github/DeepFaceLab
1
0
Fork 0
mirror of https://github.com/iperov/DeepFaceLab.git synced 2025-07-06 04:52:13 -07:00
Code Issues Projects Releases Packages Wiki Activity Actions

SAEHD: random scale increased to -0.15+0.15. Improved lr_dropout capability to reach lower value of the loss.

Browse source 
When random_warp is off, inter_AB network is no longer trained to keep the face more src-like.

added option Random hue/saturation/light intensity applied to the src face set only at the input of the neural network. Stabilizes color perturbations during face swapping. Reduces the quality of the color transfer by selecting the closest one in the src faceset. Thus the src faceset must be diverse enough. Typical fine value is 0.05.

AMP: random scale increased to -0.15+0.15. Improved lr_dropout capability to reach lower value of the loss.
This commit is contained in:
iperov 2021-10-18 11:03:23 +04:00
parent f71a9e6bbd
commit 71f22957a6
2 changed files with 54 additions and 38 deletions
Download patch file Download diff file Expand all files Collapse all files

31
models/Model_AMP/Model.py
View file

@ -292,20 +292,21 @@ class AMPModel(ModelBase):
if self.is_training:
# Initialize optimizers
clipnorm = 1.0 if self.options['clipgrad'] else 0.0
lr_dropout = 0.3 if self.options['lr_dropout'] in ['y','cpu'] else 1.0
if self.options['lr_dropout'] in ['y','cpu'] and not self.pretrain:
lr_cos = 500
lr_dropout = 0.3
else:
lr_cos = 0
lr_dropout = 1.0
self.G_weights = self.encoder.get_weights() + self.decoder.get_weights()
#if random_warp:
# self.G_weights += self.inter_src.get_weights() + self.inter_dst.get_weights()
self.src_dst_opt = nn.AdaBelief(lr=5e-5, lr_dropout=lr_dropout, clipnorm=clipnorm, name='src_dst_opt')
self.src_dst_opt = nn.AdaBelief(lr=5e-5, lr_dropout=lr_dropout, lr_cos=lr_cos, clipnorm=clipnorm, name='src_dst_opt')
self.src_dst_opt.initialize_variables (self.G_weights, vars_on_cpu=optimizer_vars_on_cpu)
self.model_filename_list += [ (self.src_dst_opt, 'src_dst_opt.npy') ]
if gan_power != 0:
self.GAN = nn.UNetPatchDiscriminator(patch_size=self.options['gan_patch_size'], in_ch=input_ch, base_ch=self.options['gan_dims'], name="GAN")
self.GAN_opt = nn.AdaBelief(lr=5e-5, lr_dropout=lr_dropout, clipnorm=clipnorm, name='GAN_opt')
self.GAN_opt = nn.AdaBelief(lr=5e-5, lr_dropout=lr_dropout, lr_cos=lr_cos, clipnorm=clipnorm, name='GAN_opt')
self.GAN_opt.initialize_variables ( self.GAN.get_weights(), vars_on_cpu=optimizer_vars_on_cpu)
self.model_filename_list += [ [self.GAN, 'GAN.npy'],
[self.GAN_opt, 'GAN_opt.npy'] ]
@ -391,15 +392,15 @@ class AMPModel(ModelBase):
if blur_out_mask:
sigma = resolution / 128
x = nn.gaussian_blur(gpu_target_src*gpu_target_srcm_anti, sigma)
y = 1-nn.gaussian_blur(gpu_target_srcm, sigma)
y = tf.where(tf.equal(y, 0), tf.ones_like(y), y)
y = 1-nn.gaussian_blur(gpu_target_srcm, sigma)
y = tf.where(tf.equal(y, 0), tf.ones_like(y), y)
gpu_target_src = gpu_target_src*gpu_target_srcm + (x/y)*gpu_target_srcm_anti
x = nn.gaussian_blur(gpu_target_dst*gpu_target_dstm_anti, sigma)
y = 1-nn.gaussian_blur(gpu_target_dstm, sigma)
y = tf.where(tf.equal(y, 0), tf.ones_like(y), y)
y = 1-nn.gaussian_blur(gpu_target_dstm, sigma)
y = tf.where(tf.equal(y, 0), tf.ones_like(y), y)
gpu_target_dst = gpu_target_dst*gpu_target_dstm + (x/y)*gpu_target_dstm_anti
gpu_target_src_masked = gpu_target_src*gpu_target_srcm_blur
@ -561,7 +562,7 @@ class AMPModel(ModelBase):
self.set_training_data_generators ([
SampleGeneratorFace(training_data_src_path, random_ct_samples_path=random_ct_samples_path, debug=self.is_debug(), batch_size=self.get_batch_size(),
sample_process_options=SampleProcessor.Options(scale_range=[-0.125, 0.125], random_flip=self.random_src_flip),
sample_process_options=SampleProcessor.Options(scale_range=[-0.15, 0.15], random_flip=self.random_src_flip),
output_sample_types = [ {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':random_warp, 'transform':True, 'channel_type' : SampleProcessor.ChannelType.BGR, 'ct_mode': ct_mode, 'face_type':face_type, 'data_format':nn.data_format, 'resolution': resolution},
{'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':False , 'transform':True, 'channel_type' : SampleProcessor.ChannelType.BGR, 'ct_mode': ct_mode, 'face_type':face_type, 'data_format':nn.data_format, 'resolution': resolution},
{'sample_type': SampleProcessor.SampleType.FACE_MASK, 'warp':False , 'transform':True, 'channel_type' : SampleProcessor.ChannelType.G, 'face_mask_type' : SampleProcessor.FaceMaskType.FULL_FACE, 'face_type':face_type, 'data_format':nn.data_format, 'resolution': resolution},
@ -571,7 +572,7 @@ class AMPModel(ModelBase):
generators_count=src_generators_count ),
SampleGeneratorFace(training_data_dst_path, debug=self.is_debug(), batch_size=self.get_batch_size(),
sample_process_options=SampleProcessor.Options(scale_range=[-0.125, 0.125], random_flip=self.random_dst_flip),
sample_process_options=SampleProcessor.Options(scale_range=[-0.15, 0.15], random_flip=self.random_dst_flip),
output_sample_types = [ {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':random_warp, 'transform':True, 'channel_type' : SampleProcessor.ChannelType.BGR, 'face_type':face_type, 'data_format':nn.data_format, 'resolution': resolution},
{'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':False , 'transform':True, 'channel_type' : SampleProcessor.ChannelType.BGR, 'face_type':face_type, 'data_format':nn.data_format, 'resolution': resolution},
{'sample_type': SampleProcessor.SampleType.FACE_MASK, 'warp':False , 'transform':True, 'channel_type' : SampleProcessor.ChannelType.G, 'face_mask_type' : SampleProcessor.FaceMaskType.FULL_FACE, 'face_type':face_type, 'data_format':nn.data_format, 'resolution': resolution},