mirror of
https://github.com/iperov/DeepFaceLab.git
synced 2025-08-19 04:59:27 -07:00
Merge remote-tracking branch 'original/master' into whitespace-fix
This commit is contained in:
camjac251
commit
56f7add24c
21 changed files with 492 additions and 549 deletions
2
.vscode/launch.json
vendored
2
.vscode/launch.json
vendored
|
|
@ -12,7 +12,7 @@
|
|||
"type": "python",
|
||||
"request": "launch",
|
||||
"program": "${env:DFL_ROOT}\\main.py",
|
||||
"pythonPath": "${env:PYTHONEXECUTABLE}",
|
||||
"python": "${env:PYTHONEXECUTABLE}",
|
||||
"cwd": "${env:WORKSPACE}",
|
||||
"console": "integratedTerminal",
|
||||
"args": ["train",
|
||||
|
|
|
|||
17
README.md
17
README.md
|
|
@ -29,8 +29,8 @@ More than 95% of deepfake videos are created with DeepFaceLab.
|
|||
|
||||
DeepFaceLab is used by such popular youtube channels as
|
||||
|
||||
| [deeptomcruise](https://www.tiktok.com/@deeptomcruise)| [1facerussia](https://www.tiktok.com/@1facerussia)| [arnoldschwarzneggar](https://www.tiktok.com/@arnoldschwarzneggar)
|
||||
|---|---|---|
|
||||
| [deeptomcruise](https://www.tiktok.com/@deeptomcruise)| [1facerussia](https://www.tiktok.com/@1facerussia)| [arnoldschwarzneggar](https://www.tiktok.com/@arnoldschwarzneggar)| [mariahcareyathome?](https://www.tiktok.com/@mariahcareyathome?)
|
||||
|---|---|---|---|
|
||||
|
||||
| [Ctrl Shift Face](https://www.youtube.com/channel/UCKpH0CKltc73e4wh0_pgL3g)| [VFXChris Ume](https://www.youtube.com/channel/UCGf4OlX_aTt8DlrgiH3jN3g/videos)| [Sham00k](https://www.youtube.com/channel/UCZXbWcv7fSZFTAZV4beckyw/videos)|
|
||||
|---|---|---|
|
||||
|
|
@ -194,7 +194,7 @@ Unfortunately, there is no "make everything ok" button in DeepFaceLab. You shoul
|
|||
</td></tr>
|
||||
|
||||
<tr><td align="right">
|
||||
<a href="https://tinyurl.com/4tb2tn4w">Windows (magnet link)</a>
|
||||
<a href="https://tinyurl.com/2afv92ay">Windows (magnet link)</a>
|
||||
</td><td align="center">Last release. Use torrent client to download.</td></tr>
|
||||
|
||||
<tr><td align="right">
|
||||
|
|
@ -305,6 +305,17 @@ QQ群1095077489
|
|||
<a href="https://www.deepfaker.xyz/">deepfaker.xyz</a>
|
||||
</td><td align="center">中文学习站（非官方)</td></tr>
|
||||
|
||||
<tr><td colspan=2 align="center">
|
||||
|
||||
## Related works
|
||||
|
||||
</td></tr>
|
||||
|
||||
<tr><td align="right">
|
||||
<a href="https://github.com/neuralchen/SimSwap">neuralchen/SimSwap</a>
|
||||
</td><td align="center">Swapping face using ONE single photo 一张图免训练换脸</td></tr>
|
||||
|
||||
</td></tr>
|
||||
</table>
|
||||
|
||||
<table align="center" border="0">
|
||||
|
|
|
|||
|
|
@ -8,12 +8,15 @@ class DeepFakeArchi(nn.ArchiBase):
|
|||
mod None - default
|
||||
'quick'
|
||||
"""
|
||||
def __init__(self, resolution, mod=None, opts=None):
|
||||
def __init__(self, resolution, use_fp16=False, mod=None, opts=None):
|
||||
super().__init__()
|
||||
|
||||
if opts is None:
|
||||
opts = ''
|
||||
|
||||
|
||||
conv_dtype = tf.float16 if use_fp16 else tf.float32
|
||||
|
||||
if mod is None:
|
||||
class Downscale(nn.ModelBase):
|
||||
def __init__(self, in_ch, out_ch, kernel_size=5, *kwargs ):
|
||||
|
|
@ -23,7 +26,7 @@ class DeepFakeArchi(nn.ArchiBase):
|
|||
super().__init__(*kwargs)
|
||||
|
||||
def on_build(self, *args, **kwargs ):
|
||||
self.conv1 = nn.Conv2D( self.in_ch, self.out_ch, kernel_size=self.kernel_size, strides=2, padding='SAME')
|
||||
self.conv1 = nn.Conv2D( self.in_ch, self.out_ch, kernel_size=self.kernel_size, strides=2, padding='SAME', dtype=conv_dtype)
|
||||
|
||||
def forward(self, x):
|
||||
x = self.conv1(x)
|
||||
|
|
@ -40,7 +43,7 @@ class DeepFakeArchi(nn.ArchiBase):
|
|||
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=kernel_size) )
|
||||
self.downs.append ( Downscale(last_ch, cur_ch, kernel_size=kernel_size))
|
||||
last_ch = self.downs[-1].get_out_ch()
|
||||
|
||||
def forward(self, inp):
|
||||
|
|
@ -50,8 +53,8 @@ class DeepFakeArchi(nn.ArchiBase):
|
|||
return x
|
||||
|
||||
class Upscale(nn.ModelBase):
|
||||
def on_build(self, in_ch, out_ch, kernel_size=3 ):
|
||||
self.conv1 = nn.Conv2D( in_ch, out_ch*4, kernel_size=kernel_size, padding='SAME')
|
||||
def on_build(self, in_ch, out_ch, kernel_size=3):
|
||||
self.conv1 = nn.Conv2D( in_ch, out_ch*4, kernel_size=kernel_size, padding='SAME', dtype=conv_dtype)
|
||||
|
||||
def forward(self, x):
|
||||
x = self.conv1(x)
|
||||
|
|
@ -60,9 +63,9 @@ class DeepFakeArchi(nn.ArchiBase):
|
|||
return x
|
||||
|
||||
class ResidualBlock(nn.ModelBase):
|
||||
def on_build(self, ch, kernel_size=3 ):
|
||||
self.conv1 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME')
|
||||
self.conv2 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME')
|
||||
def on_build(self, ch, kernel_size=3):
|
||||
self.conv1 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME', dtype=conv_dtype)
|
||||
self.conv2 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME', dtype=conv_dtype)
|
||||
|
||||
def forward(self, inp):
|
||||
x = self.conv1(inp)
|
||||
|
|
@ -76,16 +79,21 @@ class DeepFakeArchi(nn.ArchiBase):
|
|||
self.in_ch = in_ch
|
||||
self.e_ch = e_ch
|
||||
super().__init__(**kwargs)
|
||||
|
||||
def on_build(self):
|
||||
|
||||
def on_build(self):
|
||||
self.down1 = DownscaleBlock(self.in_ch, self.e_ch, n_downscales=4, kernel_size=5)
|
||||
|
||||
def forward(self, inp):
|
||||
return nn.flatten(self.down1(inp))
|
||||
|
||||
def forward(self, x):
|
||||
if use_fp16:
|
||||
x = tf.cast(x, tf.float16)
|
||||
x = nn.flatten(self.down1(x))
|
||||
if use_fp16:
|
||||
x = tf.cast(x, tf.float32)
|
||||
return x
|
||||
|
||||
def get_out_res(self, res):
|
||||
return res // (2**4)
|
||||
|
||||
|
||||
def get_out_ch(self):
|
||||
return self.e_ch * 8
|
||||
|
||||
|
|
@ -98,9 +106,10 @@ class DeepFakeArchi(nn.ArchiBase):
|
|||
|
||||
def on_build(self):
|
||||
in_ch, ae_ch, ae_out_ch = self.in_ch, self.ae_ch, self.ae_out_ch
|
||||
|
||||
if 'u' in opts:
|
||||
self.dense_norm = nn.DenseNorm()
|
||||
|
||||
|
||||
self.dense1 = nn.Dense( in_ch, ae_ch )
|
||||
self.dense2 = nn.Dense( ae_ch, lowest_dense_res * lowest_dense_res * ae_out_ch )
|
||||
self.upscale1 = Upscale(ae_out_ch, ae_out_ch)
|
||||
|
|
@ -112,6 +121,9 @@ class DeepFakeArchi(nn.ArchiBase):
|
|||
x = self.dense1(x)
|
||||
x = self.dense2(x)
|
||||
x = nn.reshape_4D (x, lowest_dense_res, lowest_dense_res, self.ae_out_ch)
|
||||
|
||||
if use_fp16:
|
||||
x = tf.cast(x, tf.float16)
|
||||
x = self.upscale1(x)
|
||||
return x
|
||||
|
||||
|
|
@ -122,7 +134,7 @@ class DeepFakeArchi(nn.ArchiBase):
|
|||
return self.ae_out_ch
|
||||
|
||||
class Decoder(nn.ModelBase):
|
||||
def on_build(self, in_ch, d_ch, d_mask_ch ):
|
||||
def on_build(self, in_ch, d_ch, d_mask_ch):
|
||||
self.upscale0 = Upscale(in_ch, d_ch*8, kernel_size=3)
|
||||
self.upscale1 = Upscale(d_ch*8, d_ch*4, kernel_size=3)
|
||||
self.upscale2 = Upscale(d_ch*4, d_ch*2, kernel_size=3)
|
||||
|
|
@ -131,25 +143,23 @@ class DeepFakeArchi(nn.ArchiBase):
|
|||
self.res1 = ResidualBlock(d_ch*4, kernel_size=3)
|
||||
self.res2 = ResidualBlock(d_ch*2, kernel_size=3)
|
||||
|
||||
self.out_conv = nn.Conv2D( d_ch*2, 3, kernel_size=1, padding='SAME')
|
||||
self.out_conv = nn.Conv2D( d_ch*2, 3, kernel_size=1, padding='SAME', dtype=conv_dtype)
|
||||
|
||||
self.upscalem0 = Upscale(in_ch, d_mask_ch*8, kernel_size=3)
|
||||
self.upscalem1 = Upscale(d_mask_ch*8, d_mask_ch*4, kernel_size=3)
|
||||
self.upscalem2 = Upscale(d_mask_ch*4, d_mask_ch*2, kernel_size=3)
|
||||
self.out_convm = nn.Conv2D( d_mask_ch*2, 1, kernel_size=1, padding='SAME')
|
||||
self.out_convm = nn.Conv2D( d_mask_ch*2, 1, kernel_size=1, padding='SAME', dtype=conv_dtype)
|
||||
|
||||
if 'd' in opts:
|
||||
self.out_conv1 = nn.Conv2D( d_ch*2, 3, kernel_size=3, padding='SAME')
|
||||
self.out_conv2 = nn.Conv2D( d_ch*2, 3, kernel_size=3, padding='SAME')
|
||||
self.out_conv3 = nn.Conv2D( d_ch*2, 3, kernel_size=3, padding='SAME')
|
||||
self.out_conv1 = nn.Conv2D( d_ch*2, 3, kernel_size=3, padding='SAME', dtype=conv_dtype)
|
||||
self.out_conv2 = nn.Conv2D( d_ch*2, 3, kernel_size=3, padding='SAME', dtype=conv_dtype)
|
||||
self.out_conv3 = nn.Conv2D( d_ch*2, 3, kernel_size=3, padding='SAME', dtype=conv_dtype)
|
||||
self.upscalem3 = Upscale(d_mask_ch*2, d_mask_ch*1, kernel_size=3)
|
||||
self.out_convm = nn.Conv2D( d_mask_ch*1, 1, kernel_size=1, padding='SAME')
|
||||
self.out_convm = nn.Conv2D( d_mask_ch*1, 1, kernel_size=1, padding='SAME', dtype=conv_dtype)
|
||||
else:
|
||||
self.out_convm = nn.Conv2D( d_mask_ch*2, 1, kernel_size=1, padding='SAME')
|
||||
|
||||
def forward(self, inp):
|
||||
z = inp
|
||||
self.out_convm = nn.Conv2D( d_mask_ch*2, 1, kernel_size=1, padding='SAME', dtype=conv_dtype)
|
||||
|
||||
def forward(self, z):
|
||||
x = self.upscale0(z)
|
||||
x = self.res0(x)
|
||||
x = self.upscale1(x)
|
||||
|
|
@ -157,40 +167,11 @@ class DeepFakeArchi(nn.ArchiBase):
|
|||
x = self.upscale2(x)
|
||||
x = self.res2(x)
|
||||
|
||||
|
||||
if 'd' in opts:
|
||||
x0 = tf.nn.sigmoid(self.out_conv(x))
|
||||
x0 = nn.upsample2d(x0)
|
||||
x1 = tf.nn.sigmoid(self.out_conv1(x))
|
||||
x1 = nn.upsample2d(x1)
|
||||
x2 = tf.nn.sigmoid(self.out_conv2(x))
|
||||
x2 = nn.upsample2d(x2)
|
||||
x3 = tf.nn.sigmoid(self.out_conv3(x))
|
||||
x3 = nn.upsample2d(x3)
|
||||
|
||||
if nn.data_format == "NHWC":
|
||||
tile_cfg = ( 1, resolution // 2, resolution //2, 1)
|
||||
else:
|
||||
tile_cfg = ( 1, 1, resolution // 2, resolution //2 )
|
||||
|
||||
z0 = tf.concat ( ( tf.concat ( ( tf.ones ( (1,1,1,1) ), tf.zeros ( (1,1,1,1) ) ), axis=nn.conv2d_spatial_axes[1] ),
|
||||
tf.concat ( ( tf.zeros ( (1,1,1,1) ), tf.zeros ( (1,1,1,1) ) ), axis=nn.conv2d_spatial_axes[1] ) ), axis=nn.conv2d_spatial_axes[0] )
|
||||
|
||||
z0 = tf.tile ( z0, tile_cfg )
|
||||
|
||||
z1 = tf.concat ( ( tf.concat ( ( tf.zeros ( (1,1,1,1) ), tf.ones ( (1,1,1,1) ) ), axis=nn.conv2d_spatial_axes[1] ),
|
||||
tf.concat ( ( tf.zeros ( (1,1,1,1) ), tf.zeros ( (1,1,1,1) ) ), axis=nn.conv2d_spatial_axes[1] ) ), axis=nn.conv2d_spatial_axes[0] )
|
||||
z1 = tf.tile ( z1, tile_cfg )
|
||||
|
||||
z2 = tf.concat ( ( tf.concat ( ( tf.zeros ( (1,1,1,1) ), tf.zeros ( (1,1,1,1) ) ), axis=nn.conv2d_spatial_axes[1] ),
|
||||
tf.concat ( ( tf.ones ( (1,1,1,1) ), tf.zeros ( (1,1,1,1) ) ), axis=nn.conv2d_spatial_axes[1] ) ), axis=nn.conv2d_spatial_axes[0] )
|
||||
z2 = tf.tile ( z2, tile_cfg )
|
||||
|
||||
z3 = tf.concat ( ( tf.concat ( ( tf.zeros ( (1,1,1,1) ), tf.zeros ( (1,1,1,1) ) ), axis=nn.conv2d_spatial_axes[1] ),
|
||||
tf.concat ( ( tf.zeros ( (1,1,1,1) ), tf.ones ( (1,1,1,1) ) ), axis=nn.conv2d_spatial_axes[1] ) ), axis=nn.conv2d_spatial_axes[0] )
|
||||
z3 = tf.tile ( z3, tile_cfg )
|
||||
|
||||
x = x0*z0 + x1*z1 + x2*z2 + x3*z3
|
||||
x = tf.nn.sigmoid( nn.depth_to_space(tf.concat( (self.out_conv(x),
|
||||
self.out_conv1(x),
|
||||
self.out_conv2(x),
|
||||
self.out_conv3(x)), nn.conv2d_ch_axis), 2) )
|
||||
else:
|
||||
x = tf.nn.sigmoid(self.out_conv(x))
|
||||
|
||||
|
|
@ -201,9 +182,13 @@ class DeepFakeArchi(nn.ArchiBase):
|
|||
if 'd' in opts:
|
||||
m = self.upscalem3(m)
|
||||
m = tf.nn.sigmoid(self.out_convm(m))
|
||||
|
||||
|
||||
if use_fp16:
|
||||
x = tf.cast(x, tf.float32)
|
||||
m = tf.cast(m, tf.float32)
|
||||
|
||||
return x, m
|
||||
|
||||
|
||||
self.Encoder = Encoder
|
||||
self.Inter = Inter
|
||||
self.Decoder = Decoder
|
||||
|
|
|
|||
|
|
@ -55,8 +55,8 @@ class Conv2D(nn.LayerBase):
|
|||
if kernel_initializer is None:
|
||||
