|
diff --git a/core/leras/models/PatchDiscriminator.py b/core/leras/models/PatchDiscriminator.py
index f13f6ec..343e000 100644
--- a/core/leras/models/PatchDiscriminator.py
+++ b/core/leras/models/PatchDiscriminator.py
@@ -94,7 +94,7 @@ class UNetPatchDiscriminator(nn.ModelBase):
ts *= s
return rf
- def find_archi(self, target_patch_size, max_layers=6):
+ def find_archi(self, target_patch_size, max_layers=9):
"""
Find the best configuration of layers using only 3x3 convs for target patch size
"""
@@ -106,12 +106,12 @@ class UNetPatchDiscriminator(nn.ModelBase):
layers = []
sum_st = 0
+ layers.append ( [3, 2])
+ sum_st += 2
for i in range(layers_count-1):
st = 1 + (1 if val & (1 << i) !=0 else 0 )
layers.append ( [3, st ])
- sum_st += st
- layers.append ( [3, 2])
- sum_st += 2
+ sum_st += st
rf = self.calc_receptive_field_size(layers)
@@ -130,7 +130,8 @@ class UNetPatchDiscriminator(nn.ModelBase):
q=x[np.abs(np.array(x)-target_patch_size).argmin()]
return s[q][2]
- def on_build(self, patch_size, in_ch):
+ def on_build(self, patch_size, in_ch, base_ch = 16):
+
class ResidualBlock(nn.ModelBase):
def on_build(self, ch, kernel_size=3 ):
self.conv1 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME')
@@ -145,12 +146,13 @@ class UNetPatchDiscriminator(nn.ModelBase):
prev_ch = in_ch
self.convs = []
- self.res = []
+ self.res1 = []
+ self.res2 = []
self.upconvs = []
- self.upres = []
+ self.upres1 = []
+ self.upres2 = []
layers = self.find_archi(patch_size)
- base_ch = 16
-
+
level_chs = { i-1:v for i,v in enumerate([ min( base_ch * (2**i), 512 ) for i in range(len(layers)+1)]) }
self.in_conv = nn.Conv2D( in_ch, level_chs[-1], kernel_size=1, padding='VALID')
@@ -158,12 +160,14 @@ class UNetPatchDiscriminator(nn.ModelBase):
for i, (kernel_size, strides) in enumerate(layers):
self.convs.append ( nn.Conv2D( level_chs[i-1], level_chs[i], kernel_size=kernel_size, strides=strides, padding='SAME') )
- self.res.append ( ResidualBlock(level_chs[i]) )
-
+ self.res1.append ( ResidualBlock(level_chs[i]) )
+ self.res2.append ( ResidualBlock(level_chs[i]) )
+
self.upconvs.insert (0, nn.Conv2DTranspose( level_chs[i]*(2 if i != len(layers)-1 else 1), level_chs[i-1], kernel_size=kernel_size, strides=strides, padding='SAME') )
- self.upres.insert (0, ResidualBlock(level_chs[i-1]*2) )
-
+ self.upres1.insert (0, ResidualBlock(level_chs[i-1]*2) )
+ self.upres2.insert (0, ResidualBlock(level_chs[i-1]*2) )
+
self.out_conv = nn.Conv2D( level_chs[-1]*2, 1, kernel_size=1, padding='VALID')
self.center_out = nn.Conv2D( level_chs[len(layers)-1], 1, kernel_size=1, padding='VALID')
@@ -171,20 +175,22 @@ class UNetPatchDiscriminator(nn.ModelBase):
def forward(self, x):
- x = tf.nn.leaky_relu( self.in_conv(x), 0.1 )
+ x = tf.nn.leaky_relu( self.in_conv(x), 0.2 )
encs = []
- for conv, res in zip(self.convs, self.res):
+ for conv, res1,res2 in zip(self.convs, self.res1, self.res2):
encs.insert(0, x)
- x = tf.nn.leaky_relu( conv(x), 0.1 )
- x = res(x)
+ x = tf.nn.leaky_relu( conv(x), 0.2 )
+ x = res1(x)
+ x = res2(x)
+
+ center_out, x = self.center_out(x), tf.nn.leaky_relu( self.center_conv(x), 0.2 )
- center_out, x = self.center_out(x), self.center_conv(x)
-
- for i, (upconv, enc, upres) in enumerate(zip(self.upconvs, encs, self.upres)):
- x = tf.nn.leaky_relu( upconv(x), 0.1 )
+ for i, (upconv, enc, upres1, upres2 ) in enumerate(zip(self.upconvs, encs, self.upres1, self.upres2)):
+ x = tf.nn.leaky_relu( upconv(x), 0.2 )
x = tf.concat( [enc, x], axis=nn.conv2d_ch_axis)
- x = upres(x)
+ x = upres1(x)
+ x = upres2(x)
return center_out, self.out_conv(x)
diff --git a/core/leras/nn.py b/core/leras/nn.py
index c4420e1..ef5c2c9 100644
--- a/core/leras/nn.py
+++ b/core/leras/nn.py
@@ -76,15 +76,26 @@ class nn():
if first_run:
io.log_info("Caching GPU kernels...")
