github/DeepFaceLab
1
0
Fork 0
mirror of https://github.com/iperov/DeepFaceLab.git synced 2025-08-22 22:34:25 -07:00
Code Issues Projects Releases Packages Wiki Activity Actions

more debugging

Browse source
This commit is contained in:
Jeremy Hummel 2019-10-20 00:28:52 -04:00
commit 39446f6188
2 changed files with 14 additions and 13 deletions
Download patch file Download diff file Expand all files Collapse all files

10
models/Model_SAEHD/Model.py
View file

@ -478,7 +478,7 @@ class SAEHDModel(ModelBase):
if self.options['ms_ssim_loss']: if self.options['ms_ssim_loss']:
# TODO - Done # TODO - Done
src_loss = K.mean(10 * MsSSIM(max_value=1.0, power_factors=(1.0,))(target_src_masked_opt, pred_src_src_masked_opt)) src_loss = K.mean(10 * MsSSIM(max_value=1.0)(target_src_masked_opt, pred_src_src_masked_opt))
else: else:
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*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 ) ) src_loss += K.mean ( 10*K.square( target_src_masked_opt - pred_src_src_masked_opt ) )
@ -495,14 +495,14 @@ class SAEHDModel(ModelBase):
if bg_style_power != 0: if bg_style_power != 0:
if self.options['ms_ssim_loss']: if self.options['ms_ssim_loss']:
# TODO - Done # TODO - Done
src_loss += K.mean(10 * bg_style_power * MsSSIM(max_value=1.0, power_factors=(1.0,))(psd_target_dst_anti_masked, target_dst_anti_masked)) src_loss += K.mean(10 * bg_style_power * MsSSIM(max_value=1.0)(psd_target_dst_anti_masked, target_dst_anti_masked))
else: else:
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)*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 )) src_loss += K.mean( (10*bg_style_power)*K.square( psd_target_dst_anti_masked - target_dst_anti_masked ))
if self.options['ms_ssim_loss']: if self.options['ms_ssim_loss']:
# TODO - Done # TODO - Done
dst_loss = K.mean(10 * MsSSIM(max_value=1.0, power_factors=(1.0,))(target_dst_masked_opt, pred_dst_dst_masked_opt)) dst_loss = K.mean(10 * MsSSIM(max_value=1.0)(target_dst_masked_opt, pred_dst_dst_masked_opt))
else: else:
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*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 ) ) dst_loss += K.mean( 10*K.square( target_dst_masked_opt - pred_dst_dst_masked_opt ) )
@ -533,8 +533,8 @@ class SAEHDModel(ModelBase):
if self.options['learn_mask']: if self.options['learn_mask']:
if self.options['ms_ssim_loss']: if self.options['ms_ssim_loss']:
# TODO - Done # TODO - Done
src_mask_loss = K.mean(MsSSIM(max_value=1.0, power_factors=(1.0,))(self.model.target_srcm, self.model.pred_src_srcm)) src_mask_loss = K.mean(MsSSIM(max_value=1.0)(self.model.target_srcm, self.model.pred_src_srcm))
dst_mask_loss = K.mean(MsSSIM(max_value=1.0, power_factors=(1.0,))(self.model.target_dstm, self.model.pred_dst_dstm)) dst_mask_loss = K.mean(MsSSIM(max_value=1.0)(self.model.target_dstm, self.model.pred_dst_dstm))
else: else:
src_mask_loss = K.mean(K.square(self.model.target_srcm-self.model.pred_src_srcm)) 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)) dst_mask_loss = K.mean(K.square(self.model.target_dstm-self.model.pred_dst_dstm))

17
nnlib/nnlib.py
View file

@ -365,14 +365,15 @@ NLayerDiscriminator = nnlib.NLayerDiscriminator
k1=self.k1, k2=self.k2) k1=self.k1, k2=self.k2)
loss = (1.0 - mssim_val) / 2.0 loss = (1.0 - mssim_val) / 2.0
return loss return loss
loss = 0.0 else:
# im_size = K.shape(y_pred)[-2] loss = 0.0
for i, weight in enumerate(self.power_factors): # im_size = K.shape(y_pred)[-2]
size = 2**i for i, weight in enumerate(self.power_factors):
dssim = self.dssim(K.pool2d(y_true, (size, size), strides=(size, size), pool_mode='avg'), size = 2**i
K.pool2d(y_pred, (size, size), strides=(size, size), pool_mode='avg')) dssim = self.dssim(K.pool2d(y_true, (size, size), strides=(size, size), pool_mode='avg'),
loss += dssim**weight K.pool2d(y_pred, (size, size), strides=(size, size), pool_mode='avg'))
return loss/len(self.power_factors) loss += dssim**weight
return loss/len(self.power_factors)
nnlib.MsSSIM = MsSSIM nnlib.MsSSIM = MsSSIM