def to_image(inp, style, channels, image_size):
# Not sure if this works for rectangular images
x = Conv2DMod(channels,
1,
kernel_initializer=VarianceScaling(200 / inp.shape[2]),
demod=False)([inp, style])
return Lambda(Generator.upsample_to_size,
output_shape=[None, image_size[0], image_size[1],
None])([x, image_size])
def c(i, style, noise, l, l_):
o = Conv2DMod(layer_dict[l_] * self.filter_num,
kernel_size=3,
strides=1)([i, style])
noise_crop = keras.layers.Cropping2D(
(input_shape[0] - l_) // 2)(noise)
o_n = conv(layer_dict[l_] * self.filter_num,
kernel_size=1,
strides=1)(noise_crop)
o = keras.layers.Add()([o, o_n])
o = act_layer(o)
o = Conv2DMod(layer_dict[l_] * self.filter_num,
kernel_size=3,
strides=1)([o, style])
o_n = conv(layer_dict[l_] * self.filter_num,
kernel_size=1,
strides=1)(noise_crop)
o = keras.layers.Add()([o, o_n])
o = act_layer(o)
return o
def _generator_block(inp,
input_style,
input_noise,
filters,
final_image_size,
upsample=True,
channels=3):
if upsample:
out = UpSampling2D(interpolation='bilinear')(inp)
else:
out = inp
image_style = Dense(filters,
kernel_initializer=VarianceScaling(
200 / out.shape[2]))(input_style)
style = Dense(inp.shape[-1],
kernel_initializer='he_uniform')(input_style)
cropped_noise = Lambda(
Generator.crop_noise_to_size)([input_noise, out])
noise = Dense(filters, kernel_initializer='zeros')(cropped_noise)
out = Conv2DMod(filters=filters,
kernel_size=3,
padding='same',
kernel_initializer='he_uniform')([out, style])
out = Add()([out, noise])
out = LeakyReLU(0.2)(out)
style = Dense(filters, kernel_initializer='he_uniform')(input_style)
noise = Dense(filters, kernel_initializer='zeros')(cropped_noise)
out = Conv2DMod(filters=filters,
kernel_size=3,
padding='same',
kernel_initializer='he_uniform')([out, style])
out = Add()([out, noise])
out = LeakyReLU(0.2)(out)
return out, Generator.to_image(out, image_style, channels,
final_image_size)
def up(i, style, noise, l, l_):
o = keras.layers.UpSampling2D(size=l_ // l,
interpolation="bilinear")(i)
o = Conv2DMod(layer_dict[l_] * self.filter_num,
kernel_size=3,
strides=1)([o, style])
noise_crop = keras.layers.Cropping2D(
(input_shape[0] - l_) // 2)(noise)
o_n = conv(layer_dict[l_] * self.filter_num,
kernel_size=1,
strides=1)(noise_crop)
o = keras.layers.Add()([o, o_n])
o = act_layer(o)
o = Conv2DMod(layer_dict[l_] * self.filter_num,
kernel_size=3,
strides=1)([o, style])
o_n = conv(layer_dict[l_] * self.filter_num,
kernel_size=1,
strides=1)(noise_crop)
o = keras.layers.Add()([o, o_n])
o = act_layer(o)
return o
def build(self, input_shape):
self.M = keras.Sequential(
[keras.layers.InputLayer(input_shape=(self.latent_dim, ))])
for _ in range(8):
self.M.add(
keras.layers.Dense(self.latent_dim,
kernel_initializer="he_normal"))
self.M.add(keras.layers.LeakyReLU(0.2))
style_list = []
inp = keras.layers.Input(shape=(self.latent_dim, ))
noise = keras.layers.Input(shape=input_shape)
outs = []
self.n = 0
for i, f in layer_dict.items():
if i == min(layer_dict.keys()):
style = keras.layers.Input(shape=(self.latent_dim, ))
style_list.append(style)
self.n += 1
o = keras.layers.Dense(self.latent_dim)(inp)
o = o[:, tf.newaxis, tf.newaxis, :]
o = up(o, size=i)
o = Conv2DMod(f * self.filter_num, kernel_size=3,
strides=1)([o, style])
noise_crop = keras.layers.Cropping2D(
(input_shape[0] - i) // 2)(noise)
o_n = conv(f * self.filter_num, kernel_size=1,
strides=1)(noise_crop)
o = keras.layers.Add()([o, o_n])
o = act_layer(o)
