mirror of
https://github.com/iperov/DeepFaceLab.git
synced 2025-07-05 20:42:11 -07:00
AMP, SAEHD: added option use_fp16: Increases training/inference speed, reduces model size. Model may crash. Enable it after 1-5k iters.
AMP: removed lr_dropout, now it is enabled by default;
This commit is contained in:
iperov
parent
ee1bc83a14
commit
35877dbfd7
4 changed files with 102 additions and 98 deletions
|
|
@ -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)
|
||||
|
|
@ -80,8 +83,13 @@ class DeepFakeArchi(nn.ArchiBase):
|
|||
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)
|
||||
|
|
@ -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))
|
||||
|
||||
|
|
@ -201,7 +182,11 @@ 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
|
||||
|
|
|
|||
|
|
@ -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)
|
||||
|
|
@ -151,20 +153,23 @@ class UNetPatchDiscriminator(nn.ModelBase):
|
|||
|
||||
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.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.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.out_conv = nn.Conv2D( level_chs[-1]*2, 1, kernel_size=1, padding='VALID')
|
||||
self.out_conv = nn.Conv2D( level_chs[-1]*2, 1, kernel_size=1, padding='VALID', dtype=conv_dtype)
|
||||
|
||||
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 = []
|
||||
|
|
@ -178,6 +183,12 @@ class UNetPatchDiscriminator(nn.ModelBase):
|
|||
x = tf.nn.leaky_relu( upconv(x), 0.2 )
|
||||
x = tf.concat( [enc, x], axis=nn.conv2d_ch_axis)
|
||||
|
||||
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
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue