mirror of
https://github.com/iperov/DeepFaceLab.git
synced 2025-07-06 13:02:15 -07:00
76ca79216e
Removed the wait at first launch for most graphics cards. Increased speed of training by 10-20%, but you have to retrain all models from scratch. SAEHD: added option 'use float16' Experimental option. Reduces the model size by half. Increases the speed of training. Decreases the accuracy of the model. The model may collapse or not train. Model may not learn the mask in large resolutions. true_face_training option is replaced by "True face power". 0.0000 .. 1.0 Experimental option. Discriminates the result face to be more like the src face. Higher value - stronger discrimination. Comparison - https://i.imgur.com/czScS9q.png
332 lines
No EOL
13 KiB
Python
332 lines
No EOL
13 KiB
Python
import numpy as np
|
|
|
|
def initialize_tensor_ops(nn):
|
|
tf = nn.tf
|
|
from tensorflow.python.ops import array_ops, random_ops, math_ops, sparse_ops, gradients
|
|
from tensorflow.python.framework import sparse_tensor
|
|
|
|
def tf_get_value(tensor):
|
|
return nn.tf_sess.run (tensor)
|
|
nn.tf_get_value = tf_get_value
|
|
|
|
|
|
def tf_batch_set_value(tuples):
|
|
if len(tuples) != 0:
|
|
with nn.tf.device('/CPU:0'):
|
|
assign_ops = []
|
|
feed_dict = {}
|
|
|
|
for x, value in tuples:
|
|
if isinstance(value, nn.tf.Operation):
|
|
assign_ops.append(value)
|
|
else:
|
|
value = np.asarray(value, dtype=x.dtype.as_numpy_dtype)
|
|
assign_placeholder = nn.tf.placeholder( x.dtype.base_dtype, shape=[None]*value.ndim )
|
|
assign_op = nn.tf.assign (x, assign_placeholder )
|
|
assign_ops.append(assign_op)
|
|
feed_dict[assign_placeholder] = value
|
|
|
|
nn.tf_sess.run(assign_ops, feed_dict=feed_dict)
|
|
nn.tf_batch_set_value = tf_batch_set_value
|
|
|
|
|
|
def tf_gradients ( loss, vars ):
|
|
grads = gradients.gradients(loss, vars, colocate_gradients_with_ops=True )
|
|
gv = [*zip(grads,vars)]
|
|
for g,v in gv:
|
|
if g is None:
|
|
raise Exception("No gradient for variable {v.name}")
|
|
return gv
|
|
nn.tf_gradients = tf_gradients
|
|
|
|
def tf_average_gv_list(grad_var_list, tf_device_string=None):
|
|
if len(grad_var_list) == 1:
|
|
return grad_var_list[0]
|
|
|
|
e = tf.device(tf_device_string) if tf_device_string is not None else None
|
|
if e is not None: e.__enter__()
|
|
result = []
|
|
for i, (gv) in enumerate(grad_var_list):
|
|
for j,(g,v) in enumerate(gv):
|
|
g = tf.expand_dims(g, 0)
|
|
if i == 0:
|
|
result += [ [[g], v] ]
|
|
else:
|
|
result[j][0] += [g]
|
|
|
|
for i,(gs,v) in enumerate(result):
|
|
result[i] = ( tf.reduce_mean( tf.concat (gs, 0), 0 ), v )
|
|
if e is not None: e.__exit__(None,None,None)
|
|
return result
|
|
nn.tf_average_gv_list = tf_average_gv_list
|
|
|
|
def tf_average_tensor_list(tensors_list, tf_device_string=None):
|
|
if len(tensors_list) == 1:
|
|
return tensors_list[0]
|
|
|
|
e = tf.device(tf_device_string) if tf_device_string is not None else None
|
|
if e is not None: e.__enter__()
|
|
result = tf.reduce_mean(tf.concat ([tf.expand_dims(t, 0) for t in tensors_list], 0), 0)
|
|
if e is not None: e.__exit__(None,None,None)
|
|
return result
|
|
nn.tf_average_tensor_list = tf_average_tensor_list
|
|
|
|
def tf_concat (tensors_list, axis):
|
|
"""
|
|
Better version.