kernel_initializer = tf.initializers.random_normal(0, 1.0, dtype=self.dtype)
|
||||
|
||||
if kernel_initializer is None:
|
||||
kernel_initializer = nn.initializers.ca()
|
||||
#if kernel_initializer is None:
|
||||
# kernel_initializer = nn.initializers.ca()
|
||||
|
||||
self.weight = tf.get_variable("weight", (self.kernel_size,self.kernel_size,self.in_ch,self.out_ch), dtype=self.dtype, initializer=kernel_initializer, trainable=self.trainable )
|
||||
|
||||
|
|
|
|||
|
|
@ -38,8 +38,8 @@ class Conv2DTranspose(nn.LayerBase):
|
|||
if kernel_initializer is None:
|
||||
kernel_initializer = tf.initializers.random_normal(0, 1.0, dtype=self.dtype)
|
||||
|
||||
if kernel_initializer is None:
|
||||
kernel_initializer = nn.initializers.ca()
|
||||
#if kernel_initializer is None:
|
||||
# kernel_initializer = nn.initializers.ca()
|
||||
self.weight = tf.get_variable("weight", (self.kernel_size,self.kernel_size,self.out_ch,self.in_ch), dtype=self.dtype, initializer=kernel_initializer, trainable=self.trainable )
|
||||
|
||||
if self.use_bias:
|
||||
|
|
|
|||
|
|
@ -68,8 +68,8 @@ class DepthwiseConv2D(nn.LayerBase):
|
|||
if kernel_initializer is None:
|
||||
kernel_initializer = tf.initializers.random_normal(0, 1.0, dtype=self.dtype)
|
||||
|
||||
if kernel_initializer is None:
|
||||
kernel_initializer = nn.initializers.ca()
|
||||
#if kernel_initializer is None:
|
||||
# kernel_initializer = nn.initializers.ca()
|
||||
|
||||
self.weight = tf.get_variable("weight", (self.kernel_size,self.kernel_size,self.in_ch,self.depth_multiplier), dtype=self.dtype, initializer=kernel_initializer, trainable=self.trainable )
|
||||
|
||||
|
|
|
|||
|
|
@ -111,7 +111,7 @@ class UNetPatchDiscriminator(nn.ModelBase):
|
|||
for i in range(layers_count-1):
|
||||
st = 1 + (1 if val & (1 << i) !=0 else 0 )
|
||||
layers.append ( [3, st ])
|
||||
sum_st += st
|
||||
sum_st += st
|
||||
|
||||
rf = self.calc_receptive_field_size(layers)
|
||||
|
||||
|
|
@ -130,12 +130,14 @@ 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, base_ch = 16):
|
||||
|
||||
def on_build(self, patch_size, in_ch, base_ch = 16, use_fp16 = False):
|
||||
self.use_fp16 = use_fp16
|
||||
conv_dtype = tf.float16 if use_fp16 else tf.float32
|
||||
|
||||
class ResidualBlock(nn.ModelBase):
|
||||
def on_build(self, ch, kernel_size=3 ):
|
||||
self.conv1 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME')
|
||||
self.conv2 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME')
|
||||
self.conv1 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME', dtype=conv_dtype)
|
||||
self.conv2 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME', dtype=conv_dtype)
|
||||
|
||||
def forward(self, inp):
|
||||
x = self.conv1(inp)
|
||||
|
|
@ -146,52 +148,47 @@ class UNetPatchDiscriminator(nn.ModelBase):
|
|||
|
||||
prev_ch = in_ch
|
||||
self.convs = []
|
||||
self.res1 = []
|
||||
self.res2 = []
|
||||
self.upconvs = []
|
||||
self.upres1 = []
|
||||
self.upres2 = []
|
||||
layers = self.find_archi(patch_size)
|
||||
|
||||
|
||||
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')
|
||||
self.in_conv = nn.Conv2D( in_ch, level_chs[-1], kernel_size=1, padding='VALID', dtype=conv_dtype)
|
||||
|
||||
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.convs.append ( nn.Conv2D( level_chs[i-1], level_chs[i], kernel_size=kernel_size, strides=strides, padding='SAME', dtype=conv_dtype) )
|
||||
|
||||
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', dtype=conv_dtype) )
|
||||
|
||||
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.out_conv = nn.Conv2D( level_chs[-1]*2, 1, kernel_size=1, padding='VALID', dtype=conv_dtype)
|
||||
|
||||
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')
|
||||
self.center_conv = nn.Conv2D( level_chs[len(layers)-1], level_chs[len(layers)-1], kernel_size=1, padding='VALID')
|
||||
self.center_out = nn.Conv2D( level_chs[len(layers)-1], 1, kernel_size=1, padding='VALID', dtype=conv_dtype)
|
||||
self.center_conv = nn.Conv2D( level_chs[len(layers)-1], level_chs[len(layers)-1], kernel_size=1, padding='VALID', dtype=conv_dtype)
|
||||
|
||||
|
||||
def forward(self, x):
|
||||
if self.use_fp16:
|
||||
x = tf.cast(x, tf.float16)
|
||||
|
||||
x = tf.nn.leaky_relu( self.in_conv(x), 0.2 )
|
||||
|
||||
encs = []
|
||||
for conv, res1,res2 in zip(self.convs, self.res1, self.res2):
|
||||
for conv in self.convs:
|
||||
encs.insert(0, 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 )
|
||||
|
||||
for i, (upconv, enc, upres1, upres2 ) in enumerate(zip(self.upconvs, encs, self.upres1, self.upres2)):
|
||||
for i, (upconv, enc) in enumerate(zip(self.upconvs, encs)):
|
||||
x = tf.nn.leaky_relu( upconv(x), 0.2 )
|
||||
x = tf.concat( [enc, x], axis=nn.conv2d_ch_axis)
|
||||
x = upres1(x)
|
||||
x = upres2(x)
|
||||
|
||||
return center_out, self.out_conv(x)
|
||||
x = self.out_conv(x)
|
||||
|
||||
if self.use_fp16:
|
||||
center_out = tf.cast(center_out, tf.float32)
|
||||
x = tf.cast(x, tf.float32)
|
||||
|
||||
return center_out, x
|
||||
|
||||
nn.UNetPatchDiscriminator = UNetPatchDiscriminator
|
||||
|
|
|
|||
|
|
@ -50,11 +50,11 @@ class AdaBelief(nn.OptimizerBase):
|
|||
updates = []
|
||||
|
||||
if self.clipnorm > 0.0:
|
||||
norm = tf.sqrt( sum([tf.reduce_sum(tf.square(g)) for g,v in grads_vars]))
|
||||
norm = tf.sqrt( sum([tf.reduce_sum(tf.square(tf.cast(g, tf.float32))) for g,v in grads_vars]))
|
||||
updates += [ state_ops.assign_add( self.iterations, 1) ]
|
||||
for i, (g,v) in enumerate(grads_vars):
|
||||
if self.clipnorm > 0.0:
|
||||
g = self.tf_clip_norm(g, self.clipnorm, norm)
|
||||
g = self.tf_clip_norm(g, self.clipnorm, tf.cast(norm, g.dtype) )
|
||||
|
||||
ms = self.ms_dict[ v.name ]
|
||||
vs = self.vs_dict[ v.name ]
|
||||
|
|
|
|||
|
|
@ -47,11 +47,11 @@ class RMSprop(nn.OptimizerBase):
|
|||
updates = []
|
||||
|
||||
if self.clipnorm > 0.0:
|
||||
norm = tf.sqrt( sum([tf.reduce_sum(tf.square(g)) for g,v in grads_vars]))
|
||||
norm = tf.sqrt( sum([tf.reduce_sum(tf.square(tf.cast(g, tf.float32))) for g,v in grads_vars]))
|
||||
updates += [ state_ops.assign_add( self.iterations, 1) ]
|
||||
for i, (g,v) in enumerate(grads_vars):
|
||||
if self.clipnorm > 0.0:
|
||||
g = self.tf_clip_norm(g, self.clipnorm, norm)
|
||||
g = self.tf_clip_norm(g, self.clipnorm, tf.cast(norm, g.dtype) )
|
||||
|
||||
a = self.accumulators_dict[ v.name ]
|
||||
|
||||
|
|
|
|||
53
main.py
53
main.py
|
|
@ -23,7 +23,7 @@ if __name__ == "__main__":
|
|||
setattr(namespace, self.dest, os.path.abspath(os.path.expanduser(values)))
|
||||
|
||||
exit_code = 0
|
||||
|
||||
|
||||
parser = argparse.ArgumentParser()
|
||||
subparsers = parser.add_subparsers()
|
||||
|
||||
|
|
@ -52,9 +52,9 @@ if __name__ == "__main__":
|
|||
p.add_argument('--output-debug', action="store_true", dest="output_debug", default=None, help="Writes debug images to <output-dir>_debug\ directory.")
|
||||
p.add_argument('--no-output-debug', action="store_false", dest="output_debug", default=None, help="Don't writes debug images to <output-dir>_debug\ directory.")
|
||||
p.add_argument('--face-type', dest="face_type", choices=['half_face', 'full_face', 'whole_face', 'head', 'mark_only'], default=None)
|
||||
p.add_argument('--max-faces-from-image', type=int, dest="max_faces_from_image", default=None, help="Max faces from image.")
|
||||
p.add_argument('--max-faces-from-image', type=int, dest="max_faces_from_image", default=None, help="Max faces from image.")
|
||||
p.add_argument('--image-size', type=int, dest="image_size", default=None, help="Output image size.")
|
||||
p.add_argument('--jpeg-quality', type=int, dest="jpeg_quality", default=None, help="Jpeg quality.")
|
||||
p.add_argument('--jpeg-quality', type=int, dest="jpeg_quality", default=None, help="Jpeg quality.")
|
||||
p.add_argument('--manual-fix', action="store_true", dest="manual_fix", default=False, help="Enables manual extract only frames where faces were not recognized.")
|
||||
p.add_argument('--manual-output-debug-fix', action="store_true", dest="manual_output_debug_fix", default=False, help="Performs manual reextract input-dir frames which were deleted from [output_dir]_debug\ dir.")
|
||||
p.add_argument('--manual-window-size', type=int, dest="manual_window_size", default=1368, help="Manual fix window size. Default: 1368.")
|
||||
|
|
@ -127,7 +127,6 @@ if __name__ == "__main__":
|
|||
'silent_start' : arguments.silent_start,
|
||||
'execute_programs' : [ [int(x[0]), x[1] ] for x in arguments.execute_program ],
|
||||
'debug' : arguments.debug,
|
||||
'dump_ckpt' : arguments.dump_ckpt,
|
||||
}
|
||||
from mainscripts import Trainer
|
||||
Trainer.main(**kwargs)
|
||||
|
|
@ -145,11 +144,19 @@ if __name__ == "__main__":
|
|||
p.add_argument('--cpu-only', action="store_true", dest="cpu_only", default=False, help="Train on CPU.")
|
||||
p.add_argument('--force-gpu-idxs', dest="force_gpu_idxs", default=None, help="Force to choose GPU indexes separated by comma.")
|
||||
p.add_argument('--silent-start', action="store_true", dest="silent_start", default=False, help="Silent start. Automatically chooses Best GPU and last used model.")
|
||||
p.add_argument('--dump-ckpt', action="store_true", dest="dump_ckpt", default=False, help="Dump the model to ckpt format.")
|
||||
|
||||
|
||||
|
||||
p.add_argument('--execute-program', dest="execute_program", default=[], action='append', nargs='+')
|
||||
p.set_defaults (func=process_train)
|
||||
|
||||
def process_exportdfm(arguments):
|
||||
osex.set_process_lowest_prio()
|
||||
from mainscripts import ExportDFM
|
||||
ExportDFM.main(model_class_name = arguments.model_name, saved_models_path = Path(arguments.model_dir))
|
||||
|
||||
p = subparsers.add_parser( "exportdfm", help="Export model to use in DeepFaceLive.")
|
||||
p.add_argument('--model-dir', required=True, action=fixPathAction, dest="model_dir", help="Saved models dir.")
|
||||
p.add_argument('--model', required=True, dest="model_name", choices=pathex.get_all_dir_names_startswith ( Path(__file__).parent / 'models' , 'Model_'), help="Model class name.")
|
||||
p.set_defaults (func=process_exportdfm)
|
||||
|
||||
def process_merge(arguments):
|
||||
osex.set_process_lowest_prio()
|
||||
|
|
@ -254,8 +261,8 @@ if __name__ == "__main__":
|
|||
p.add_argument('--force-gpu-idxs', dest="force_gpu_idxs", default=None, help="Force to choose GPU indexes separated by comma.")
|
||||
|
||||
p.set_defaults(func=process_faceset_enhancer)
|
||||
|
||||
|
||||
|
||||
|
||||
p = facesettool_parser.add_parser ("resize", help="Resize DFL faceset.")
|
||||
p.add_argument('--input-dir', required=True, action=fixPathAction, dest="input_dir", help="Input directory of aligned faces.")
|
||||
|
||||
|
|
@ -264,7 +271,7 @@ if __name__ == "__main__":
|
|||
from mainscripts import FacesetResizer
|
||||
FacesetResizer.process_folder ( Path(arguments.input_dir) )
|
||||
p.set_defaults(func=process_faceset_resizer)
|
||||
|
||||
|
||||
def process_dev_test(arguments):
|
||||
osex.set_process_lowest_prio()
|
||||
from mainscripts import dev_misc
|
||||
|
|
@ -273,10 +280,10 @@ if __name__ == "__main__":
|
|||
p = subparsers.add_parser( "dev_test", help="")
|
||||
p.add_argument('--input-dir', required=True, action=fixPathAction, dest="input_dir")
|
||||
p.set_defaults (func=process_dev_test)
|
||||
|
||||
|
||||
# ========== XSeg
|
||||
xseg_parser = subparsers.add_parser( "xseg", help="XSeg tools.").add_subparsers()
|
||||
|
||||
|
||||
p = xseg_parser.add_parser( "editor", help="XSeg editor.")
