github/DeepFaceLab
1
0
Fork 0
mirror of https://github.com/iperov/DeepFaceLab.git synced 2025-07-07 05:22:06 -07:00
Code Issues Projects Releases Packages Wiki Activity Actions

SAE: model will not work with this update, you should to restart training,

Browse source 
applied paper https://arxiv.org/abs/1807.01442 https://github.com/aditya-grover/uae
This commit is contained in:
iperov 2019-04-20 22:33:28 +04:00
parent 7c5f762e78
commit 2cdf2745a2
1 changed files with 27 additions and 22 deletions
Download patch file Download diff file Expand all files Collapse all files

49
models/Model_SAE/Model.py
View file

@ -128,7 +128,7 @@ class SAEModel(ModelBase):
padding = 'reflect' if self.options['remove_gray_border'] else 'zero'
common_flow_kwargs = { 'padding': padding,
'norm': 'bn',
'act':'prelu' }
'act':'' }
weights_to_load = []
if self.options['archi'] == 'liae':
@ -370,8 +370,8 @@ class SAEModel(ModelBase):
dst_samples = generators_samples[1]
feed = [src_samples[0], dst_samples[0] ] + \
src_samples[1:1+self.ms_count*2] + \
dst_samples[1:1+self.ms_count*2]
src_samples[1:1+self.ms_count*2] + \
dst_samples[1:1+self.ms_count*2]
src_loss, dst_loss, = self.src_dst_train (feed)
@ -467,7 +467,7 @@ class SAEModel(ModelBase):
return LeakyReLU(alpha=lrelu_alpha)
class ResidualBlock(object):
def __init__(self, filters, kernel_size=3, padding='zero', use_reflection_padding=False, norm='', act='', **kwargs):
def __init__(self, filters, kernel_size=3, padding='zero', norm='', act='', **kwargs):
self.filters = filters
self.kernel_size = kernel_size
self.padding = padding
@ -486,9 +486,9 @@ class SAEModel(ModelBase):
return x
SAEModel.ResidualBlock = ResidualBlock
def downscale (dim, padding='zero', norm='', act='', **kwargs):
def downscale (dim, padding='zero', norm='', act='', kernel_regularizer=None, **kwargs):
def func(x):
return Norm(norm)( Act(act) (Conv2D(dim, kernel_size=5, strides=2, padding=padding)(x)) )
return Norm(norm)( Act(act) (Conv2D(dim, kernel_size=5, strides=2, padding=padding, kernel_regularizer=kernel_regularizer)(x)) )
return func
SAEModel.downscale = downscale
@ -508,8 +508,8 @@ class SAEModel(ModelBase):
def LIAEEncFlow(resolution, ch_dims, **kwargs):
exec (nnlib.import_all(), locals(), globals())
upscale = partial(SAEModel.upscale, **kwargs)
downscale = partial(SAEModel.downscale, **kwargs)
downscale = partial(SAEModel.downscale, kernel_regularizer=keras.regularizers.l2(0.0), **kwargs)
def func(input):
dims = K.int_shape(input)[-1]*ch_dims
@ -531,8 +531,12 @@ class SAEModel(ModelBase):
def func(input):
x = input[0]
x = Dense(ae_dims)(x)
x = Dense(lowest_dense_res * lowest_dense_res * ae_dims*2)(x)
#https://arxiv.org/abs/1807.01442 https://github.com/aditya-grover/uae
x = Dense(ae_dims, use_bias=False)(x)
x = Lambda ( lambda x: x + 0.1*K.random_normal(K.shape(x), 0, 1) , output_shape=(None,ae_dims) ) (x)
x = Dense(lowest_dense_res * lowest_dense_res * ae_dims*2, use_bias=False)(x)
x = Reshape((lowest_dense_res, lowest_dense_res, ae_dims*2))(x)
x = upscale(ae_dims*2)(x)
return x
@ -580,37 +584,38 @@ class SAEModel(ModelBase):
return func
@staticmethod
def DFEncFlow(resolution, ae_dims, ch_dims, padding='zero', **kwargs):
def DFEncFlow(resolution, ae_dims, ch_dims, **kwargs):
exec (nnlib.import_all(), locals(), globals())
upscale = partial(SAEModel.upscale, padding=padding)
downscale = partial(SAEModel.downscale, padding=padding)
upscale = partial(SAEModel.upscale, **kwargs)
downscale = partial(SAEModel.downscale, kernel_regularizer=keras.regularizers.l2(0.0), **kwargs)
lowest_dense_res = resolution // 16
def func(input):
x = input
dims = K.int_shape(input)[-1]*ch_dims
x = downscale(dims)(x)
x = downscale(dims*2)(x)
x = downscale(dims*4)(x)
x = downscale(dims*8)(x)
x = Dense(ae_dims)(Flatten()(x))
x = Dense(lowest_dense_res * lowest_dense_res * ae_dims)(x)
#https://arxiv.org/abs/1807.01442 https://github.com/aditya-grover/uae
x = Dense(ae_dims, use_bias=False)(Flatten()(x))
x = Lambda ( lambda x: x + 0.1*K.random_normal(K.shape(x), 0, 1) , output_shape=(None,ae_dims) ) (x)
x = Dense(lowest_dense_res * lowest_dense_res * ae_dims, use_bias=False)(x)
x = Reshape((lowest_dense_res, lowest_dense_res, ae_dims))(x)
x = upscale(ae_dims)(x)
return x
return func
@staticmethod
def DFDecFlow(output_nc, ch_dims, multiscale_count=1, add_residual_blocks=False, padding='zero', **kwargs):
def DFDecFlow(output_nc, ch_dims, multiscale_count=1, add_residual_blocks=False, **kwargs):
exec (nnlib.import_all(), locals(), globals())
upscale = partial(SAEModel.upscale, padding=padding)
to_bgr = partial(SAEModel.to_bgr, padding=padding)
upscale = partial(SAEModel.upscale, **kwargs)
to_bgr = partial(SAEModel.to_bgr, **kwargs)
dims = output_nc * ch_dims
ResidualBlock = partial(SAEModel.ResidualBlock, padding=padding)
ResidualBlock = partial(SAEModel.ResidualBlock, **kwargs)
def func(input):
x = input[0]