github/DeepFaceLab
1
0
Fork 0
mirror of https://github.com/iperov/DeepFaceLab.git synced 2025-08-22 06:23:20 -07:00
Code Issues Projects Releases Packages Wiki Activity Actions

Update DeepFakeArchi.py

Browse source
This commit is contained in:
TalosOfCrete 2020-05-21 18:52:09 -05:00
commit c29982222b
1 changed files with 41 additions and 14 deletions
Download patch file Download diff file Expand all files Collapse all files

55
core/leras/archis/DeepFakeArchi.py
View file

@ -335,7 +335,7 @@ class DeepFakeArchi(nn.ArchiBase):
elif mod == 'uhd':
class Downscale(nn.ModelBase):
def __init__(self, in_ch, out_ch, kernel_size=5, depth_multiplier=1, dilations=1, subpixel=True, use_activator=True, *kwargs ):
def __init__(self, in_ch, out_ch, kernel_size=3, depth_multiplier=1, dilations=1, subpixel=True, use_activator=True, *kwargs ):
self.in_ch = in_ch
self.out_ch = out_ch
self.kernel_size = kernel_size
@ -367,9 +367,9 @@ class DeepFakeArchi(nn.ArchiBase):
def get_out_ch(self):
return (self.out_ch // 4) * 4 if self.subpixel else self.out_ch
class DownscaleBlock(nn.ModelBase):
def on_build(self, in_ch, ch, n_downscales, kernel_size, dilations=1, subpixel=True, use_activator=True):
def on_build(self, in_ch, ch, n_downscales, kernel_size, dilations=1, subpixel=True, use_activator=False):
self.downs = []
last_ch = in_ch
@ -377,9 +377,7 @@ class DeepFakeArchi(nn.ArchiBase):
cur_ch = ch*( min(2**i, 8) )
self.downs.append ( Downscale(last_ch, cur_ch, kernel_size=kernel_size, dilations=dilations, subpixel=subpixel, use_activator=use_activator) )
last_ch = self.downs[-1].get_out_ch()
self.bp1 = nn.BlurPool(kernel_size)
def forward(self, inp):
x = inp
for down in self.downs:
@ -387,6 +385,38 @@ class DeepFakeArchi(nn.ArchiBase):
x = self.bp1(x)
return x
class DecayingDownscaleBlock(nn.ModelBase):
def on_build(self, in_ch, ch, n_downscales=4, init_kernel_size=5, kernel_floor=3, dilations=1, subpixel=True, use_activator=True):
self.downs = []
if init_kernel_size % 2 == 0:
init_kernel_size += 1
print("Initial kernel size has been adjusted up by 1 as it was an even number.")
if kernel_floor % 2 == 0:
kernel_floor += 1
print("Kernel floor has been adjusted up by 1 as it was an even number.")
if not init_kernel_size > kernel_floor:
raise ValueError("The initial kernel size must be larger than the kernel floor.")
cur_kernel_size = init_kernel_size
last_ch = in_ch
for i in range(n_downscales):
cur_ch = ch*( min(2**i, 8) )
self.downs.append ( Downscale(last_ch, cur_ch, kernel_size=cur_kernel_size, dilations=dilations, subpixel=subpixel, use_activator=use_activator) )
last_ch = self.downs[-1].get_out_ch()
if cur_kernel_size != kernel_floor and cur_kernel_size-2 != 1:
cur_kernel_size -= 2
def forward(self, inp):
x = inp
for down in self.downs:
x = down(x)
return x
class Upscale(nn.ModelBase):
def on_build(self, in_ch, out_ch, kernel_size=3, depth_multiplier=1 ):
self.conv1 = nn.SeparableConv2D( in_ch, out_ch*4, kernel_size=kernel_size, depth_multiplier=depth_multiplier, padding='SAME')
@ -441,17 +471,14 @@ class DeepFakeArchi(nn.ArchiBase):
"""
class Encoder(nn.ModelBase):
def on_build(self, in_ch, e_ch, **kwargs):
self.down1 = DownscaleBlock(in_ch, e_ch, n_downscales=4, kernel_size=5, dilations=1, use_activator=False, subpixel=False)
#self.down2 = DownscaleBlock(in_ch, e_ch//2, n_downscales=4, kernel_size=5, dilations=1, use_activator=False)
#self.down3 = DownscaleBlock(in_ch, e_ch//2, n_downscales=4, kernel_size=5, dilations=2, use_activator=False)
#self.down4 = DownscaleBlock(in_ch, e_ch//2, n_downscales=4, kernel_size=7, dilations=2, use_activator=False)
self.down1 = DecayingDownscaleBlock(in_ch, e_ch*2, n_downscales=6, dilations=1, use_activator=False)
self.down2 = DecayingDownscaleBlock(in_ch, e_ch//2, n_downscales=6, dilations=2, use_activator=False)
def forward(self, inp):
x = nn.flatten(self.down1(inp))
#x = tf.concat([ nn.flatten(self.down1(inp)),
#nn.flatten(self.down2(inp)) ], -1),
#nn.flatten(self.down3(inp)),
#nn.flatten(self.down4(inp)) ], -1 )
#x = nn.flatten(self.down1(inp))
x = tf.concat([ nn.flatten(self.down1(inp)),
nn.flatten(self.down2(inp)) ], -1 )
return x
lowest_dense_res = resolution // 16