|
||||
|
||||
def process_xsegeditor(arguments):
|
||||
|
|
@ -284,11 +291,11 @@ if __name__ == "__main__":
|
|||
from XSegEditor import XSegEditor
|
||||
global exit_code
|
||||
exit_code = XSegEditor.start (Path(arguments.input_dir))
|
||||
|
||||
|
||||
p.add_argument('--input-dir', required=True, action=fixPathAction, dest="input_dir")
|
||||
|
||||
p.set_defaults (func=process_xsegeditor)
|
||||
|
||||
|
||||
p = xseg_parser.add_parser( "apply", help="Apply trained XSeg model to the extracted faces.")
|
||||
|
||||
def process_xsegapply(arguments):
|
||||
|
|
@ -298,8 +305,8 @@ if __name__ == "__main__":
|
|||
p.add_argument('--input-dir', required=True, action=fixPathAction, dest="input_dir")
|
||||
p.add_argument('--model-dir', required=True, action=fixPathAction, dest="model_dir")
|
||||
p.set_defaults (func=process_xsegapply)
|
||||
|
||||
|
||||
|
||||
|
||||
p = xseg_parser.add_parser( "remove", help="Remove applied XSeg masks from the extracted faces.")
|
||||
def process_xsegremove(arguments):
|
||||
osex.set_process_lowest_prio()
|
||||
|
|
@ -307,8 +314,8 @@ if __name__ == "__main__":
|
|||
XSegUtil.remove_xseg (Path(arguments.input_dir) )
|
||||
p.add_argument('--input-dir', required=True, action=fixPathAction, dest="input_dir")
|
||||
p.set_defaults (func=process_xsegremove)
|
||||
|
||||
|
||||
|
||||
|
||||
p = xseg_parser.add_parser( "remove_labels", help="Remove XSeg labels from the extracted faces.")
|
||||
def process_xsegremovelabels(arguments):
|
||||
osex.set_process_lowest_prio()
|
||||
|
|
@ -316,8 +323,8 @@ if __name__ == "__main__":
|
|||
XSegUtil.remove_xseg_labels (Path(arguments.input_dir) )
|
||||
p.add_argument('--input-dir', required=True, action=fixPathAction, dest="input_dir")
|
||||
p.set_defaults (func=process_xsegremovelabels)
|
||||
|
||||
|
||||
|
||||
|
||||
p = xseg_parser.add_parser( "fetch", help="Copies faces containing XSeg polygons in <input_dir>_xseg dir.")
|
||||
|
||||
def process_xsegfetch(arguments):
|
||||
|
|
@ -326,7 +333,7 @@ if __name__ == "__main__":
|
|||
XSegUtil.fetch_xseg (Path(arguments.input_dir) )
|
||||
p.add_argument('--input-dir', required=True, action=fixPathAction, dest="input_dir")
|
||||
p.set_defaults (func=process_xsegfetch)
|
||||
|
||||
|
||||
def bad_args(arguments):
|
||||
parser.print_help()
|
||||
exit(0)
|
||||
|
|
@ -337,9 +344,9 @@ if __name__ == "__main__":
|
|||
|
||||
if exit_code == 0:
|
||||
print ("Done.")
|
||||
|
||||
|
||||
exit(exit_code)
|
||||
|
||||
|
||||
'''
|
||||
import code
|
||||
code.interact(local=dict(globals(), **locals()))
|
||||
|
|
|
|||
22
mainscripts/ExportDFM.py
Normal file
22
mainscripts/ExportDFM.py
Normal file
|
|
@ -0,0 +1,22 @@
|
|||
import os
|
||||
import sys
|
||||
import traceback
|
||||
import queue
|
||||
import threading
|
||||
import time
|
||||
import numpy as np
|
||||
import itertools
|
||||
from pathlib import Path
|
||||
from core import pathex
|
||||
from core import imagelib
|
||||
import cv2
|
||||
import models
|
||||
from core.interact import interact as io
|
||||
|
||||
|
||||
def main(model_class_name, saved_models_path):
|
||||
model = models.import_model(model_class_name)(
|
||||
is_exporting=True,
|
||||
saved_models_path=saved_models_path,
|
||||
cpu_only=True)
|
||||
model.export_dfm ()
|
||||
|
|
@ -79,79 +79,79 @@ class FacesetResizerSubprocessor(Subprocessor):
|
|||
h,w = img.shape[:2]
|
||||
if h != w:
|
||||
raise Exception(f'w != h in {filepath}')
|
||||
|
||||
|
||||
image_size = self.image_size
|
||||
face_type = self.face_type
|
||||
output_filepath = self.output_dirpath / filepath.name
|
||||
|
||||
|
||||
if face_type is not None:
|
||||
lmrks = dflimg.get_landmarks()
|
||||
mat = LandmarksProcessor.get_transform_mat(lmrks, image_size, face_type)
|
||||
|
||||
|
||||
img = cv2.warpAffine(img, mat, (image_size, image_size), flags=cv2.INTER_LANCZOS4 )
|
||||
img = np.clip(img, 0, 255).astype(np.uint8)
|
||||
|
||||
|
||||
cv2_imwrite ( str(output_filepath), img, [int(cv2.IMWRITE_JPEG_QUALITY), 100] )
|
||||
|
||||
dfl_dict = dflimg.get_dict()
|
||||
dflimg = DFLIMG.load (output_filepath)
|
||||
dflimg.set_dict(dfl_dict)
|
||||
|
||||
|
||||
xseg_mask = dflimg.get_xseg_mask()
|
||||
if xseg_mask is not None:
|
||||
xseg_res = 256
|
||||
|
||||
|
||||
xseg_lmrks = lmrks.copy()
|
||||
xseg_lmrks *= (xseg_res / w)
|
||||
xseg_mat = LandmarksProcessor.get_transform_mat(xseg_lmrks, xseg_res, face_type)
|
||||
|
||||
|
||||
xseg_mask = cv2.warpAffine(xseg_mask, xseg_mat, (xseg_res, xseg_res), flags=cv2.INTER_LANCZOS4 )
|
||||
xseg_mask[xseg_mask < 0.5] = 0
|
||||
xseg_mask[xseg_mask >= 0.5] = 1
|
||||
|
||||
dflimg.set_xseg_mask(xseg_mask)
|
||||
|
||||
|
||||
seg_ie_polys = dflimg.get_seg_ie_polys()
|
||||
|
||||
|
||||
for poly in seg_ie_polys.get_polys():
|
||||
poly_pts = poly.get_pts()
|
||||
poly_pts = LandmarksProcessor.transform_points(poly_pts, mat)
|
||||
poly.set_points(poly_pts)
|
||||
|
||||
|
||||
dflimg.set_seg_ie_polys(seg_ie_polys)
|
||||
|
||||
|
||||
lmrks = LandmarksProcessor.transform_points(lmrks, mat)
|
||||
dflimg.set_landmarks(lmrks)
|
||||
|
||||
|
||||
image_to_face_mat = dflimg.get_image_to_face_mat()
|
||||
if image_to_face_mat is not None:
|
||||
image_to_face_mat = LandmarksProcessor.get_transform_mat ( dflimg.get_source_landmarks(), image_size, face_type )
|
||||
dflimg.set_image_to_face_mat(image_to_face_mat)
|
||||
dflimg.set_face_type( FaceType.toString(face_type) )
|
||||
dflimg.save()
|
||||
|
||||
|
||||
else:
|
||||
dfl_dict = dflimg.get_dict()
|
||||
|
||||
|
||||
scale = w / image_size
|
||||
|
||||
img = cv2.resize(img, (image_size, image_size), interpolation=cv2.INTER_LANCZOS4)
|
||||
|
||||
|
||||
img = cv2.resize(img, (image_size, image_size), interpolation=cv2.INTER_LANCZOS4)
|
||||
|
||||
cv2_imwrite ( str(output_filepath), img, [int(cv2.IMWRITE_JPEG_QUALITY), 100] )
|
||||
|
||||
dflimg = DFLIMG.load (output_filepath)
|
||||
dflimg.set_dict(dfl_dict)
|
||||
|
||||
lmrks = dflimg.get_landmarks()
|
||||
|
||||
lmrks = dflimg.get_landmarks()
|
||||
lmrks /= scale
|
||||
dflimg.set_landmarks(lmrks)
|
||||
|
||||
|
||||
seg_ie_polys = dflimg.get_seg_ie_polys()
|
||||
seg_ie_polys.mult_points( 1.0 / scale)
|
||||
dflimg.set_seg_ie_polys(seg_ie_polys)
|
||||
|
||||
|
||||
image_to_face_mat = dflimg.get_image_to_face_mat()
|
||||
|
||||
|
||||
if image_to_face_mat is not None:
|
||||
face_type = FaceType.fromString ( dflimg.get_face_type() )
|
||||
image_to_face_mat = LandmarksProcessor.get_transform_mat ( dflimg.get_source_landmarks(), image_size, face_type )
|
||||
|
|
@ -165,9 +165,9 @@ class FacesetResizerSubprocessor(Subprocessor):
|
|||
return (0, filepath, None)
|
||||
|
||||
def process_folder ( dirpath):
|
||||
|
||||
image_size = io.input_int(f"New image size", 512, valid_range=[256,2048])
|
||||
|
||||
|
||||
image_size = io.input_int(f"New image size", 512, valid_range=[128,2048])
|
||||
|
||||
face_type = io.input_str ("Change face type", 'same', ['h','mf','f','wf','head','same']).lower()
|
||||
if face_type == 'same':
|
||||
face_type = None
|
||||
|
|
@ -177,7 +177,7 @@ def process_folder ( dirpath):
|
|||
'f' : FaceType.FULL,
|
||||
'wf' : FaceType.WHOLE_FACE,
|
||||
'head' : FaceType.HEAD}[face_type]
|
||||
|
||||
|
||||
|
||||
output_dirpath = dirpath.parent / (dirpath.name + '_resized')
|
||||
output_dirpath.mkdir (exist_ok=True, parents=True)
|
||||
|
|
|
|||
|
|
@ -49,6 +49,7 @@ def main (model_class_name=None,
|
|||
model = models.import_model(model_class_name)(is_training=False,
|
||||
saved_models_path=saved_models_path,
|
||||
force_gpu_idxs=force_gpu_idxs,
|
||||
force_model_name=force_model_name,
|
||||
cpu_only=cpu_only)
|
||||
|
||||
predictor_func, predictor_input_shape, cfg = model.get_MergerConfig()
|
||||
|
|
|
|||
|
|
@ -27,7 +27,6 @@ def trainerThread (s2c, c2s, e,
|
|||
silent_start=False,
|
||||
execute_programs = None,
|
||||
debug=False,
|
||||
dump_ckpt=False,
|
||||
**kwargs):
|
||||
while True:
|
||||
try:
|
||||
|
|
@ -43,12 +42,9 @@ def trainerThread (s2c, c2s, e,
|
|||
|
||||
if not saved_models_path.exists():
|
||||
saved_models_path.mkdir(exist_ok=True, parents=True)
|
||||
|
||||
if dump_ckpt:
|
||||
cpu_only=True
|
||||
|
||||
|
||||
model = models.import_model(model_class_name)(
|
||||
is_training=not dump_ckpt,
|
||||
is_training=True,
|
||||
saved_models_path=saved_models_path,
|
||||
training_data_src_path=training_data_src_path,
|
||||
training_data_dst_path=training_data_dst_path,
|
||||
|
|
@ -61,11 +57,6 @@ def trainerThread (s2c, c2s, e,
|
|||
silent_start=silent_start,
|
||||
debug=debug)
|
||||
|
||||
if dump_ckpt:
|
||||
e.set()
|
||||
model.dump_ckpt()
|
||||
break
|
||||
|
||||
is_reached_goal = model.is_reached_iter_goal()
|
||||
|
||||
shared_state = { 'after_save' : False }
|
||||
|
|
@ -76,10 +67,10 @@ def trainerThread (s2c, c2s, e,
|
|||
io.log_info ("Saving....", end='\r')
|
||||
model.save()
|
||||
shared_state['after_save'] = True
|
||||
|
||||
|
||||
def model_backup():
|
||||
if not debug and not is_reached_goal:
|
||||
model.create_backup()
|
||||
model.create_backup()
|
||||
|
||||
def send_preview():
|
||||
if not debug:
|
||||
|
|
@ -128,7 +119,7 @@ def trainerThread (s2c, c2s, e,
|
|||
io.log_info("")
|
||||
io.log_info("Trying to do the first iteration. If an error occurs, reduce the model parameters.")
|
||||
io.log_info("")
|
||||
|
||||
|
||||
if sys.platform[0:3] == 'win':
|
||||
io.log_info("!!!")
|
||||
io.log_info("Windows 10 users IMPORTANT notice. You should set this setting in order to work correctly.")