|
|
"""
|
|
if len(tensors_list) == 1:
|
|
return tensors_list[0]
|
|
return tf.concat(tensors_list, axis)
|
|
nn.tf_concat = tf_concat
|
|
|
|
def tf_gelu(x):
|
|
cdf = 0.5 * (1.0 + tf.nn.tanh((np.sqrt(2 / np.pi) * (x + 0.044715 * tf.pow(x, 3)))))
|
|
return x * cdf
|
|
nn.tf_gelu = tf_gelu
|
|
|
|
def tf_upsample2d(x, size=2):
|
|
if nn.data_format == "NCHW":
|
|
b,c,h,w = x.shape.as_list()
|
|
x = tf.reshape (x, (-1,c,h,1,w,1) )
|
|
x = tf.tile(x, (1,1,1,size,1,size) )
|
|
x = tf.reshape (x, (-1,c,h*size,w*size) )
|
|
return x
|
|
else:
|
|
return tf.image.resize_nearest_neighbor(x, (x.shape[1]*size, x.shape[2]*size) )
|
|
nn.tf_upsample2d = tf_upsample2d
|
|
|
|
def tf_upsample2d_bilinear(x, size=2):
|
|
return tf.image.resize_images(x, (x.shape[1]*size, x.shape[2]*size) )
|
|
nn.tf_upsample2d_bilinear = tf_upsample2d_bilinear
|
|
|
|
def tf_flatten(x):
|
|
if nn.data_format == "NHWC":
|
|
# match NCHW version in order to switch data_format without problems
|
|
x = tf.transpose(x, (0,3,1,2) )
|
|
return tf.reshape (x, (-1, np.prod(x.shape[1:])) )
|
|
|
|
nn.tf_flatten = tf_flatten
|
|
|
|
def tf_reshape_4D(x, w,h,c):
|
|
if nn.data_format == "NHWC":
|
|
# match NCHW version in order to switch data_format without problems
|
|
x = tf.reshape (x, (-1,c,h,w))
|
|
x = tf.transpose(x, (0,2,3,1) )
|
|
return x
|
|
else:
|
|
return tf.reshape (x, (-1,c,h,w))
|
|
nn.tf_reshape_4D = tf_reshape_4D
|
|
|
|
def tf_random_binomial(shape, p=0.0, dtype=None, seed=None):
|
|
if dtype is None:
|
|
dtype=tf.float32
|
|
|
|
if seed is None:
|
|
seed = np.random.randint(10e6)
|
|
return array_ops.where(
|
|
random_ops.random_uniform(shape, dtype=tf.float16, seed=seed) < p,
|
|
array_ops.ones(shape, dtype=dtype), array_ops.zeros(shape, dtype=dtype))
|
|
nn.tf_random_binomial = tf_random_binomial
|
|
|
|
def tf_gaussian_blur(input, radius=2.0):
|
|
def gaussian(x, mu, sigma):
|
|
return np.exp(-(float(x) - float(mu)) ** 2 / (2 * sigma ** 2))
|
|
|
|
def make_kernel(sigma):
|
|
kernel_size = max(3, int(2 * 2 * sigma + 1))
|
|
mean = np.floor(0.5 * kernel_size)
|
|
kernel_1d = np.array([gaussian(x, mean, sigma) for x in range(kernel_size)])
|
|
np_kernel = np.outer(kernel_1d, kernel_1d).astype(np.float32)
|
|
kernel = np_kernel / np.sum(np_kernel)
|
|
return kernel, kernel_size
|
|
|
|
gauss_kernel, kernel_size = make_kernel(radius)
|
|
padding = kernel_size//2
|
|
if padding != 0:
|
|
if nn.data_format == "NHWC":
|
|
padding = [ [0,0], [padding,padding], [padding,padding], [0,0] ]
|
|
else:
|
|
padding = [ [0,0], [0,0], [padding,padding], [padding,padding] ]
|
|
else:
|
|
padding = None
|
|
gauss_kernel = gauss_kernel[:,:,None,None]
|
|
|
|
outputs = []
|
|
for i in range(input.shape[nn.conv2d_ch_axis]):
|
|
x = input[:,:,:,i:i+1] if nn.data_format == "NHWC" \
|
|
else input[:,i:i+1,:,:]
|
|
|
|
if padding is not None:
|
|
x = tf.pad (x, padding)
|
|
outputs += [ tf.nn.conv2d(x, tf.constant(gauss_kernel, dtype=input.dtype ), strides=[1,1,1,1], padding="VALID", data_format=nn.data_format) ]
|
|
|
|
return tf.concat (outputs, axis=nn.conv2d_ch_axis)
|
|
nn.tf_gaussian_blur = tf_gaussian_blur
|
|
|
|