- #import tensorflow as tf
- import tensorflow.compat.v1 as tf
+ import tensorflow
+ tf_version = getattr(tensorflow,'VERSION', None)
+ if tf_version is None:
+ tf_version = tensorflow.version.GIT_VERSION
+ if tf_version[0] == 'v':
+ tf_version = tf_version[1:]
+
+ if tf_version[0] == '2':
+ tf = tensorflow.compat.v1
+ else:
+ tf = tensorflow
+
import logging
# Disable tensorflow warnings
tf_logger = logging.getLogger('tensorflow')
tf_logger.setLevel(logging.ERROR)
- tf.disable_v2_behavior()
+ if tf_version[0] == '2':
+ tf.disable_v2_behavior()
nn.tf = tf
# Initialize framework
diff --git a/doc/political_speech1.jpg b/doc/political_speech1.jpg
index 9775e75..33ae2ab 100644
Binary files a/doc/political_speech1.jpg and b/doc/political_speech1.jpg differ
diff --git a/merger/gfx/help_merger_masked.jpg b/merger/gfx/help_merger_masked.jpg
index df22ffb..f1822f3 100644
Binary files a/merger/gfx/help_merger_masked.jpg and b/merger/gfx/help_merger_masked.jpg differ
diff --git a/merger/gfx/help_merger_masked_source.psd b/merger/gfx/help_merger_masked_source.psd
index 437a410..b88c8ff 100644
Binary files a/merger/gfx/help_merger_masked_source.psd and b/merger/gfx/help_merger_masked_source.psd differ
diff --git a/models/Model_SAEHD/Model.py b/models/Model_SAEHD/Model.py
index cc71149..7f351c0 100644
--- a/models/Model_SAEHD/Model.py
+++ b/models/Model_SAEHD/Model.py
@@ -77,7 +77,7 @@ class SAEHDModel(ModelBase):
resolution = np.clip ( (resolution // 16) * 16, min_res, max_res)
self.options['resolution'] = resolution
self.options['face_type'] = io.input_str ("Face type", default_face_type, ['h','mf','f','wf','head', 'custom'], help_message="Half / mid face / full face / whole face / head / custom. Half face has better resolution, but covers less area of cheeks. Mid face is 30% wider than half face. 'Whole face' covers full area of face include forehead. 'head' covers full head, but requires XSeg for src and dst faceset.").lower()
-
+
while True:
archi = io.input_str ("AE architecture", default_archi, help_message=\
"""
@@ -138,7 +138,11 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
self.options['mouth_prio'] = io.input_bool ("Mouth priority", default_mouth_prio, help_message='Helps to fix mouth problems during training by forcing the neural network to train mouth with higher priority similar to eyes ')
self.options['uniform_yaw'] = io.input_bool ("Uniform yaw distribution of samples", default_uniform_yaw, help_message='Helps to fix blurry side faces due to small amount of them in the faceset.')
-
+
+ default_gan_power = self.options['gan_power'] = self.load_or_def_option('gan_power', 0.0)
+ default_gan_patch_size = self.options['gan_patch_size'] = self.load_or_def_option('gan_patch_size', self.options['resolution'] // 8)
+ default_gan_dims = self.options['gan_dims'] = self.load_or_def_option('gan_dims', 16)
+
if self.is_first_run() or ask_override:
self.options['models_opt_on_gpu'] = io.input_bool ("Place models and optimizer on GPU", default_models_opt_on_gpu, help_message="When you train on one GPU, by default model and optimizer weights are placed on GPU to accelerate the process. You can place they on CPU to free up extra VRAM, thus set bigger dimensions.")
@@ -148,11 +152,15 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
self.options['random_warp'] = io.input_bool ("Enable random warp of samples", default_random_warp, help_message="Random warp is required to generalize facial expressions of both faces. When the face is trained enough, you can disable it to get extra sharpness and reduce subpixel shake for less amount of iterations.")