|
||||
|
|
@ -146,7 +137,7 @@ def trainerThread (s2c, c2s, e,
|
|||
|
||||
if shared_state['after_save']:
|
||||
shared_state['after_save'] = False
|
||||
|
||||
|
||||
mean_loss = np.mean ( loss_history[save_iter:iter], axis=0)
|
||||
|
||||
for loss_value in mean_loss:
|
||||
|
|
|
|||
|
|
@ -22,6 +22,7 @@ from samplelib import SampleGeneratorBase
|
|||
|
||||
class ModelBase(object):
|
||||
def __init__(self, is_training=False,
|
||||
is_exporting=False,
|
||||
saved_models_path=None,
|
||||
training_data_src_path=None,
|
||||
training_data_dst_path=None,
|
||||
|
|
@ -36,6 +37,7 @@ class ModelBase(object):
|
|||
silent_start=False,
|
||||
**kwargs):
|
||||
self.is_training = is_training
|
||||
self.is_exporting = is_exporting
|
||||
self.saved_models_path = saved_models_path
|
||||
self.training_data_src_path = training_data_src_path
|
||||
self.training_data_dst_path = training_data_dst_path
|
||||
|
|
@ -232,7 +234,7 @@ class ModelBase(object):
|
|||
preview_id_counter = 0
|
||||
while not choosed:
|
||||
self.sample_for_preview = self.generate_next_samples()
|
||||
previews = self.get_static_previews()
|
||||
previews = self.get_history_previews()
|
||||
|
||||
io.show_image( wnd_name, ( previews[preview_id_counter % len(previews) ][1] *255).astype(np.uint8) )
|
||||
|
||||
|
|
@ -258,7 +260,7 @@ class ModelBase(object):
|
|||
self.sample_for_preview = self.generate_next_samples()
|
||||
|
||||
try:
|
||||
self.get_static_previews()
|
||||
self.get_history_previews()
|
||||
except:
|
||||
self.sample_for_preview = self.generate_next_samples()
|
||||
|
||||
|
|
@ -347,7 +349,7 @@ class ModelBase(object):
|
|||
return ( ('loss_src', 0), ('loss_dst', 0) )
|
||||
|
||||
#overridable
|
||||
def onGetPreview(self, sample):
|
||||
def onGetPreview(self, sample, for_history=False):
|
||||
#you can return multiple previews
|
||||
#return [ ('preview_name',preview_rgb), ... ]
|
||||
return []
|
||||
|
|
@ -377,8 +379,8 @@ class ModelBase(object):
|
|||
def get_previews(self):
|
||||
return self.onGetPreview ( self.last_sample )
|
||||
|
||||
def get_static_previews(self):
|
||||
return self.onGetPreview (self.sample_for_preview)
|
||||
def get_history_previews(self):
|
||||
return self.onGetPreview (self.sample_for_preview, for_history=True)
|
||||
|
||||
def get_preview_history_writer(self):
|
||||
if self.preview_history_writer is None:
|
||||
|
|
@ -484,7 +486,7 @@ class ModelBase(object):
|
|||
plist += [ (bgr, self.get_strpath_storage_for_file('preview_%s.jpg' % (name) ) ) ]
|
||||
|
||||
if self.write_preview_history:
|
||||
previews = self.get_static_previews()
|
||||
previews = self.get_history_previews()
|
||||
for i in range(len(previews)):
|
||||
name, bgr = previews[i]
|
||||
path = self.preview_history_path / name
|
||||
|
|
|
|||
|
|
@ -18,41 +18,26 @@ class AMPModel(ModelBase):
|
|||
def on_initialize_options(self):
|
||||
device_config = nn.getCurrentDeviceConfig()
|
||||
|
||||
lowest_vram = 2
|
||||
if len(device_config.devices) != 0:
|
||||
lowest_vram = device_config.devices.get_worst_device().total_mem_gb
|
||||
|
||||
if lowest_vram >= 4:
|
||||
suggest_batch_size = 8
|
||||
else:
|
||||
suggest_batch_size = 4
|
||||
|
||||
yn_str = {True:'y',False:'n'}
|
||||
min_res = 64
|
||||
max_res = 640
|
||||
|
||||
default_resolution = self.options['resolution'] = self.load_or_def_option('resolution', 224)
|
||||
default_face_type = self.options['face_type'] = self.load_or_def_option('face_type', 'wf')
|
||||
default_models_opt_on_gpu = self.options['models_opt_on_gpu'] = self.load_or_def_option('models_opt_on_gpu', True)
|
||||
|
||||
default_ae_dims = self.options['ae_dims'] = self.load_or_def_option('ae_dims', 256)
|
||||
|
||||
inter_dims = self.load_or_def_option('inter_dims', None)
|
||||
if inter_dims is None:
|
||||
inter_dims = self.options['ae_dims']
|
||||
default_inter_dims = self.options['inter_dims'] = inter_dims
|
||||
|
||||
default_e_dims = self.options['e_dims'] = self.load_or_def_option('e_dims', 64)
|
||||
default_d_dims = self.options['d_dims'] = self.options.get('d_dims', None)
|
||||
default_d_mask_dims = self.options['d_mask_dims'] = self.options.get('d_mask_dims', None)
|
||||
default_morph_factor = self.options['morph_factor'] = self.options.get('morph_factor', 0.33)
|
||||
default_masked_training = self.options['masked_training'] = self.load_or_def_option('masked_training', True)
|
||||
default_eyes_mouth_prio = self.options['eyes_mouth_prio'] = self.load_or_def_option('eyes_mouth_prio', True)
|
||||
default_morph_factor = self.options['morph_factor'] = self.options.get('morph_factor', 0.5)
|
||||
default_uniform_yaw = self.options['uniform_yaw'] = self.load_or_def_option('uniform_yaw', False)
|
||||
|
||||
lr_dropout = self.load_or_def_option('lr_dropout', 'n')
|
||||
lr_dropout = {True:'y', False:'n'}.get(lr_dropout, lr_dropout) #backward comp
|
||||
default_lr_dropout = self.options['lr_dropout'] = lr_dropout
|
||||
|
||||
default_random_warp = self.options['random_warp'] = self.load_or_def_option('random_warp', True)
|
||||
default_ct_mode = self.options['ct_mode'] = self.load_or_def_option('ct_mode', 'none')
|
||||
default_clipgrad = self.options['clipgrad'] = self.load_or_def_option('clipgrad', False)
|
||||
default_pretrain = self.options['pretrain'] = self.load_or_def_option('pretrain', False)
|
||||
|
||||
|
||||
ask_override = self.ask_override()
|
||||
if self.is_first_run() or ask_override:
|
||||
|
|
@ -61,13 +46,13 @@ class AMPModel(ModelBase):
|
|||
self.ask_target_iter()
|
||||
self.ask_random_src_flip()
|
||||
self.ask_random_dst_flip()
|
||||
self.ask_batch_size(suggest_batch_size)
|
||||
self.ask_batch_size(8)
|
||||
|
||||
if self.is_first_run():
|
||||
resolution = io.input_int("Resolution", default_resolution, add_info="64-640", help_message="More resolution requires more VRAM and time to train. Value will be adjusted to multiple of 32 .")
|
||||
resolution = np.clip ( (resolution // 32) * 32, min_res, max_res)
|
||||
resolution = np.clip ( (resolution // 32) * 32, 64, 640)
|
||||
self.options['resolution'] = resolution
|
||||
self.options['face_type'] = io.input_str ("Face type", default_face_type, ['wf','head'], help_message="whole face / head").lower()
|
||||
self.options['face_type'] = io.input_str ("Face type", default_face_type, ['f','wf','head'], help_message="whole face / head").lower()
|
||||
|
||||
|
||||
default_d_dims = self.options['d_dims'] = self.load_or_def_option('d_dims', 64)
|
||||
|
|
@ -77,7 +62,8 @@ class AMPModel(ModelBase):
|
|||
default_d_mask_dims = self.options['d_mask_dims'] = self.load_or_def_option('d_mask_dims', default_d_mask_dims)
|
||||
|
||||
if self.is_first_run():
|
||||
self.options['ae_dims'] = np.clip ( io.input_int("AutoEncoder dimensions", default_ae_dims, add_info="32-1024", help_message="All face information will packed to AE dims. If amount of AE dims are not enough, then for example closed eyes will not be recognized. More dims are better, but require more VRAM. You can fine-tune model size to fit your GPU." ), 32, 1024 )
|
||||
self.options['ae_dims'] = np.clip ( io.input_int("AutoEncoder dimensions", default_ae_dims, add_info="32-1024", help_message="All face information will packed to AE dims. If amount of AE dims are not enough, then for example closed eyes will not be recognized. More dims are better, but require more VRAM. You can fine-tune model size to fit your GPU." ), 32, 1024 )
|
||||
self.options['inter_dims'] = np.clip ( io.input_int("Inter dimensions", default_inter_dims, add_info="32-2048", help_message="Should be equal or more than AutoEncoder dimensions. More dims are better, but require more VRAM. You can fine-tune model size to fit your GPU." ), 32, 2048 )
|
||||
|
||||
e_dims = np.clip ( io.input_int("Encoder dimensions", default_e_dims, add_info="16-256", help_message="More dims help to recognize more facial features and achieve sharper result, but require more VRAM. You can fine-tune model size to fit your GPU." ), 16, 256 )
|
||||
self.options['e_dims'] = e_dims + e_dims % 2
|
||||
|
|
@ -88,15 +74,10 @@ class AMPModel(ModelBase):
|
|||
d_mask_dims = np.clip ( io.input_int("Decoder mask dimensions", default_d_mask_dims, add_info="16-256", help_message="Typical mask dimensions = decoder dimensions / 3. If you manually cut out obstacles from the dst mask, you can increase this parameter to achieve better quality." ), 16, 256 )
|
||||
self.options['d_mask_dims'] = d_mask_dims + d_mask_dims % 2
|
||||
|
||||
morph_factor = np.clip ( io.input_number ("Morph factor.", default_morph_factor, add_info="0.1 .. 0.5", help_message="The smaller the value, the more src-like facial expressions will appear. The larger the value, the less space there is to train a large dst faceset in the neural network. Typical fine value is 0.33"), 0.1, 0.5 )
|
||||
morph_factor = np.clip ( io.input_number ("Morph factor.", default_morph_factor, add_info="0.1 .. 0.5", help_message="Typical fine value is 0.5"), 0.1, 0.5 )
|
||||
self.options['morph_factor'] = morph_factor
|
||||
|
||||
|
||||
if self.is_first_run() or ask_override:
|
||||
if self.options['face_type'] == 'wf' or self.options['face_type'] == 'head':
|
||||
self.options['masked_training'] = io.input_bool ("Masked training", default_masked_training, help_message="This option is available only for 'whole_face' or 'head' type. Masked training clips training area to full_face mask or XSeg mask, thus network will train the faces properly.")
|
||||
|
||||
self.options['eyes_mouth_prio'] = io.input_bool ("Eyes and mouth priority", default_eyes_mouth_prio, help_message='Helps to fix eye problems during training like "alien eyes" and wrong eyes direction. Also makes the detail of the teeth higher.')
|
||||
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)
|
||||
|
|
@ -106,26 +87,21 @@ class AMPModel(ModelBase):
|
|||
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.")
|
||||
|
||||
self.options['lr_dropout'] = io.input_str (f"Use learning rate dropout", default_lr_dropout, ['n','y','cpu'], help_message="When the face is trained enough, you can enable this option to get extra sharpness and reduce subpixel shake for less amount of iterations. Enabled it before `disable random warp` and before GAN. \nn - disabled.\ny - enabled\ncpu - enabled on CPU. This allows not to use extra VRAM, sacrificing 20% time of iteration.")
|
||||
|
||||
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 .. 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 )
|
||||
self.options['gan_power'] = np.clip ( io.input_number ("GAN power", default_gan_power, add_info="0.0 .. 5.0", help_message="Forces the neural network to learn small details of the face. Enable it only when the face is trained enough with 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, 5.0 )
|
||||
|
||||
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 )
|
||||
gan_dims = np.clip ( io.input_int("GAN dimensions", default_gan_dims, add_info="4-512", 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, 512 )
|
||||
self.options['gan_dims'] = gan_dims
|
||||
|
||||
self.options['ct_mode'] = io.input_str (f"Color transfer for src faceset", default_ct_mode, ['none','rct','lct','mkl','idt','sot'], help_message="Change color distribution of src samples close to dst samples. Try all modes to find the best.")
|
||||
self.options['clipgrad'] = io.input_bool ("Enable gradient clipping", default_clipgrad, help_message="Gradient clipping reduces chance of model collapse, sacrificing speed of training.")
|
||||
|
||||
self.options['pretrain'] = io.input_bool ("Enable pretraining mode", default_pretrain, help_message="Pretrain the model with large amount of various faces. After that, model can be used to train the fakes more quickly. Forces random_warp=N, random_flips=Y, gan_power=0.0, lr_dropout=N, uniform_yaw=Y")
|
||||
|
||||
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)
|
||||
|
||||
#override
|
||||
def on_initialize(self):
|
||||
|
|
@ -135,42 +111,47 @@ class AMPModel(ModelBase):
|
|||
nn.initialize(data_format=self.model_data_format)
|
||||
tf = nn.tf
|
||||
|
||||
self.resolution = resolution = self.options['resolution']
|
||||
input_ch=3
|
||||
resolution = self.resolution = self.options['resolution']
|
||||
e_dims = self.options['e_dims']
|
||||
ae_dims = self.options['ae_dims']
|
||||
inter_dims = self.inter_dims = self.options['inter_dims']
|
||||
inter_res = self.inter_res = resolution // 32
|
||||
d_dims = self.options['d_dims']
|
||||
d_mask_dims = self.options['d_mask_dims']
|
||||
