm[0] * mask_src_dstm[0]
|
|
return bgr[0], mask[...,0]
|
|
else:
|
|
bgr, = self.AE_convert ([face[np.newaxis,...]])
|
|
return bgr[0]
|
|
|
|
#override
|
|
def get_ConverterConfig(self):
|
|
if self.options['face_type'] == 'h':
|
|
face_type = FaceType.HALF
|
|
elif self.options['face_type'] == 'mf':
|
|
face_type = FaceType.MID_FULL
|
|
elif self.options['face_type'] == 'f':
|
|
face_type = FaceType.FULL
|
|
|
|
import converters
|
|
return self.predictor_func, (self.options['resolution'], self.options['resolution'], 3), converters.ConverterConfigMasked(face_type=face_type,
|
|
default_mode = 1 if self.options['apply_random_ct'] or self.options['face_style_power'] or self.options['bg_style_power'] else 4,
|
|
clip_hborder_mask_per=0.0625 if (face_type != FaceType.HALF) else 0,
|
|
)
|
|
|
|
Model = SAEv2Model
|