- self.options['gan_power'] = np.clip ( io.input_number ("GAN power", default_gan_power, add_info="0.0 .. 10.0", help_message="Train the network in Generative Adversarial manner. Forces the neural network to learn small details of the face. Enable it only when the face is trained enough and don't disable. Typical fine value is 0.05"), 0.0, 10.0 )
-
- if (self.options['gan_power'] != 0):
- self.options['gan_old'] = io.input_bool ("Use old GAN version", default_gan_old, help_message="Use older version of GAN." )
-
+ self.options['gan_power'] = np.clip ( io.input_number ("GAN power", default_gan_power, add_info="0.0 .. 1.0", help_message="Forces the neural network to learn small details of the face. Enable it only when the face is trained enough with lr_dropout(on) and random_warp(off), and don't disable. The higher the value, the higher the chances of artifacts. Typical fine value is 0.1"), 0.0, 1.0 )
+
+ if self.options['gan_power'] != 0.0:
+ gan_patch_size = np.clip ( io.input_int("GAN patch size", default_gan_patch_size, add_info="3-640", help_message="The higher patch size, the higher the quality, the more VRAM is required. You can get sharper edges even at the lowest setting. Typical fine value is resolution / 8." ), 3, 640 )
+ self.options['gan_patch_size'] = gan_patch_size
+
+ gan_dims = np.clip ( io.input_int("GAN dimensions", default_gan_dims, add_info="4-64", help_message="The dimensions of the GAN network. The higher dimensions, the more VRAM is required. You can get sharper edges even at the lowest setting. Typical fine value is 16." ), 4, 64 )
+ self.options['gan_dims'] = gan_dims
+
if 'df' in self.options['archi']:
self.options['true_face_power'] = np.clip ( io.input_number ("'True face' power.", default_true_face_power, add_info="0.0000 .. 1.0", help_message="Experimental option. Discriminates result face to be more like src face. Higher value - stronger discrimination. Typical value is 0.01 . Comparison - https://i.imgur.com/czScS9q.png"), 0.0, 1.0 )
else:
@@ -168,6 +176,8 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
if self.options['pretrain'] and self.get_pretraining_data_path() is None:
raise Exception("pretraining_data_path is not defined")
+
+ self.gan_model_changed = (default_gan_patch_size != self.options['gan_patch_size']) or (default_gan_dims != self.options['gan_dims'])
self.pretrain_just_disabled = (default_pretrain == True and self.options['pretrain'] == False)
@@ -289,14 +299,8 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
if self.is_training:
if gan_power != 0:
- if gan_old:
- self.D_src = nn.PatchDiscriminator(patch_size=resolution//16, in_ch=input_ch, name="D_src")
- self.D_src_x2 = nn.PatchDiscriminator(patch_size=resolution//32, in_ch=input_ch, name="D_src_x2")
- self.model_filename_list += [ [self.D_src, 'D_src.npy'] ]
- self.model_filename_list += [ [self.D_src_x2, 'D_src_x2.npy'] ]
- else:
- self.D_src = nn.UNetPatchDiscriminator(patch_size=resolution//16, in_ch=input_ch, name="D_src")
- self.model_filename_list += [ [self.D_src, 'D_src_v2.npy'] ]
+ self.D_src = nn.UNetPatchDiscriminator(patch_size=self.options['gan_patch_size'], in_ch=input_ch, base_ch=self.options['gan_dims'], name="D_src")
+ self.model_filename_list += [ [self.D_src, 'GAN.npy'] ]
# Initialize optimizers
lr=5e-5
@@ -321,14 +325,9 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
self.model_filename_list += [ (self.D_code_opt, 'D_code_opt.npy') ]
if gan_power != 0:
- self.D_src_dst_opt = OptimizerClass(lr=lr, lr_dropout=lr_dropout, clipnorm=clipnorm, name='D_src_dst_opt')
-
- if gan_old:
- self.D_src_dst_opt.initialize_variables ( self.D_src.get_weights()+self.D_src_x2.get_weights(), vars_on_cpu=optimizer_vars_on_cpu, lr_dropout_on_cpu=self.options['lr_dropout']=='cpu')