face_type = self.face_type = {'f' : FaceType.FULL,
|
||||
'wf' : FaceType.WHOLE_FACE,
|
||||
'head' : FaceType.HEAD}[ self.options['face_type'] ]
|
||||
morph_factor = self.options['morph_factor']
|
||||
gan_power = self.gan_power = self.options['gan_power']
|
||||
random_warp = self.options['random_warp']
|
||||
|
||||
lowest_dense_res = self.lowest_dense_res = resolution // 32
|
||||
ct_mode = self.options['ct_mode']
|
||||
if ct_mode == 'none':
|
||||
ct_mode = None
|
||||
|
||||
use_fp16 = self.is_exporting
|
||||
conv_dtype = tf.float16 if use_fp16 else tf.float32
|
||||
|
||||
class Downscale(nn.ModelBase):
|
||||
def __init__(self, in_ch, out_ch, kernel_size=5, *kwargs ):
|
||||
self.in_ch = in_ch
|
||||
self.out_ch = out_ch
|
||||
self.kernel_size = kernel_size
|
||||
super().__init__(*kwargs)
|
||||
|
||||
def on_build(self, *args, **kwargs ):
|
||||
self.conv1 = nn.Conv2D( self.in_ch, self.out_ch, kernel_size=self.kernel_size, strides=2, padding='SAME')
|
||||
def on_build(self, in_ch, out_ch, kernel_size=5 ):
|
||||
self.conv1 = nn.Conv2D( in_ch, out_ch, kernel_size=kernel_size, strides=2, padding='SAME', dtype=conv_dtype)
|
||||
|
||||
def forward(self, x):
|
||||
x = self.conv1(x)
|
||||
x = tf.nn.leaky_relu(x, 0.1)
|
||||
return x
|
||||
|
||||
def get_out_ch(self):
|
||||
return self.out_ch
|
||||
return tf.nn.leaky_relu(self.conv1(x), 0.1)
|
||||
|
||||
class Upscale(nn.ModelBase):
|
||||
def on_build(self, in_ch, out_ch, kernel_size=3 ):
|
||||
self.conv1 = nn.Conv2D( in_ch, out_ch*4, kernel_size=kernel_size, padding='SAME')
|
||||
self.conv1 = nn.Conv2D(in_ch, out_ch*4, kernel_size=kernel_size, padding='SAME', dtype=conv_dtype)
|
||||
|
||||
def forward(self, x):
|
||||
x = self.conv1(x)
|
||||
x = tf.nn.leaky_relu(x, 0.1)
|
||||
x = nn.depth_to_space(x, 2)
|
||||
x = nn.depth_to_space(tf.nn.leaky_relu(self.conv1(x), 0.1), 2)
|
||||
return x
|
||||
|
||||
class ResidualBlock(nn.ModelBase):
|
||||
def on_build(self, ch, kernel_size=3 ):
|
||||
self.conv1 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME')
|
||||
self.conv2 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME')
|
||||
self.conv1 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME', dtype=conv_dtype)
|
||||
self.conv2 = nn.Conv2D( ch, ch, kernel_size=kernel_size, padding='SAME', dtype=conv_dtype)
|
||||
|
||||
def forward(self, inp):
|
||||
x = self.conv1(inp)
|
||||
|
|
@ -180,18 +161,19 @@ class AMPModel(ModelBase):
|
|||
return x
|
||||
|
||||
class Encoder(nn.ModelBase):
|
||||
def on_build(self, in_ch, e_ch, ae_ch):
|
||||
self.down1 = Downscale(in_ch, e_ch, kernel_size=5)
|
||||
self.res1 = ResidualBlock(e_ch)
|
||||
self.down2 = Downscale(e_ch, e_ch*2, kernel_size=5)
|
||||
self.down3 = Downscale(e_ch*2, e_ch*4, kernel_size=5)
|
||||
self.down4 = Downscale(e_ch*4, e_ch*8, kernel_size=5)
|
||||
self.down5 = Downscale(e_ch*8, e_ch*8, kernel_size=5)
|
||||
self.res5 = ResidualBlock(e_ch*8)
|
||||
self.dense1 = nn.Dense( lowest_dense_res*lowest_dense_res*e_ch*8, ae_ch )
|
||||
def on_build(self):
|
||||
self.down1 = Downscale(input_ch, e_dims, kernel_size=5)
|
||||
self.res1 = ResidualBlock(e_dims)
|
||||
self.down2 = Downscale(e_dims, e_dims*2, kernel_size=5)
|
||||
self.down3 = Downscale(e_dims*2, e_dims*4, kernel_size=5)
|
||||
self.down4 = Downscale(e_dims*4, e_dims*8, kernel_size=5)
|
||||
self.down5 = Downscale(e_dims*8, e_dims*8, kernel_size=5)
|
||||
self.res5 = ResidualBlock(e_dims*8)
|
||||
self.dense1 = nn.Dense( (( resolution//(2**5) )**2) * e_dims*8, ae_dims )
|
||||
|
||||
def forward(self, inp):
|
||||
x = inp
|
||||
def forward(self, x):
|
||||
if use_fp16:
|
||||
x = tf.cast(x, tf.float16)
|
||||
x = self.down1(x)
|
||||
x = self.res1(x)
|
||||
x = self.down2(x)
|
||||
|
|
@ -199,56 +181,51 @@ class AMPModel(ModelBase):
|
|||
x = self.down4(x)
|
||||
x = self.down5(x)
|
||||
x = self.res5(x)
|
||||
x = nn.flatten(x)
|
||||
x = nn.pixel_norm(x, axes=-1)
|
||||
if use_fp16:
|
||||
x = tf.cast(x, tf.float32)
|
||||
x = nn.pixel_norm(nn.flatten(x), axes=-1)
|
||||
x = self.dense1(x)
|
||||
return x
|
||||
|
||||
|
||||
class Inter(nn.ModelBase):
|
||||
def __init__(self, ae_ch, ae_out_ch, **kwargs):
|
||||
self.ae_ch, self.ae_out_ch = ae_ch, ae_out_ch
|
||||
super().__init__(**kwargs)
|
||||
|
||||
def on_build(self):
|
||||
ae_ch, ae_out_ch = self.ae_ch, self.ae_out_ch
|
||||
self.dense2 = nn.Dense( ae_ch, lowest_dense_res * lowest_dense_res * ae_out_ch )
|
||||
self.dense2 = nn.Dense(ae_dims, inter_res * inter_res * inter_dims)
|
||||
|
||||
def forward(self, inp):
|
||||
x = inp
|
||||
x = self.dense2(x)
|
||||
x = nn.reshape_4D (x, lowest_dense_res, lowest_dense_res, self.ae_out_ch)
|
||||
x = nn.reshape_4D (x, inter_res, inter_res, inter_dims)
|
||||
return x
|
||||
|
||||
def get_out_ch(self):
|
||||
return self.ae_out_ch
|
||||
|
||||
class Decoder(nn.ModelBase):
|
||||
def on_build(self, in_ch, d_ch, d_mask_ch ):
|
||||
self.upscale0 = Upscale(in_ch, d_ch*8, kernel_size=3)
|
||||
self.upscale1 = Upscale(d_ch*8, d_ch*8, kernel_size=3)
|
||||
self.upscale2 = Upscale(d_ch*8, d_ch*4, kernel_size=3)
|
||||
self.upscale3 = Upscale(d_ch*4, d_ch*2, kernel_size=3)
|
||||
def on_build(self ):
|
||||
self.upscale0 = Upscale(inter_dims, d_dims*8, kernel_size=3)
|
||||
self.upscale1 = Upscale(d_dims*8, d_dims*8, kernel_size=3)
|
||||
self.upscale2 = Upscale(d_dims*8, d_dims*4, kernel_size=3)
|
||||
self.upscale3 = Upscale(d_dims*4, d_dims*2, kernel_size=3)
|
||||
|
||||
self.res0 = ResidualBlock(d_ch*8, kernel_size=3)
|
||||
self.res1 = ResidualBlock(d_ch*8, kernel_size=3)
|
||||
self.res2 = ResidualBlock(d_ch*4, kernel_size=3)
|
||||
self.res3 = ResidualBlock(d_ch*2, kernel_size=3)
|
||||
self.res0 = ResidualBlock(d_dims*8, kernel_size=3)
|
||||
self.res1 = ResidualBlock(d_dims*8, kernel_size=3)
|
||||
self.res2 = ResidualBlock(d_dims*4, kernel_size=3)
|
||||
self.res3 = ResidualBlock(d_dims*2, kernel_size=3)
|
||||
|
||||
self.upscalem0 = Upscale(in_ch, d_mask_ch*8, kernel_size=3)
|
||||
self.upscalem1 = Upscale(d_mask_ch*8, d_mask_ch*8, kernel_size=3)
|
||||
self.upscalem2 = Upscale(d_mask_ch*8, d_mask_ch*4, kernel_size=3)
|
||||
self.upscalem3 = Upscale(d_mask_ch*4, d_mask_ch*2, kernel_size=3)
|
||||
self.upscalem4 = Upscale(d_mask_ch*2, d_mask_ch*1, kernel_size=3)
|
||||
self.out_convm = nn.Conv2D( d_mask_ch*1, 1, kernel_size=1, padding='SAME')
|
||||
self.upscalem0 = Upscale(inter_dims, d_mask_dims*8, kernel_size=3)
|
||||
self.upscalem1 = Upscale(d_mask_dims*8, d_mask_dims*8, kernel_size=3)
|
||||
self.upscalem2 = Upscale(d_mask_dims*8, d_mask_dims*4, kernel_size=3)
|
||||
self.upscalem3 = Upscale(d_mask_dims*4, d_mask_dims*2, kernel_size=3)
|
||||
self.upscalem4 = Upscale(d_mask_dims*2, d_mask_dims*1, kernel_size=3)
|
||||
self.out_convm = nn.Conv2D( d_mask_dims*1, 1, kernel_size=1, padding='SAME', dtype=conv_dtype)
|
||||
|
||||
self.out_conv = nn.Conv2D( d_ch*2, 3, kernel_size=1, padding='SAME')
|
||||
self.out_conv1 = nn.Conv2D( d_ch*2, 3, kernel_size=3, padding='SAME')
|
||||
self.out_conv2 = nn.Conv2D( d_ch*2, 3, kernel_size=3, padding='SAME')
|
||||
self.out_conv3 = nn.Conv2D( d_ch*2, 3, kernel_size=3, padding='SAME')
|
||||
self.out_conv = nn.Conv2D( d_dims*2, 3, kernel_size=1, padding='SAME', dtype=conv_dtype)
|
||||
self.out_conv1 = nn.Conv2D( d_dims*2, 3, kernel_size=3, padding='SAME', dtype=conv_dtype)
|
||||
self.out_conv2 = nn.Conv2D( d_dims*2, 3, kernel_size=3, padding='SAME', dtype=conv_dtype)
|
||||
self.out_conv3 = nn.Conv2D( d_dims*2, 3, kernel_size=3, padding='SAME', dtype=conv_dtype)
|
||||
|
||||
def forward(self, inp):
|
||||
z = inp
|
||||
def forward(self, z):
|
||||
if use_fp16:
|
||||
z = tf.cast(z, tf.float16)
|
||||
|
||||
x = self.upscale0(z)
|
||||
x = self.res0(x)
|
||||
|
|
@ -263,54 +240,22 @@ class AMPModel(ModelBase):
|
|||
self.out_conv1(x),
|
||||
self.out_conv2(x),
|
||||
self.out_conv3(x)), nn.conv2d_ch_axis), 2) )
|
||||
|
||||
m = self.upscalem0(z)
|
||||
m = self.upscalem1(m)
|
||||
m = self.upscalem2(m)
|
||||
m = self.upscalem3(m)
|
||||
m = self.upscalem4(m)
|
||||
m = tf.nn.sigmoid(self.out_convm(m))
|
||||
|
||||
if use_fp16:
|
||||
x = tf.cast(x, tf.float32)
|
||||
m = tf.cast(m, tf.float32)
|
||||
return x, m
|
||||
|
||||
self.face_type = {'wf' : FaceType.WHOLE_FACE,
|
||||
'head' : FaceType.HEAD}[ self.options['face_type'] ]
|
||||
|
||||
if 'eyes_prio' in self.options:
|
||||
self.options.pop('eyes_prio')
|
||||
|
||||
eyes_mouth_prio = self.options['eyes_mouth_prio']
|
||||
|
||||
ae_dims = self.ae_dims = self.options['ae_dims']
|
||||
e_dims = self.options['e_dims']
|
||||
d_dims = self.options['d_dims']
|
||||
d_mask_dims = self.options['d_mask_dims']
|
||||
morph_factor = self.options['morph_factor']
|
||||
|
||||
pretrain = self.pretrain = self.options['pretrain']
|
||||
if self.pretrain_just_disabled:
|
||||
self.set_iter(0)
|
||||
|
||||
self.gan_power = gan_power = 0.0 if self.pretrain else self.options['gan_power']
|
||||
random_warp = False if self.pretrain else self.options['random_warp']
|
||||
random_src_flip = self.random_src_flip if not self.pretrain else True
|
||||
random_dst_flip = self.random_dst_flip if not self.pretrain else True
|
||||
|
||||
if self.pretrain:
|
||||
self.options_show_override['gan_power'] = 0.0
|
||||
self.options_show_override['random_warp'] = False
|
||||
self.options_show_override['lr_dropout'] = 'n'
|
||||
self.options_show_override['uniform_yaw'] = True
|
||||
|
||||
masked_training = self.options['masked_training']
|
||||
ct_mode = self.options['ct_mode']
|
||||
if ct_mode == 'none':
|
||||
ct_mode = None
|
||||
|
||||
models_opt_on_gpu = False if len(devices) == 0 else self.options['models_opt_on_gpu']
|
||||
models_opt_device = nn.tf_default_device_name if models_opt_on_gpu and self.is_training else '/CPU:0'
|
||||
optimizer_vars_on_cpu = models_opt_device=='/CPU:0'
|
||||
|
||||
input_ch=3
|
||||
bgr_shape = self.bgr_shape = nn.get4Dshape(resolution,resolution,input_ch)
|
||||
mask_shape = nn.get4Dshape(resolution,resolution,1)
|
||||
self.model_filename_list = []
|
||||
|
|
@ -331,12 +276,11 @@ class AMPModel(ModelBase):
|
|||
self.morph_value_t = tf.placeholder (nn.floatx, (1,), name='morph_value_t')
|
||||
|
||||
# Initializing model classes
|
||||
|
||||
with tf.device (models_opt_device):
|
||||
self.encoder = Encoder(in_ch=input_ch, e_ch=e_dims, ae_ch=ae_dims, name='encoder')
|
||||
self.inter_src = Inter(ae_ch=ae_dims, ae_out_ch=ae_dims, name='inter_src')
|
||||
self.inter_dst = Inter(ae_ch=ae_dims, ae_out_ch=ae_dims, name='inter_dst')
|
||||
self.decoder = Decoder(in_ch=ae_dims, d_ch=d_dims, d_mask_ch=d_mask_dims, name='decoder')
|
||||
self.encoder = Encoder(name='encoder')
|
||||
self.inter_src = Inter(name='inter_src')
|
||||
self.inter_dst = Inter(name='inter_dst')
|
||||
self.decoder = Decoder(name='decoder')
|
||||
|
||||
self.model_filename_list += [ [self.encoder, 'encoder.npy'],
|
||||
[self.inter_src, 'inter_src.npy'],
|
||||
|
|
@ -344,30 +288,21 @@ class AMPModel(ModelBase):
|
|||
[self.decoder , 'decoder.npy'] ]
|
||||
|
||||
if self.is_training:
|
||||
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.model_filename_list += [ [self.GAN, 'GAN.npy'] ]
|
||||
|
||||
# Initialize optimizers
|
||||
lr=5e-5
|
||||
lr_dropout = 0.3 if self.options['lr_dropout'] in ['y','cpu'] and not self.pretrain else 1.0
|
||||
|
||||
clipnorm = 1.0 if self.options['clipgrad'] else 0.0
|
||||
|
||||
self.all_weights = self.encoder.get_weights() + self.inter_src.get_weights() + self.inter_dst.get_weights() + self.decoder.get_weights()
|
||||
if pretrain:
|
||||
self.trainable_weights = self.encoder.get_weights() + self.inter_dst.get_weights() + self.decoder.get_weights()
|
||||
else:
|
||||
self.trainable_weights = self.encoder.get_weights() + self.inter_src.get_weights() + self.inter_dst.get_weights() + self.decoder.get_weights()
|
||||
self.all_weights = self.encoder.get_weights() + self.decoder.get_weights()
|
||||
|
||||
self.src_dst_opt = nn.AdaBelief(lr=lr, lr_dropout=lr_dropout, clipnorm=clipnorm, name='src_dst_opt')
|
||||
self.src_dst_opt.initialize_variables (self.all_weights, vars_on_cpu=optimizer_vars_on_cpu, lr_dropout_on_cpu=self.options['lr_dropout']=='cpu')
|
||||
self.src_dst_opt = nn.AdaBelief(lr=5e-5, lr_dropout=0.3, clipnorm=clipnorm, name='src_dst_opt')
|
||||
self.src_dst_opt.initialize_variables (self.all_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_opt = nn.AdaBelief(lr=lr, lr_dropout=lr_dropout, clipnorm=clipnorm, name='GAN_opt')
|
||||