def tf_style_loss(target, style, gaussian_blur_radius=0.0, loss_weight=1.0, step_size=1):
|
|
def sd(content, style, loss_weight):
|
|
content_nc = content.shape[ nn.conv2d_ch_axis ]
|
|
style_nc = style.shape[nn.conv2d_ch_axis]
|
|
if content_nc != style_nc:
|
|
raise Exception("style_loss() content_nc != style_nc")
|
|
c_mean, c_var = tf.nn.moments(content, axes=nn.conv2d_spatial_axes, keep_dims=True)
|
|
s_mean, s_var = tf.nn.moments(style, axes=nn.conv2d_spatial_axes, keep_dims=True)
|
|
c_std, s_std = tf.sqrt(c_var + 1e-5), tf.sqrt(s_var + 1e-5)
|
|
mean_loss = tf.reduce_sum(tf.square(c_mean-s_mean), axis=[1,2,3])
|
|
std_loss = tf.reduce_sum(tf.square(c_std-s_std), axis=[1,2,3])
|
|
return (mean_loss + std_loss) * ( loss_weight / content_nc.value )
|
|
|
|
if gaussian_blur_radius > 0.0:
|
|
target = tf_gaussian_blur(target, gaussian_blur_radius)
|
|
style = tf_gaussian_blur(style, gaussian_blur_radius)
|
|
|
|
return sd( target, style, loss_weight=loss_weight )
|
|
|
|
nn.tf_style_loss = tf_style_loss
|
|
|
|
def tf_dssim(img1,img2, max_val, filter_size=11, filter_sigma=1.5, k1=0.01, k2=0.03):
|
|
if img1.dtype != img2.dtype:
|
|
raise ValueError("img1.dtype != img2.dtype")
|
|
|
|
not_float32 = img1.dtype != tf.float32
|
|
|
|
if not_float32:
|
|
img_dtype = img1.dtype
|
|
img1 = tf.cast(img1, tf.float32)
|
|
img2 = tf.cast(img2, tf.float32)
|
|
|
|
kernel = np.arange(0, filter_size, dtype=np.float32)
|
|
kernel -= (filter_size - 1 ) / 2.0
|
|
kernel = kernel**2
|
|
kernel *= ( -0.5 / (filter_sigma**2) )
|
|
kernel = np.reshape (kernel, (1,-1)) + np.reshape(kernel, (-1,1) )
|
|
kernel = tf.constant ( np.reshape (kernel, (1,-1)), dtype=tf.float32 )
|
|
kernel = tf.nn.softmax(kernel)
|
|
kernel = tf.reshape (kernel, (filter_size, filter_size, 1, 1))
|
|
kernel = tf.tile (kernel, (1,1, img1.shape[ nn.conv2d_ch_axis ] ,1))
|
|
|
|
def reducer(x):
|
|
return tf.nn.depthwise_conv2d(x, kernel, strides=[1,1,1,1], padding='VALID', data_format=nn.data_format)
|
|
|
|
c1 = (k1 * max_val) ** 2
|
|
c2 = (k2 * max_val) ** 2
|
|
|
|
mean0 = reducer(img1)
|
|
mean1 = reducer(img2)
|
|
num0 = mean0 * mean1 * 2.0
|
|
den0 = tf.square(mean0) + tf.square(mean1)
|
|
luminance = (num0 + c1) / (den0 + c1)
|
|
|
|
num1 = reducer(img1 * img2) * 2.0
|
|
den1 = reducer(tf.square(img1) + tf.square(img2))
|
|
c2 *= 1.0 #compensation factor
|
|
cs = (num1 - num0 + c2) / (den1 - den0 + c2)
|
|
|
|
ssim_val = tf.reduce_mean(luminance * cs, axis=nn.conv2d_spatial_axes )
|
|
dssim = (1.0 - ssim_val ) / 2.0
|
|
|
|
if not_float32:
|
|
dssim = tf.cast(dssim, img_dtype)
|
|
return dssim
|
|
|
|
nn.tf_dssim = tf_dssim
|
|
|
|
def tf_space_to_depth(x, size):
|
|
if nn.data_format == "NHWC":
|
|
# match NCHW version in order to switch data_format without problems
|
|
b,h,w,c = x.shape.as_list()
|
|
oh, ow = h // size, w // size
|
|
x = tf.reshape(x, (-1, size, oh, size, ow, c))
|
|
x = tf.transpose(x, (0, 2, 4, 1, 3, 5))
|
|
x = tf.reshape(x, (-1, oh, ow, size* size* c ))
|
|
return x
|
|
else:
|
|
return tf.space_to_depth(x, size, data_format=nn.data_format)
|
|
nn.tf_space_to_depth = tf_space_to_depth
|
|
|
|
def tf_depth_to_space(x, size):
|
|
