mirror of
https://github.com/iperov/DeepFaceLab.git
synced 2024-03-22 13:10:55 +08:00
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
|
|
@ -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
|
||||
|
|
|
|||
|
|
@ -30,7 +30,8 @@ class AMPModel(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', 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)
|
||||
|
|
@ -50,10 +51,6 @@ class AMPModel(ModelBase):
|
|||
default_eyes_mouth_prio = self.options['eyes_mouth_prio'] = self.load_or_def_option('eyes_mouth_prio', True)
|
||||
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)
|
||||
|
|
@ -68,7 +65,8 @@ class AMPModel(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 32 .")
|
||||
resolution = np.clip ( (resolution // 32) * 32, min_res, max_res)
|
||||
|
|
@ -113,11 +111,9 @@ 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 .. 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 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:
|
||||
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 )
|
||||
|
|
@ -151,17 +147,19 @@ class AMPModel(ModelBase):
|
|||
d_dims = self.options['d_dims']
|
||||
d_mask_dims = self.options['d_mask_dims']
|
||||
inter_res = self.inter_res = resolution // 32
|
||||
|
||||
use_fp16 = self.options['use_fp16']
|
||||
conv_dtype = tf.float16 if use_fp16 else tf.float32
|
||||
|
||||
class Downscale(nn.ModelBase):
|
||||
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')
|
||||
self.conv1 = nn.Conv2D( in_ch, out_ch, kernel_size=kernel_size, strides=2, padding='SAME', dtype=conv_dtype)
|
||||
|
||||
def forward(self, x):
|
||||
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 = nn.depth_to_space(tf.nn.leaky_relu(self.conv1(x), 0.1), 2)
|
||||
|
|
@ -169,8 +167,8 @@ class AMPModel(ModelBase):
|
|||
|
||||
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)
|
||||
|
|
@ -190,8 +188,9 @@ class AMPModel(ModelBase):
|
|||
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)
|
||||
|
|
@ -199,6 +198,8 @@ class AMPModel(ModelBase):
|
|||
x = self.down4(x)
|
||||
x = self.down5(x)
|
||||
x = self.res5(x)
|
||||
if use_fp16:
|
||||
x = tf.cast(x, tf.float32)
|
||||
x = nn.pixel_norm(nn.flatten(x), axes=-1)
|
||||
x = self.dense1(x)
|
||||
return x
|
||||
|
|
@ -232,15 +233,16 @@ class AMPModel(ModelBase):
|
|||
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')
|
||||
self.out_convm = nn.Conv2D( d_mask_dims*1, 1, kernel_size=1, padding='SAME', dtype=conv_dtype)
|
||||
|
||||
self.out_conv = nn.Conv2D( d_dims*2, 3, kernel_size=1, padding='SAME')
|
||||
self.out_conv1 = nn.Conv2D( d_dims*2, 3, kernel_size=3, padding='SAME')
|
||||
self.out_conv2 = nn.Conv2D( d_dims*2, 3, kernel_size=3, padding='SAME')
|
||||
self.out_conv3 = nn.Conv2D( d_dims*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)
|
||||
|
|
@ -262,6 +264,10 @@ class AMPModel(ModelBase):
|
|||
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 = {'f' : FaceType.FULL,
|
||||
|
|
@ -339,12 +345,12 @@ class AMPModel(ModelBase):
|
|||
|
||||
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.GAN = nn.UNetPatchDiscriminator(patch_size=self.options['gan_patch_size'], in_ch=input_ch, base_ch=self.options['gan_dims'], use_fp16=use_fp16, 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'] else 1.0 #and not self.pretrain
|
||||
lr_dropout = 0.3
|
||||
|
||||
clipnorm = 1.0 if self.options['clipgrad'] else 0.0
|
||||
|
||||
|
|
@ -355,12 +361,12 @@ class AMPModel(ModelBase):
|
|||
self.trainable_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.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.GAN_opt.initialize_variables ( self.GAN.get_weights(), vars_on_cpu=optimizer_vars_on_cpu)#+self.D_src_x2.get_weights()
|
||||
self.model_filename_list += [ (self.GAN_opt, 'GAN_opt.npy') ]
|
||||
|
||||
if self.is_training:
|
||||
|
|
|
|||
|
|
@ -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)
|
||||
|
|
@ -68,7 +69,8 @@ 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)
|
||||
|
|
@ -260,7 +262,7 @@ 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=self.options['use_fp16'], opts=archi_opts)
|
||||
|
||||
with tf.device (models_opt_device):
|
||||
if 'df' in archi_type:
|
||||
|
|
@ -301,7 +303,7 @@ Examples: df, liae, df-d, df-ud, liae-ud, ...
|
|||
|
||||
if self.is_training:
|
||||
if gan_power != 0:
|
||||
self.D_src = nn.UNetPatchDiscriminator(patch_size=self.options['gan_patch_size'], in_ch=input_ch, base_ch=self.options['gan_dims'], name="D_src")
|
||||
self.D_src = nn.UNetPatchDiscriminator(patch_size=self.options['gan_patch_size'], in_ch=input_ch, base_ch=self.options['gan_dims'], use_fp16=self.options['use_fp16'], name="D_src")
|
||||
self.model_filename_list += [ [self.D_src, 'GAN.npy'] ]
|
||||
|
||||
# Initialize optimizers
|
||||
|
|
|
|||
Loading…
Reference in New Issue
Block a user