self.GAN_opt.initialize_variables ( self.GAN.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.GAN_opt, 'GAN_opt.npy') ]
|
||||
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=0.3, 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'] ]
|
||||
|
||||
if self.is_training:
|
||||
# Adjust batch size for multiple GPU
|
||||
|
|
@ -385,10 +320,8 @@ class AMPModel(ModelBase):
|
|||
|
||||
gpu_src_losses = []
|
||||
gpu_dst_losses = []
|
||||
gpu_G_loss_gvs = []
|
||||
gpu_GAN_loss_gvs = []
|
||||
gpu_D_code_loss_gvs = []
|
||||
gpu_D_src_dst_loss_gvs = []
|
||||
gpu_G_loss_gradients = []
|
||||
gpu_GAN_loss_grads = []
|
||||
|
||||
for gpu_id in range(gpu_count):
|
||||
with tf.device( f'/{devices[gpu_id].tf_dev_type}:{gpu_id}' if len(devices) != 0 else f'/CPU:0' ):
|
||||
|
|
@ -408,86 +341,66 @@ class AMPModel(ModelBase):
|
|||
gpu_src_code = self.encoder (gpu_warped_src)
|
||||
gpu_dst_code = self.encoder (gpu_warped_dst)
|
||||
|
||||
if pretrain:
|
||||
gpu_src_inter_src_code = self.inter_src (gpu_src_code)
|
||||
gpu_dst_inter_dst_code = self.inter_dst (gpu_dst_code)
|
||||
gpu_src_code = gpu_src_inter_src_code * nn.random_binomial( [bs_per_gpu, gpu_src_inter_src_code.shape.as_list()[1], 1,1] , p=morph_factor)
|
||||
gpu_dst_code = gpu_src_dst_code = gpu_dst_inter_dst_code * nn.random_binomial( [bs_per_gpu, gpu_dst_inter_dst_code.shape.as_list()[1], 1,1] , p=0.25)
|
||||
else:
|
||||
gpu_src_inter_src_code = self.inter_src (gpu_src_code)
|
||||
gpu_src_inter_dst_code = self.inter_dst (gpu_src_code)
|
||||
gpu_dst_inter_src_code = self.inter_src (gpu_dst_code)
|
||||
gpu_dst_inter_dst_code = self.inter_dst (gpu_dst_code)
|
||||
gpu_src_inter_src_code, gpu_src_inter_dst_code = self.inter_src (gpu_src_code), self.inter_dst (gpu_src_code)
|
||||
gpu_dst_inter_src_code, gpu_dst_inter_dst_code = self.inter_src (gpu_dst_code), self.inter_dst (gpu_dst_code)
|
||||
|
||||
inter_rnd_binomial = nn.random_binomial( [bs_per_gpu, gpu_src_inter_src_code.shape.as_list()[1], 1,1] , p=morph_factor)
|
||||
gpu_src_code = gpu_src_inter_src_code * inter_rnd_binomial + gpu_src_inter_dst_code * (1-inter_rnd_binomial)
|
||||
gpu_dst_code = gpu_dst_inter_dst_code
|
||||
inter_rnd_binomial = nn.random_binomial( [bs_per_gpu, gpu_src_inter_src_code.shape.as_list()[1], 1,1] , p=morph_factor)
|
||||
gpu_src_code = gpu_src_inter_src_code * inter_rnd_binomial + gpu_src_inter_dst_code * (1-inter_rnd_binomial)
|
||||
gpu_dst_code = gpu_dst_inter_dst_code
|
||||
|
||||
ae_dims_slice = tf.cast(ae_dims*self.morph_value_t[0], tf.int32)
|
||||
gpu_src_dst_code = tf.concat( (tf.slice(gpu_dst_inter_src_code, [0,0,0,0], [-1, ae_dims_slice , lowest_dense_res, lowest_dense_res]),
|
||||
tf.slice(gpu_dst_inter_dst_code, [0,ae_dims_slice,0,0], [-1,ae_dims-ae_dims_slice, lowest_dense_res,lowest_dense_res]) ), 1 )
|
||||
inter_dims_slice = tf.cast(inter_dims*self.morph_value_t[0], tf.int32)
|
||||
gpu_src_dst_code = tf.concat( (tf.slice(gpu_dst_inter_src_code, [0,0,0,0], [-1, inter_dims_slice , inter_res, inter_res]),
|
||||
tf.slice(gpu_dst_inter_dst_code, [0,inter_dims_slice,0,0], [-1,inter_dims-inter_dims_slice, inter_res,inter_res]) ), 1 )
|
||||
|
||||
gpu_pred_src_src, gpu_pred_src_srcm = self.decoder(gpu_src_code)
|
||||
gpu_pred_dst_dst, gpu_pred_dst_dstm = self.decoder(gpu_dst_code)
|
||||
gpu_pred_src_dst, gpu_pred_src_dstm = self.decoder(gpu_src_dst_code)
|
||||
|
||||
gpu_pred_src_src_list.append(gpu_pred_src_src)
|
||||
gpu_pred_dst_dst_list.append(gpu_pred_dst_dst)
|
||||
gpu_pred_src_dst_list.append(gpu_pred_src_dst)
|
||||
gpu_pred_src_src_list.append(gpu_pred_src_src), gpu_pred_src_srcm_list.append(gpu_pred_src_srcm)
|
||||
gpu_pred_dst_dst_list.append(gpu_pred_dst_dst), gpu_pred_dst_dstm_list.append(gpu_pred_dst_dstm)
|
||||
gpu_pred_src_dst_list.append(gpu_pred_src_dst), gpu_pred_src_dstm_list.append(gpu_pred_src_dstm)
|
||||
|
||||
gpu_pred_src_srcm_list.append(gpu_pred_src_srcm)
|
||||
gpu_pred_dst_dstm_list.append(gpu_pred_dst_dstm)
|
||||
gpu_pred_src_dstm_list.append(gpu_pred_src_dstm)
|
||||
gpu_target_srcm_blur = tf.clip_by_value( nn.gaussian_blur(gpu_target_srcm, max(1, resolution // 32) ), 0, 0.5) * 2
|
||||
gpu_target_dstm_blur = tf.clip_by_value(nn.gaussian_blur(gpu_target_dstm, max(1, resolution // 32) ), 0, 0.5) * 2
|
||||
|
||||
gpu_target_srcm_blur = nn.gaussian_blur(gpu_target_srcm, max(1, resolution // 32) )
|
||||
gpu_target_srcm_blur = tf.clip_by_value(gpu_target_srcm_blur, 0, 0.5) * 2
|
||||
gpu_target_srcm_anti_blur = 1.0-gpu_target_srcm_blur
|
||||
gpu_target_dstm_anti_blur = 1.0-gpu_target_dstm_blur
|
||||
|
||||
gpu_target_dstm_blur = nn.gaussian_blur(gpu_target_dstm, max(1, resolution // 32) )
|
||||
gpu_target_dstm_blur = tf.clip_by_value(gpu_target_dstm_blur, 0, 0.5) * 2
|
||||
gpu_target_src_masked = gpu_target_src*gpu_target_srcm_blur
|
||||
gpu_target_dst_masked = gpu_target_dst*gpu_target_dstm_blur
|
||||
gpu_target_src_anti_masked = gpu_target_src*gpu_target_srcm_anti_blur
|
||||
gpu_target_dst_anti_masked = gpu_target_dst*gpu_target_dstm_anti_blur
|
||||
|
||||
gpu_target_dst_anti_masked = gpu_target_dst*(1.0-gpu_target_dstm_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*gpu_target_dstm_blur if masked_training else gpu_target_dst
|
||||
gpu_pred_src_src_masked = gpu_pred_src_src*gpu_target_srcm_blur
|
||||
gpu_pred_dst_dst_masked = gpu_pred_dst_dst*gpu_target_dstm_blur
|
||||
gpu_pred_src_src_anti_masked = gpu_pred_src_src*gpu_target_srcm_anti_blur
|
||||
gpu_pred_dst_dst_anti_masked = gpu_pred_dst_dst*gpu_target_dstm_anti_blur
|
||||
|
||||
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_pred_dst_dst_anti_masked = gpu_pred_dst_dst*(1.0-gpu_target_dstm_blur)
|
||||
# Structural loss
|
||||
gpu_src_loss = tf.reduce_mean (5*nn.dssim(gpu_target_src_masked, gpu_pred_src_src_masked, max_val=1.0, filter_size=int(resolution/11.6)), axis=[1])
|
||||
gpu_src_loss += tf.reduce_mean (5*nn.dssim(gpu_target_src_masked, gpu_pred_src_src_masked, max_val=1.0, filter_size=int(resolution/23.2)), axis=[1])
|
||||
gpu_dst_loss = tf.reduce_mean (5*nn.dssim(gpu_target_dst_masked, gpu_pred_dst_dst_masked, max_val=1.0, filter_size=int(resolution/11.6) ), axis=[1])
|
||||
gpu_dst_loss += tf.reduce_mean (5*nn.dssim(gpu_target_dst_masked, gpu_pred_dst_dst_masked, max_val=1.0, filter_size=int(resolution/23.2) ), axis=[1])
|
||||
|
||||
if resolution < 256:
|
||||
gpu_dst_loss = tf.reduce_mean ( 10*nn.dssim(gpu_target_dst_masked_opt, gpu_pred_dst_dst_masked_opt, max_val=1.0, filter_size=int(resolution/11.6) ), axis=[1])
|
||||
else:
|
||||
gpu_dst_loss = tf.reduce_mean ( 5*nn.dssim(gpu_target_dst_masked_opt, gpu_pred_dst_dst_masked_opt, max_val=1.0, filter_size=int(resolution/11.6) ), axis=[1])
|
||||
gpu_dst_loss += tf.reduce_mean ( 5*nn.dssim(gpu_target_dst_masked_opt, gpu_pred_dst_dst_masked_opt, max_val=1.0, filter_size=int(resolution/23.2) ), axis=[1])
|
||||
gpu_dst_loss += tf.reduce_mean ( 10*tf.square( gpu_target_dst_masked_opt- gpu_pred_dst_dst_masked_opt ), axis=[1,2,3])
|
||||
if eyes_mouth_prio:
|
||||
gpu_dst_loss += tf.reduce_mean ( 300*tf.abs ( gpu_target_dst*gpu_target_dstm_em - gpu_pred_dst_dst*gpu_target_dstm_em ), axis=[1,2,3])
|
||||
# Pixel loss
|
||||
gpu_src_loss += tf.reduce_mean (10*tf.square(gpu_target_src_masked-gpu_pred_src_src_masked), axis=[1,2,3])
|
||||
gpu_dst_loss += tf.reduce_mean (10*tf.square(gpu_target_dst_masked-gpu_pred_dst_dst_masked), axis=[1,2,3])
|
||||
|
||||
# Eyes+mouth prio loss
|
||||
gpu_src_loss += tf.reduce_mean (300*tf.abs (gpu_target_src*gpu_target_srcm_em-gpu_pred_src_src*gpu_target_srcm_em), axis=[1,2,3])
|
||||
gpu_dst_loss += tf.reduce_mean (300*tf.abs (gpu_target_dst*gpu_target_dstm_em-gpu_pred_dst_dst*gpu_target_dstm_em), axis=[1,2,3])
|
||||
|
||||
# Mask loss
|
||||
gpu_src_loss += tf.reduce_mean ( 10*tf.square( gpu_target_srcm - gpu_pred_src_srcm ),axis=[1,2,3] )
|
||||
gpu_dst_loss += tf.reduce_mean ( 10*tf.square( gpu_target_dstm - gpu_pred_dst_dstm ),axis=[1,2,3] )
|
||||
gpu_dst_loss += 0.1*tf.reduce_mean(tf.square(gpu_pred_dst_dst_anti_masked-gpu_target_dst_anti_masked),axis=[1,2,3] )
|
||||
gpu_dst_losses += [gpu_dst_loss]
|
||||
|
||||
if not pretrain:
|
||||
if resolution < 256:
|
||||
gpu_src_loss = tf.reduce_mean ( 10*nn.dssim(gpu_target_src_masked_opt, gpu_pred_src_src_masked_opt, max_val=1.0, filter_size=int(resolution/11.6)), axis=[1])
|
||||
else:
|
||||
gpu_src_loss = tf.reduce_mean ( 5*nn.dssim(gpu_target_src_masked_opt, gpu_pred_src_src_masked_opt, max_val=1.0, filter_size=int(resolution/11.6)), axis=[1])
|
||||
gpu_src_loss += tf.reduce_mean ( 5*nn.dssim(gpu_target_src_masked_opt, gpu_pred_src_src_masked_opt, max_val=1.0, filter_size=int(resolution/23.2)), axis=[1])
|
||||
gpu_src_loss += tf.reduce_mean ( 10*tf.square ( gpu_target_src_masked_opt - gpu_pred_src_src_masked_opt ), axis=[1,2,3])
|
||||
|
||||
if eyes_mouth_prio:
|
||||
gpu_src_loss += tf.reduce_mean ( 300*tf.abs ( gpu_target_src*gpu_target_srcm_em - gpu_pred_src_src*gpu_target_srcm_em ), axis=[1,2,3])
|
||||
|
||||
gpu_src_loss += tf.reduce_mean ( 10*tf.square( gpu_target_srcm - gpu_pred_src_srcm ),axis=[1,2,3] )
|
||||
else:
|
||||
gpu_src_loss = gpu_dst_loss
|
||||
# dst-dst background weak loss
|
||||
gpu_dst_loss += tf.reduce_mean(0.1*tf.square(gpu_pred_dst_dst_anti_masked-gpu_target_dst_anti_masked),axis=[1,2,3] )
|
||||
gpu_dst_loss += 0.000001*nn.total_variation_mse(gpu_pred_dst_dst_anti_masked)
|
||||
|
||||
gpu_src_losses += [gpu_src_loss]
|
||||
|
||||
if pretrain:
|
||||
gpu_G_loss = gpu_dst_loss
|
||||
else:
|
||||
gpu_G_loss = gpu_src_loss + gpu_dst_loss
|
||||
gpu_dst_losses += [gpu_dst_loss]
|
||||
gpu_G_loss = gpu_src_loss + gpu_dst_loss
|
||||
|
||||
def DLossOnes(logits):
|
||||
return tf.reduce_mean( tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.ones_like(logits), logits=logits), axis=[1,2,3])
|
||||
|
|
@ -496,30 +409,28 @@ class AMPModel(ModelBase):
|
|||
return tf.reduce_mean( tf.nn.sigmoid_cross_entropy_with_logits(labels=tf.zeros_like(logits), logits=logits), axis=[1,2,3])
|
||||
|
||||
if gan_power != 0:
|
||||
gpu_pred_src_src_d, gpu_pred_src_src_d2 = self.GAN(gpu_pred_src_src_masked_opt)
|
||||
gpu_pred_dst_dst_d, gpu_pred_dst_dst_d2 = self.GAN(gpu_pred_dst_dst_masked_opt)
|
||||
gpu_target_src_d, gpu_target_src_d2 = self.GAN(gpu_target_src_masked_opt)
|
||||
gpu_target_dst_d, gpu_target_dst_d2 = self.GAN(gpu_target_dst_masked_opt)
|
||||
gpu_pred_src_src_d, gpu_pred_src_src_d2 = self.GAN(gpu_pred_src_src_masked)
|
||||
gpu_pred_dst_dst_d, gpu_pred_dst_dst_d2 = self.GAN(gpu_pred_dst_dst_masked)
|
||||
gpu_target_src_d, gpu_target_src_d2 = self.GAN(gpu_target_src_masked)
|
||||
gpu_target_dst_d, gpu_target_dst_d2 = self.GAN(gpu_target_dst_masked)
|
||||
|
||||
gpu_D_src_dst_loss = (DLossOnes (gpu_target_src_d) + DLossOnes (gpu_target_src_d2) + \
|
||||
DLossZeros(gpu_pred_src_src_d) + DLossZeros(gpu_pred_src_src_d2) + \
|
||||
DLossOnes (gpu_target_dst_d) + DLossOnes (gpu_target_dst_d2) + \
|
||||
DLossZeros(gpu_pred_dst_dst_d) + DLossZeros(gpu_pred_dst_dst_d2)
|
||||
) * ( 1.0 / 8)
|
||||
gpu_GAN_loss = (DLossOnes (gpu_target_src_d) + DLossOnes (gpu_target_src_d2) + \
|
||||
DLossZeros(gpu_pred_src_src_d) + DLossZeros(gpu_pred_src_src_d2) + \
|
||||
DLossOnes (gpu_target_dst_d) + DLossOnes (gpu_target_dst_d2) + \
|
||||
DLossZeros(gpu_pred_dst_dst_d) + DLossZeros(gpu_pred_dst_dst_d2)
|
||||
) * (1.0 / 8)
|
||||
|
||||
gpu_D_src_dst_loss_gvs += [ nn.gradients (gpu_D_src_dst_loss, self.GAN.get_weights() ) ]
|
||||
gpu_GAN_loss_grads += [ nn.gradients (gpu_GAN_loss, self.GAN.get_weights() ) ]
|
||||
|
||||
gpu_G_loss += (DLossOnes(gpu_pred_src_src_d) + DLossOnes(gpu_pred_src_src_d2) + \
|
||||
DLossOnes(gpu_pred_dst_dst_d) + DLossOnes(gpu_pred_dst_dst_d2)
|
||||
) * gan_power
|
||||
|
||||
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.trainable_weights ) ]
|
||||
# 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_gradients += [ nn.gradients ( gpu_G_loss, self.encoder.get_weights() + self.decoder.get_weights() ) ]
|
||||
|
||||
# Average losses and gradients, and create optimizer update ops
|
||||
with tf.device(f'/CPU:0'):
|
||||
|
|
@ -533,17 +444,15 @@ class AMPModel(ModelBase):
|
|||
with tf.device (models_opt_device):
|
||||
src_loss = tf.concat(gpu_src_losses, 0)
|
||||
dst_loss = tf.concat(gpu_dst_losses, 0)
|
||||
src_dst_loss_gv_op = self.src_dst_opt.get_update_op (nn.average_gv_list (gpu_G_loss_gvs))
|
||||
train_op = self.src_dst_opt.get_update_op (nn.average_gv_list (gpu_G_loss_gradients))
|
||||
|
||||
if gan_power != 0:
|
||||
src_D_src_dst_loss_gv_op = self.GAN_opt.get_update_op (nn.average_gv_list(gpu_D_src_dst_loss_gvs) )