- self.model_filename_list += [ (self.D_src_dst_opt, 'D_src_dst_opt.npy') ]
- else:
- self.D_src_dst_opt.initialize_variables ( self.D_src.get_weights(), vars_on_cpu=optimizer_vars_on_cpu, lr_dropout_on_cpu=self.options['lr_dropout']=='cpu')#+self.D_src_x2.get_weights()
- self.model_filename_list += [ (self.D_src_dst_opt, 'D_src_v2_opt.npy') ]
+ self.D_src_dst_opt = OptimizerClass(lr=lr, lr_dropout=lr_dropout, clipnorm=clipnorm, name='GAN_opt')
+ self.D_src_dst_opt.initialize_variables ( self.D_src.get_weights(), vars_on_cpu=optimizer_vars_on_cpu, lr_dropout_on_cpu=self.options['lr_dropout']=='cpu')#+self.D_src_x2.get_weights()
+ self.model_filename_list += [ (self.D_src_dst_opt, 'GAN_opt.npy') ]
if self.is_training:
# Adjust batch size for multiple GPU
@@ -412,14 +411,16 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
gpu_target_dstm_style_blur = gpu_target_dstm_blur #default style mask is 0.5 on boundary
gpu_target_dstm_blur = tf.clip_by_value(gpu_target_dstm_blur, 0, 0.5) * 2
- gpu_target_dst_masked = gpu_target_dst*gpu_target_dstm_blur
+ gpu_target_dst_masked = gpu_target_dst*gpu_target_dstm_blur
gpu_target_dst_style_masked = gpu_target_dst*gpu_target_dstm_style_blur
gpu_target_dst_style_anti_masked = gpu_target_dst*(1.0 - gpu_target_dstm_style_blur)
+ gpu_target_src_anti_masked = gpu_target_src*(1.0-gpu_target_srcm_blur)
gpu_target_src_masked_opt = gpu_target_src*gpu_target_srcm_blur if masked_training else gpu_target_src
gpu_target_dst_masked_opt = gpu_target_dst_masked if masked_training else gpu_target_dst
gpu_pred_src_src_masked_opt = gpu_pred_src_src*gpu_target_srcm_blur if masked_training else gpu_pred_src_src
+ gpu_pred_src_src_anti_masked = gpu_pred_src_src*(1.0-gpu_target_srcm_blur)
gpu_pred_dst_dst_masked_opt = gpu_pred_dst_dst*gpu_target_dstm_blur if masked_training else gpu_pred_dst_dst
gpu_psd_target_dst_style_masked = gpu_pred_src_dst*gpu_target_dstm_style_blur
@@ -533,19 +534,16 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
gpu_target_src_d, \
gpu_target_src_d2 = self.D_src(gpu_target_src_masked_opt)
- gpu_target_src_d_ones = tf.ones_like(gpu_target_src_d)
- gpu_target_src_d2_ones = tf.ones_like(gpu_target_src_d2)
-
- gpu_D_src_dst_loss = (DLoss(gpu_target_src_d_ones , gpu_target_src_d) + \
- DLoss(gpu_pred_src_src_d_zeros , gpu_pred_src_src_d) ) * 0.5 + \
- (DLoss(gpu_target_src_d2_ones , gpu_target_src_d2) + \
- DLoss(gpu_pred_src_src_d2_zeros , gpu_pred_src_src_d2) ) * 0.5
-
- gpu_D_src_dst_loss_gvs += [ nn.gradients (gpu_D_src_dst_loss, self.D_src.get_weights() ) ]#+self.D_src_x2.get_weights()
-
- gpu_G_loss += gan_power*(DLoss(gpu_pred_src_src_d_ones, gpu_pred_src_src_d) + \
- DLoss(gpu_pred_src_src_d2_ones, gpu_pred_src_src_d2))
-
+ gpu_G_loss += gan_power*(DLoss(gpu_pred_src_src_d_ones, gpu_pred_src_src_d) + \
+ DLoss(gpu_pred_src_src_d2_ones, gpu_pred_src_src_d2))
+
+
+
+ if masked_training:
+ # Minimal src-src-bg rec with total_variation_mse to suppress random bright dots from gan
+ gpu_G_loss += 0.000001*nn.total_variation_mse(gpu_pred_src_src)
+ gpu_G_loss += 0.02*tf.reduce_mean(tf.square(gpu_pred_src_src_anti_masked-gpu_target_src_anti_masked),axis=[1,2,3] )
+
gpu_G_loss_gvs += [ nn.gradients ( gpu_G_loss, self.src_dst_trainable_weights ) ]
@@ -646,6 +644,9 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
do_init = True
else:
do_init = self.is_first_run()
+ if self.is_training and gan_power != 0 and model == self.D_src:
+ if self.gan_model_changed:
+ do_init = True
if not do_init:
do_init = not model.load_weights( self.get_strpath_storage_for_file(filename) )
|
- https://www.youtube.com/watch?v=2Z1oA3GYPaY
+ https://www.youtube.com/watch?v=rYKkQ3BOo_E
@@ -192,7 +199,7 @@ Unfortunately, there is no "make everything ok" button in DeepFaceLab. You shoul