if nn.data_format == "NHWC":
|
|
# match NCHW version in order to switch data_format without problems
|
|
|
|
b,h,w,c = x.shape.as_list()
|
|
oh, ow = h * size, w * size
|
|
oc = c // (size * size)
|
|
|
|
x = tf.reshape(x, (-1, h, w, size, size, oc, ) )
|
|
x = tf.transpose(x, (0, 1, 3, 2, 4, 5))
|
|
x = tf.reshape(x, (-1, oh, ow, oc, ))
|
|
return x
|
|
else:
|
|
return tf.depth_to_space(x, size, data_format=nn.data_format)
|
|
nn.tf_depth_to_space = tf_depth_to_space
|
|
|
|
def tf_rgb_to_lab(srgb):
|
|
srgb_pixels = tf.reshape(srgb, [-1, 3])
|
|
linear_mask = tf.cast(srgb_pixels <= 0.04045, dtype=tf.float32)
|
|
exponential_mask = tf.cast(srgb_pixels > 0.04045, dtype=tf.float32)
|
|
rgb_pixels = (srgb_pixels / 12.92 * linear_mask) + (((srgb_pixels + 0.055) / 1.055) ** 2.4) * exponential_mask
|
|
rgb_to_xyz = tf.constant([
|
|
# X Y Z
|
|
[0.412453, 0.212671, 0.019334], # R
|
|
[0.357580, 0.715160, 0.119193], # G
|
|
[0.180423, 0.072169, 0.950227], # B
|
|
])
|
|
xyz_pixels = tf.matmul(rgb_pixels, rgb_to_xyz)
|
|
|
|
xyz_normalized_pixels = tf.multiply(xyz_pixels, [1/0.950456, 1.0, 1/1.088754])
|
|
|
|
epsilon = 6/29
|
|
linear_mask = tf.cast(xyz_normalized_pixels <= (epsilon**3), dtype=tf.float32)
|
|
exponential_mask = tf.cast(xyz_normalized_pixels > (epsilon**3), dtype=tf.float32)
|
|
fxfyfz_pixels = (xyz_normalized_pixels / (3 * epsilon**2) + 4/29) * linear_mask + (xyz_normalized_pixels ** (1/3)) * exponential_mask
|
|
|
|
fxfyfz_to_lab = tf.constant([
|
|
# l a b
|
|
[ 0.0, 500.0, 0.0], # fx
|
|
[116.0, -500.0, 200.0], # fy
|
|
[ 0.0, 0.0, -200.0], # fz
|
|
])
|
|
lab_pixels = tf.matmul(fxfyfz_pixels, fxfyfz_to_lab) + tf.constant([-16.0, 0.0, 0.0])
|
|
return tf.reshape(lab_pixels, tf.shape(srgb))
|
|
nn.tf_rgb_to_lab = tf_rgb_to_lab
|
|
|
|
def tf_suppress_lower_mean(t, eps=0.00001):
|
|
if t.shape.ndims != 1:
|
|
raise ValueError("tf_suppress_lower_mean: t rank must be 1")
|
|
t_mean_eps = tf.reduce_mean(t) - eps
|
|
q = tf.clip_by_value(t, t_mean_eps, tf.reduce_max(t) )
|
|
q = tf.clip_by_value(q-t_mean_eps, 0, eps)
|
|
q = q * (t/eps)
|
|
return q
|
|
"""
|
|
class GeLU(KL.Layer):
|
|
Gaussian Error Linear Unit.
|
|
A smoother version of ReLU generally used
|
|
in the BERT or BERT architecture based models.
|
|
Original paper: https://arxiv.org/abs/1606.08415
|
|
Input shape:
|
|
Arbitrary. Use the keyword argument `input_shape`
|
|
(tuple of integers, does not include the samples axis)
|
|
when using this layer as the first layer in a model.
|
|
Output shape:
|
|
Same shape as the input.
|
|
|
|
def __init__(self, approximate=True, **kwargs):
|
|
super(GeLU, self).__init__(**kwargs)
|
|
self.approximate = approximate
|
|
self.supports_masking = True
|
|
|
|
def call(self, inputs):
|
|
cdf = 0.5 * (1.0 + K.tanh((np.sqrt(2 / np.pi) * (inputs + 0.044715 * K.pow(inputs, 3)))))
|
|
return inputs * cdf
|
|
|
|
def get_config(self):
|
|
config = {'approximate': self.approximate}
|
|
base_config = super(GeLU, self).get_config()
|
|
return dict(list(base_config.items()) + list(config.items()))
|
|
|
|
def compute_output_shape(self, input_shape):
|
|
return input_shape
|
|
nn.GeLU = GeLU
|
|
""" |