|
||||
#GAN_loss_gv_op = self.src_dst_opt.get_update_op (nn.average_gv_list(gpu_GAN_loss_gvs) )
|
||||
|
||||
GAN_train_op = self.GAN_opt.get_update_op (nn.average_gv_list(gpu_GAN_loss_grads) )
|
||||
|
||||
# Initializing training and view functions
|
||||
def src_dst_train(warped_src, target_src, target_srcm, target_srcm_em, \
|
||||
def train(warped_src, target_src, target_srcm, target_srcm_em, \
|
||||
warped_dst, target_dst, target_dstm, target_dstm_em, ):
|
||||
s, d, _ = nn.tf_sess.run ( [ src_loss, dst_loss, src_dst_loss_gv_op],
|
||||
s, d, _ = nn.tf_sess.run ([src_loss, dst_loss, train_op],
|
||||
feed_dict={self.warped_src :warped_src,
|
||||
self.target_src :target_src,
|
||||
self.target_srcm:target_srcm,
|
||||
|
|
@ -554,21 +463,20 @@ class AMPModel(ModelBase):
|
|||
self.target_dstm_em:target_dstm_em,
|
||||
})
|
||||
return s, d
|
||||
self.src_dst_train = src_dst_train
|
||||
self.train = train
|
||||
|
||||
if gan_power != 0:
|
||||
def D_src_dst_train(warped_src, target_src, target_srcm, target_srcm_em, \
|
||||
warped_dst, target_dst, target_dstm, target_dstm_em, ):
|
||||
nn.tf_sess.run ([src_D_src_dst_loss_gv_op], feed_dict={self.warped_src :warped_src,
|
||||
self.target_src :target_src,
|
||||
self.target_srcm:target_srcm,
|
||||
self.target_srcm_em:target_srcm_em,
|
||||
self.warped_dst :warped_dst,
|
||||
self.target_dst :target_dst,
|
||||
self.target_dstm:target_dstm,
|
||||
self.target_dstm_em:target_dstm_em})
|
||||
self.D_src_dst_train = D_src_dst_train
|
||||
|
||||
def GAN_train(warped_src, target_src, target_srcm, target_srcm_em, \
|
||||
warped_dst, target_dst, target_dstm, target_dstm_em, ):
|
||||
nn.tf_sess.run ([GAN_train_op], feed_dict={self.warped_src :warped_src,
|
||||
self.target_src :target_src,
|
||||
self.target_srcm:target_srcm,
|
||||
self.target_srcm_em:target_srcm_em,
|
||||
self.warped_dst :warped_dst,
|
||||
self.target_dst :target_dst,
|
||||
self.target_dstm:target_dstm,
|
||||
self.target_dstm_em:target_dstm_em})
|
||||
self.GAN_train = GAN_train
|
||||
|
||||
def AE_view(warped_src, warped_dst, morph_value):
|
||||
return nn.tf_sess.run ( [pred_src_src, pred_dst_dst, pred_dst_dstm, pred_src_dst, pred_src_dstm],
|
||||
|
|
@ -579,12 +487,12 @@ class AMPModel(ModelBase):
|
|||
#Initializing merge function
|
||||
with tf.device( nn.tf_default_device_name if len(devices) != 0 else f'/CPU:0'):
|
||||
gpu_dst_code = self.encoder (self.warped_dst)
|
||||
gpu_dst_inter_src_code = self.inter_src ( gpu_dst_code)
|
||||
gpu_dst_inter_dst_code = self.inter_dst ( gpu_dst_code)
|
||||
gpu_dst_inter_src_code = self.inter_src (gpu_dst_code)
|
||||
gpu_dst_inter_dst_code = self.inter_dst (gpu_dst_code)
|
||||
|
||||
ae_dims_slice = tf.cast(ae_dims*self.morph_value_t[0], tf.int32)
|
||||
gpu_src_dst_code = tf.concat( ( tf.slice(gpu_dst_inter_src_code, [0,0,0,0], [-1, ae_dims_slice , lowest_dense_res, lowest_dense_res]),
|
||||
tf.slice(gpu_dst_inter_dst_code, [0,ae_dims_slice,0,0], [-1,ae_dims-ae_dims_slice, lowest_dense_res,lowest_dense_res]) ), 1 )
|
||||
inter_dims_slice = tf.cast(inter_dims*self.morph_value_t[0], tf.int32)
|
||||
gpu_src_dst_code = tf.concat( ( tf.slice(gpu_dst_inter_src_code, [0,0,0,0], [-1, inter_dims_slice , inter_res, inter_res]),
|
||||
tf.slice(gpu_dst_inter_dst_code, [0,inter_dims_slice,0,0], [-1,inter_dims-inter_dims_slice, inter_res,inter_res]) ), 1 )
|
||||
|
||||
gpu_pred_src_dst, gpu_pred_src_dstm = self.decoder(gpu_src_dst_code)
|
||||
_, gpu_pred_dst_dstm = self.decoder(gpu_dst_inter_dst_code)
|
||||
|
|
@ -596,31 +504,22 @@ class AMPModel(ModelBase):
|
|||
|
||||
# Loading/initializing all models/optimizers weights
|
||||
for model, filename in io.progress_bar_generator(self.model_filename_list, "Initializing models"):
|
||||
if self.pretrain_just_disabled:
|
||||
do_init = False
|
||||
if model == self.inter_src or model == self.inter_dst:
|
||||
do_init = self.is_first_run()
|
||||
if self.is_training and gan_power != 0 and model == self.GAN:
|
||||
if self.gan_model_changed:
|
||||
do_init = True
|
||||
else:
|
||||
do_init = self.is_first_run()
|
||||
if self.is_training and gan_power != 0 and model == self.GAN:
|
||||
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) )
|
||||
if do_init:
|
||||
model.init_weights()
|
||||
|
||||
|
||||
###############
|
||||
|
||||
# initializing sample generators
|
||||
if self.is_training:
|
||||
training_data_src_path = self.training_data_src_path if not self.pretrain else self.get_pretraining_data_path()
|
||||
training_data_dst_path = self.training_data_dst_path if not self.pretrain else self.get_pretraining_data_path()
|
||||
|
||||
random_ct_samples_path=training_data_dst_path if ct_mode is not None and not self.pretrain else None
|
||||
training_data_src_path = self.training_data_src_path #if not self.pretrain else self.get_pretraining_data_path()
|
||||
training_data_dst_path = self.training_data_dst_path #if not self.pretrain else self.get_pretraining_data_path()
|
||||
|
||||
random_ct_samples_path=training_data_dst_path if ct_mode is not None else None #and not self.pretrain
|
||||
|
||||
cpu_count = min(multiprocessing.cpu_count(), 8)
|
||||
src_generators_count = cpu_count // 2
|
||||
|
|
@ -630,33 +529,34 @@ 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(random_flip=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':self.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':self.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':self.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.EYES_MOUTH, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution},
|
||||
sample_process_options=SampleProcessor.Options(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},
|
||||
{'sample_type': SampleProcessor.SampleType.FACE_MASK, 'warp':False , 'transform':True, 'channel_type' : SampleProcessor.ChannelType.G, 'face_mask_type' : SampleProcessor.FaceMaskType.EYES_MOUTH, 'face_type':face_type, 'data_format':nn.data_format, 'resolution': resolution},
|
||||
],
|
||||
uniform_yaw_distribution=self.options['uniform_yaw'] or self.pretrain,
|
||||
uniform_yaw_distribution=self.options['uniform_yaw'],# or self.pretrain,
|
||||
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(random_flip=random_dst_flip),
|
||||
output_sample_types = [ {'sample_type': SampleProcessor.SampleType.FACE_IMAGE,'warp':random_warp, 'transform':True, 'channel_type' : SampleProcessor.ChannelType.BGR, 'face_type':self.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':self.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':self.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.EYES_MOUTH, 'face_type':self.face_type, 'data_format':nn.data_format, 'resolution': resolution},
|
||||
sample_process_options=SampleProcessor.Options(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},
|
||||
{'sample_type': SampleProcessor.SampleType.FACE_MASK, 'warp':False , 'transform':True, 'channel_type' : SampleProcessor.ChannelType.G, 'face_mask_type' : SampleProcessor.FaceMaskType.EYES_MOUTH, 'face_type':face_type, 'data_format':nn.data_format, 'resolution': resolution},
|
||||
],
|
||||
uniform_yaw_distribution=self.options['uniform_yaw'] or self.pretrain,
|
||||
uniform_yaw_distribution=self.options['uniform_yaw'],# or self.pretrain,
|
||||
generators_count=dst_generators_count )
|
||||
])
|
||||
|
||||
self.last_src_samples_loss = []
|
||||
self.last_dst_samples_loss = []
|
||||
if self.pretrain_just_disabled:
|
||||
self.update_sample_for_preview(force_new=True)
|
||||
|
||||
def export_dfm (self):
|
||||
output_path=self.get_strpath_storage_for_file('model.dfm')
|
||||
|
||||
io.log_info(f'Dumping .dfm to {output_path}')
|
||||
|
||||
def dump_ckpt(self):
|
||||
tf = nn.tf
|
||||
with tf.device (nn.tf_default_device_name):
|
||||
warped_dst = tf.placeholder (nn.floatx, (None, self.resolution, self.resolution, 3), name='in_face')
|
||||
|
|
@ -667,9 +567,9 @@ class AMPModel(ModelBase):
|
|||
gpu_dst_inter_src_code = self.inter_src ( gpu_dst_code)
|
||||
gpu_dst_inter_dst_code = self.inter_dst ( gpu_dst_code)
|
||||
|
||||
ae_dims_slice = tf.cast(self.ae_dims*morph_value[0], tf.int32)
|
||||
gpu_src_dst_code = tf.concat( (tf.slice(gpu_dst_inter_src_code, [0,0,0,0], [-1, ae_dims_slice , self.lowest_dense_res, self.lowest_dense_res]),
|
||||
tf.slice(gpu_dst_inter_dst_code, [0,ae_dims_slice,0,0], [-1,self.ae_dims-ae_dims_slice, self.lowest_dense_res,self.lowest_dense_res]) ), 1 )
|
||||
inter_dims_slice = tf.cast(self.inter_dims*morph_value[0], tf.int32)
|
||||
gpu_src_dst_code = tf.concat( (tf.slice(gpu_dst_inter_src_code, [0,0,0,0], [-1, inter_dims_slice , self.inter_res, self.inter_res]),
|
||||
tf.slice(gpu_dst_inter_dst_code, [0,inter_dims_slice,0,0], [-1,self.inter_dims-inter_dims_slice, self.inter_res,self.inter_res]) ), 1 )
|
||||
|
||||
gpu_pred_src_dst, gpu_pred_src_dstm = self.decoder(gpu_src_dst_code)
|
||||
_, gpu_pred_dst_dstm = self.decoder(gpu_dst_inter_dst_code)
|
||||
|
|
@ -688,9 +588,15 @@ class AMPModel(ModelBase):
|
|||
['out_face_mask','out_celeb_face','out_celeb_face_mask']
|
||||
)
|
||||
|
||||
pb_filepath = self.get_strpath_storage_for_file('.pb')
|
||||
with tf.gfile.GFile(pb_filepath, "wb") as f:
|
||||
f.write(output_graph_def.SerializeToString())
|
||||
import tf2onnx
|
||||
with tf.device("/CPU:0"):
|
||||
model_proto, _ = tf2onnx.convert._convert_common(
|
||||
output_graph_def,
|
||||
name='AMP',
|
||||
input_names=['in_face:0','morph_value:0'],
|
||||
output_names=['out_face_mask:0','out_celeb_face:0','out_celeb_face_mask:0'],
|
||||
opset=13,
|
||||
output_path=output_path)
|
||||
|
||||
#override
|
||||
def get_model_filename_list(self):
|
||||
|
|
@ -713,35 +619,37 @@ class AMPModel(ModelBase):
|
|||
( (warped_src, target_src, target_srcm, target_srcm_em), \
|
||||
(warped_dst, target_dst, target_dstm, target_dstm_em) ) = self.generate_next_samples()
|
||||
|
||||
src_loss, dst_loss = self.src_dst_train (warped_src, target_src, target_srcm, target_srcm_em, warped_dst, target_dst, target_dstm, target_dstm_em)
|
||||
src_loss, dst_loss = self.train (warped_src, target_src, target_srcm, target_srcm_em, warped_dst, target_dst, target_dstm, target_dstm_em)
|
||||
|
||||
for i in range(bs):
|
||||
self.last_src_samples_loss.append ( (target_src[i], target_srcm[i], target_srcm_em[i], src_loss[i] ) )
|
||||
self.last_dst_samples_loss.append ( (target_dst[i], target_dstm[i], target_dstm_em[i], dst_loss[i] ) )
|
||||
self.last_src_samples_loss.append ( (src_loss[i], warped_src[i], target_src[i], target_srcm[i], target_srcm_em[i]) )
|
||||
self.last_dst_samples_loss.append ( (dst_loss[i], warped_dst[i], target_dst[i], target_dstm[i], target_dstm_em[i]) )
|
||||
|
||||
if len(self.last_src_samples_loss) >= bs*16:
|
||||
src_samples_loss = sorted(self.last_src_samples_loss, key=operator.itemgetter(3), reverse=True)
|
||||
dst_samples_loss = sorted(self.last_dst_samples_loss, key=operator.itemgetter(3), reverse=True)
|
||||
src_samples_loss = sorted(self.last_src_samples_loss, key=operator.itemgetter(0), reverse=True)
|
||||
dst_samples_loss = sorted(self.last_dst_samples_loss, key=operator.itemgetter(0), reverse=True)
|
||||
|
||||
target_src = np.stack( [ x[0] for x in src_samples_loss[:bs] ] )
|
||||
target_srcm = np.stack( [ x[1] for x in src_samples_loss[:bs] ] )
|
||||
target_srcm_em = np.stack( [ x[2] for x in src_samples_loss[:bs] ] )
|
||||
warped_src = np.stack( [ x[1] for x in src_samples_loss[:bs] ] )
|
||||
target_src = np.stack( [ x[2] for x in src_samples_loss[:bs] ] )
|
||||
target_srcm = np.stack( [ x[3] for x in src_samples_loss[:bs] ] )
|
||||
target_srcm_em = np.stack( [ x[4] for x in src_samples_loss[:bs] ] )
|
||||
|
||||
target_dst = np.stack( [ x[0] for x in dst_samples_loss[:bs] ] )
|
||||
target_dstm = np.stack( [ x[1] for x in dst_samples_loss[:bs] ] )
|
||||
target_dstm_em = np.stack( [ x[2] for x in dst_samples_loss[:bs] ] )
|
||||
warped_dst = np.stack( [ x[1] for x in dst_samples_loss[:bs] ] )
|
||||
target_dst = np.stack( [ x[2] for x in dst_samples_loss[:bs] ] )
|
||||
target_dstm = np.stack( [ x[3] for x in dst_samples_loss[:bs] ] )
|
||||
target_dstm_em = np.stack( [ x[4] for x in dst_samples_loss[:bs] ] )
|
||||
|
||||
src_loss, dst_loss = self.src_dst_train (target_src, target_src, target_srcm, target_srcm_em, target_dst, target_dst, target_dstm, target_dstm_em)
|
||||
src_loss, dst_loss = self.train (warped_src, target_src, target_srcm, target_srcm_em, warped_dst, target_dst, target_dstm, target_dstm_em)
|
||||
self.last_src_samples_loss = []
|
||||
self.last_dst_samples_loss = []
|
||||
|
||||
if self.gan_power != 0:
|
||||
self.D_src_dst_train (warped_src, target_src, target_srcm, target_srcm_em, warped_dst, target_dst, target_dstm, target_dstm_em)
|
||||
self.GAN_train (warped_src, target_src, target_srcm, target_srcm_em, warped_dst, target_dst, target_dstm, target_dstm_em)
|
||||
|
||||
return ( ('src_loss', np.mean(src_loss) ), ('dst_loss', np.mean(dst_loss) ), )
|
||||
|
||||
#override
|
||||
def onGetPreview(self, samples):
|
||||
def onGetPreview(self, samples, for_history=False):
|
||||
( (warped_src, target_src, target_srcm, target_srcm_em),
|
||||
(warped_dst, target_dst, target_dstm, target_dstm_em) ) = samples
|
||||
|
||||
|
|
@ -771,18 +679,17 @@ class AMPModel(ModelBase):
|
|||
|
||||
result = []
|
||||
|
||||
i = np.random.randint(n_samples)
|
||||
i = np.random.randint(n_samples) if not for_history else 0
|
||||
|
||||
st = [ np.concatenate ((S[i], D[i], DD[i]*DDM_000[i]), axis=1) ]
|
||||
st += [ np.concatenate ((SS[i], DD[i], SD_075[i] ), axis=1) ]
|
||||
st += [ np.concatenate ((SS[i], DD[i], SD_100[i] ), axis=1) ]
|
||||
|
||||
result += [ ('AMP morph 0.75', np.concatenate (st, axis=0 )), ]
|
||||
result += [ ('AMP morph 1.0', np.concatenate (st, axis=0 )), ]
|
||||
|
||||
st = [ np.concatenate ((DD[i], SD_025[i], SD_050[i]), axis=1) ]
|
||||
st += [ np.concatenate ((SD_065[i], SD_075[i], SD_100[i]), axis=1) ]
|
||||
result += [ ('AMP morph list', np.concatenate (st, axis=0 )), ]
|
||||
|
||||
|
||||
st = [ np.concatenate ((DD[i], SD_025[i]*DDM_025[i]*SDM_025[i], SD_050[i]*DDM_050[i]*SDM_050[i]), axis=1) ]
|
||||
st += [ np.concatenate ((SD_065[i]*DDM_065[i]*SDM_065[i], SD_075[i]*DDM_075[i]*SDM_075[i], SD_100[i]*DDM_100[i]*SDM_100[i]), axis=1) ]
|
||||
result += [ ('AMP morph list masked', np.concatenate (st, axis=0 )), ]
|
||||
|
|
@ -798,7 +705,7 @@ class AMPModel(ModelBase):
|
|||
|
||||
#override
|
||||
def get_MergerConfig(self):
|
||||
morph_factor = np.clip ( io.input_number ("Morph factor", 0.75, add_info="0.0 .. 1.0"), 0.0, 1.0 )
|
||||
morph_factor = np.clip ( io.input_number ("Morph factor", 1.0, add_info="0.0 .. 1.0"), 0.0, 1.0 )
|
||||
|
||||
def predictor_morph(face):
|
||||
return self.predictor_func(face, morph_factor)
|
||||
|
|
|
|||
|
|
@ -278,7 +278,7 @@ class QModel(ModelBase):
|
|||
return ( ('src_loss', src_loss), ('dst_loss', dst_loss), )
|
||||
|
||||
#override
|
||||
def onGetPreview(self, samples):
|
||||
def onGetPreview(self, samples, for_history=False):
|
||||
( (warped_src, target_src, target_srcm),
|
||||
(warped_dst, target_dst, target_dstm) ) = samples
|
||||
|
||||
|
|
|
|||
|
|
@ -29,7 +29,8 @@ class SAEHDModel(ModelBase):
|
|||
yn_str = {True:'y',False:'n'}
|
||||
min_res = 64
|
||||
max_res = 640
|
||||
|
||||
|
||||
#default_usefp16 = self.options['use_fp16'] = self.load_or_def_option('use_fp16', False)
|
||||
default_resolution = self.options['resolution'] = self.load_or_def_option('resolution', 128)
|
||||
default_face_type = self.options['face_type'] = self.load_or_def_option('face_type', 'f')
|
||||
default_models_opt_on_gpu = self.options['models_opt_on_gpu'] = self.load_or_def_option('models_opt_on_gpu', True)
|
||||
|
|
@ -68,14 +69,15 @@ class SAEHDModel(ModelBase):
|
|||
self.ask_random_src_flip()
|
||||
self.ask_random_dst_flip()
|
||||
self.ask_batch_size(suggest_batch_size)
|
||||
|
||||
#self.options['use_fp16'] = io.input_bool ("Use fp16", default_usefp16, help_message='Increases training/inference speed, reduces model size. Model may crash. Enable it after 1-5k iters.')
|
||||
|
||||
if self.is_first_run():
|
||||
resolution = io.input_int("Resolution", default_resolution, add_info="64-640", help_message="More resolution requires more VRAM and time to train. Value will be adjusted to multiple of 16 and 32 for -d archi.")
|
||||
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'], help_message="Half / mid face / full face / whole face / head. 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:
|
||||
|
|
@ -136,11 +138,11 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
|
|||
|
||||
self.options['eyes_mouth_prio'] = io.input_bool ("Eyes and mouth priority", default_eyes_mouth_prio, help_message='Helps to fix eye problems during training like "alien eyes" and wrong eyes direction. Also makes the detail of the teeth higher.')
|
||||
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.")
|
||||
|
||||
|
|
@ -150,15 +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 .. 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:
|
||||
self.options['gan_power'] = np.clip ( io.input_number ("GAN power", default_gan_power, add_info="0.0 .. 5.0", 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, 5.0 )
|
||||
|
||||
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 )
|
||||
|
||||
gan_dims = np.clip ( io.input_int("GAN dimensions", default_gan_dims, add_info="4-512", 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, 512 )
|
||||
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:
|
||||
|
|
@ -174,7 +176,7 @@ 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)
|
||||
|
|
@ -196,7 +198,7 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
|
|||
|
||||
if 'eyes_prio' in self.options:
|
||||
self.options.pop('eyes_prio')
|
||||
|
||||
|
||||
eyes_mouth_prio = self.options['eyes_mouth_prio']
|
||||
|
||||
archi_split = self.options['archi'].split('-')
|
||||
|
|
@ -205,7 +207,7 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
|
|||
archi_type, archi_opts = archi_split
|
||||
elif len(archi_split) == 1:
|
||||
archi_type, archi_opts = archi_split[0], None
|
||||
|
||||
|
||||
self.archi_type = archi_type
|
||||
|
||||
ae_dims = self.options['ae_dims']
|
||||
|
|
@ -217,12 +219,13 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
|
|||
self.set_iter(0)
|
||||
|
||||
adabelief = self.options['adabelief']
|
||||
|
||||
use_fp16 = False#self.options['use_fp16']
|
||||
|
||||
self.gan_power = gan_power = 0.0 if self.pretrain else self.options['gan_power']
|
||||
random_warp = False if self.pretrain else self.options['random_warp']
|
||||
random_src_flip = self.random_src_flip if not self.pretrain else True
|
||||
random_dst_flip = self.random_dst_flip if not self.pretrain else True
|
||||
|
||||
|
||||
if self.pretrain:
|
||||
self.options_show_override['gan_power'] = 0.0
|
||||
self.options_show_override['random_warp'] = False
|
||||
|
|
@ -235,8 +238,8 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
|
|||
ct_mode = self.options['ct_mode']
|
||||
if ct_mode == 'none':
|
||||
ct_mode = None
|
||||
|
||||
|
||||
|
||||
|
||||
models_opt_on_gpu = False if len(devices) == 0 else self.options['models_opt_on_gpu']
|
||||
models_opt_device = nn.tf_default_device_name if models_opt_on_gpu and self.is_training else '/CPU:0'
|
||||
optimizer_vars_on_cpu = models_opt_device=='/CPU:0'
|
||||
|
|
@ -260,7 +263,7 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
|
|||
self.target_dstm_em = tf.placeholder (nn.floatx, mask_shape, name='target_dstm_em')
|
||||
|
||||
# Initializing model classes
|
||||
model_archi = nn.DeepFakeArchi(resolution, opts=archi_opts)
|
||||
model_archi = nn.DeepFakeArchi(resolution, use_fp16=use_fp16, opts=archi_opts)
|
||||
|
||||
with tf.device (models_opt_device):
|
||||
if 'df' in archi_type:
|
||||
|
|
@ -350,7 +353,7 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
|
|||
gpu_G_loss_gvs = []
|
||||
gpu_D_code_loss_gvs = []
|
||||
gpu_D_src_dst_loss_gvs = []
|
||||
|
||||
|
||||
for gpu_id in range(gpu_count):
|
||||
with tf.device( f'/{devices[gpu_id].tf_dev_type}:{gpu_id}' if len(devices) != 0 else f'/CPU:0' ):
|
||||
with tf.device(f'/CPU:0'):
|
||||
|
|
@ -402,7 +405,7 @@ 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)
|
||||
|
||||
|
|
@ -467,7 +470,7 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
|
|||
|
||||
gpu_G_loss += self.options['true_face_power']*DLoss(gpu_src_code_d_ones, gpu_src_code_d)
|
||||
|
||||
gpu_D_code_loss = (DLoss(gpu_src_code_d_ones , gpu_dst_code_d) + \
|
||||
gpu_D_code_loss = (DLoss(gpu_dst_code_d_ones , gpu_dst_code_d) + \
|
||||
DLoss(gpu_src_code_d_zeros, gpu_src_code_d) ) * 0.5
|
||||
|
||||
gpu_D_code_loss_gvs += [ nn.gradients (gpu_D_code_loss, self.code_discriminator.get_weights() ) ]
|
||||
|
|
@ -497,14 +500,14 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
|
|||
|
||||
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 ) ]
|
||||
|
||||
|
||||
|
|
@ -614,10 +617,10 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
|
|||
|
||||
if do_init:
|
||||
model.init_weights()
|
||||
|
||||
|
||||
|
||||
|
||||
###############
|
||||
|
||||
|
||||
# initializing sample generators
|
||||
if self.is_training:
|
||||
training_data_src_path = self.training_data_src_path if not self.pretrain else self.get_pretraining_data_path()
|
||||
|
|
@ -658,16 +661,20 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
|
|||
|
||||
if self.pretrain_just_disabled:
|
||||
self.update_sample_for_preview(force_new=True)
|
||||
|
||||
def dump_ckpt(self):
|
||||
|
||||
def export_dfm (self):
|
||||
output_path=self.get_strpath_storage_for_file('model.dfm')
|
||||
|
||||
io.log_info(f'Dumping .dfm to {output_path}')
|
||||
|
||||
tf = nn.tf
|
||||
|
||||
|
||||
with tf.device ('/CPU:0'):
|
||||
nn.set_data_format('NCHW')
|
||||
|
||||
with tf.device (nn.tf_default_device_name):
|
||||
warped_dst = tf.placeholder (nn.floatx, (None, self.resolution, self.resolution, 3), name='in_face')
|
||||
warped_dst = tf.transpose(warped_dst, (0,3,1,2))
|
||||
|
||||
|
||||
|
||||
|
||||
if 'df' in self.archi_type:
|
||||
gpu_dst_code = self.inter(self.encoder(warped_dst))
|
||||
gpu_pred_src_dst, gpu_pred_src_dstm = self.decoder_src(gpu_dst_code)
|
||||
|
|
@ -682,20 +689,31 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
|
|||
|
||||
gpu_pred_src_dst, gpu_pred_src_dstm = self.decoder(gpu_src_dst_code)
|
||||
_, gpu_pred_dst_dstm = self.decoder(gpu_dst_code)
|
||||
|
||||
|
||||
gpu_pred_src_dst = tf.transpose(gpu_pred_src_dst, (0,2,3,1))
|
||||
gpu_pred_dst_dstm = tf.transpose(gpu_pred_dst_dstm, (0,2,3,1))
|
||||
gpu_pred_src_dstm = tf.transpose(gpu_pred_src_dstm, (0,2,3,1))
|
||||
|
||||
|
||||
saver = tf.train.Saver()
|
||||
tf.identity(gpu_pred_dst_dstm, name='out_face_mask')
|
||||
tf.identity(gpu_pred_src_dst, name='out_celeb_face')
|
||||
tf.identity(gpu_pred_src_dstm, name='out_celeb_face_mask')
|
||||
|
||||
saver.save(nn.tf_sess, self.get_strpath_storage_for_file('.ckpt') )
|
||||
|
||||
|
||||
tf.identity(gpu_pred_src_dstm, name='out_celeb_face_mask')
|
||||
|
||||
output_graph_def = tf.graph_util.convert_variables_to_constants(
|
||||
nn.tf_sess,
|
||||
tf.get_default_graph().as_graph_def(),
|
||||
['out_face_mask','out_celeb_face','out_celeb_face_mask']
|
||||
)
|
||||
|
||||
import tf2onnx
|
||||
with tf.device("/CPU:0"):
|
||||
model_proto, _ = tf2onnx.convert._convert_common(
|
||||
output_graph_def,
|
||||
name='SAEHD',
|
||||
input_names=['in_face:0'],
|
||||
output_names=['out_face_mask:0','out_celeb_face:0','out_celeb_face_mask:0'],
|
||||
opset=13,
|
||||
output_path=output_path)
|
||||
|
||||
#override
|
||||
def get_model_filename_list(self):
|
||||
return self.model_filename_list
|
||||
|
|
@ -751,7 +769,7 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
|
|||
return ( ('src_loss', np.mean(src_loss) ), ('dst_loss', np.mean(dst_loss) ), )
|
||||
|
||||
#override
|
||||
def onGetPreview(self, samples):
|
||||
def onGetPreview(self, samples, for_history=False):
|
||||
( (warped_src, target_src, target_srcm, target_srcm_em),
|
||||
(warped_dst, target_dst, target_dstm, target_dstm_em) ) = samples
|
||||
|
||||
|
|
|
|||
|
|
@ -164,7 +164,7 @@ class XSegModel(ModelBase):
|
|||
return ( ('loss', np.mean(loss) ), )
|
||||
|
||||
#override
|
||||
def onGetPreview(self, samples):
|
||||
def onGetPreview(self, samples, for_history=False):
|
||||
n_samples = min(4, self.get_batch_size(), 800 // self.resolution )
|
||||
|
||||
srcdst_samples, src_samples, dst_samples = samples
|
||||
|
|
|
|||
|
|
@ -1,9 +1,10 @@
|
|||
tqdm
|
||||
numpy==1.19.3
|
||||
h5py==2.9.0
|
||||
h5py==2.10.0
|
||||
opencv-python==4.1.0.25
|
||||
ffmpeg-python==0.1.17
|
||||
scikit-image==0.14.2
|
||||
scipy==1.4.1
|
||||
colorama
|
||||
tensorflow-gpu==2.3.1
|
||||
tensorflow-gpu==2.4.0
|
||||
tf2onnx==1.8.4
|
||||
|
|
@ -8,3 +8,4 @@ scipy==1.4.1
|
|||
colorama
|
||||
tensorflow-gpu==2.4.0
|
||||
pyqt5
|
||||
tf2onnx==1.8.4
|
||||
Loading…
Add table
Add a link
Reference in a new issue