o = Conv2DMod(f * self.filter_num, kernel_size=3,
strides=1)([o, style])
o_n = conv(f * self.filter_num, kernel_size=1,
strides=1)(noise_crop)
o = keras.layers.Add()([o, o_n])
o = act_layer(o)
rgb = Conv2DMod(input_shape[-1],
kernel_size=1,
strides=1,
demod=False)([o, style])
outs.append(up(rgb, size=input_shape[0] // i))
else:
style = keras.layers.Input(shape=(self.latent_dim, ))
style_list.append(style)
self.n += 1
o = up(o, size=2)
o = Conv2DMod(f * self.filter_num, kernel_size=3,
strides=1)([o, style])
noise_crop = keras.layers.Cropping2D(
(input_shape[0] - i) // 2)(noise)
o_n = conv(f * self.filter_num, kernel_size=1,
strides=1)(noise_crop)
o = keras.layers.Add()([o, o_n])
o = act_layer(o)
o = Conv2DMod(f * self.filter_num, kernel_size=3,
strides=1)([o, style])
o_n = conv(f * self.filter_num, kernel_size=1,
strides=1)(noise_crop)
o = keras.layers.Add()([o, o_n])
o = act_layer(o)
rgb = Conv2DMod(input_shape[-1],
kernel_size=1,
strides=1,
demod=False)([o, style])
outs.append(up(rgb, size=input_shape[0] // i))
if i == input_shape[0]:
self.m = self.n
break
o = keras.layers.Add()(outs)
# o = keras.layers.Activation("tanh")(o)
self._gen = keras.Model(inputs=[inp, noise] + style_list, outputs=o)
style_list_n = []
o_list = []
inp = keras.layers.Input(shape=(self.latent_dim, ))
noise = keras.layers.Input(shape=input_shape)
for _ in style_list:
style = keras.layers.Input(shape=(self.latent_dim, ))
style_list_n.append(style)
o_list.append(self.M(style))
o = self._gen([inp, noise] + o_list)
self.gen = keras.Model(inputs=[inp, noise] + style_list_n, outputs=o)
img = keras.layers.Input(shape=input_shape)
o = AugmentLayer()(img)
for i, f in reversed(layer_dict.items()):
if i < input_shape[0]:
res = conv(f * self.filter_num, kernel_size=3, strides=2)(o)
o = conv(f * self.filter_num, kernel_size=3, strides=2)(o)
o = norm_layer(o)
o = act_layer(o)
o = conv(f * self.filter_num, kernel_size=3, strides=1)(o)
o = norm_layer(o)
o = keras.layers.Add()([res, o])
o = act_layer(o)
o = keras.layers.Flatten()(o)
self.dis = keras.Model(inputs=img, outputs=o)
self.perform_gp = True
self.perform_pl = True
self.pl_mean = 0
self.pl_length = 0.
def build(self, input_shape):
assert input_shape[0] >= 32
self.M = keras.Sequential(
[keras.layers.InputLayer(input_shape=(self.latent_dim, ))])
for _ in range(5):
self.M.add(
keras.layers.Dense(self.latent_dim,
kernel_initializer="he_normal"))
self.M.add(keras.layers.LeakyReLU(0.2))
def up(i, style, noise, l, l_):
o = keras.layers.UpSampling2D(size=l_ // l,
interpolation="bilinear")(i)
o = Conv2DMod(layer_dict[l_] * self.filter_num,
kernel_size=3,
strides=1)([o, style])
noise_crop = keras.layers.Cropping2D(
(input_shape[0] - l_) // 2)(noise)
o_n = conv(layer_dict[l_] * self.filter_num,
kernel_size=1,
strides=1)(noise_crop)
o = keras.layers.Add()([o, o_n])
o = act_layer(o)
o = Conv2DMod(layer_dict[l_] * self.filter_num,
kernel_size=3,
strides=1)([o, style])
o_n = conv(layer_dict[l_] * self.filter_num,
kernel_size=1,
strides=1)(noise_crop)
o = keras.layers.Add()([o, o_n])
o = act_layer(o)
return o
def down(i, style, noise, l, l_):
o = Conv2DMod(layer_dict[l_] * self.filter_num,
kernel_size=3,
strides=l // l_)([i, style])
noise_crop = keras.layers.Cropping2D(
(input_shape[0] - l_) // 2)(noise)
o_n = conv(layer_dict[l_] * self.filter_num,
kernel_size=1,
strides=1)(noise_crop)
o = keras.layers.Add()([o, o_n])
o = act_layer(o)
o = Conv2DMod(layer_dict[l_] * self.filter_num,
kernel_size=3,
strides=1)([o, style])
o_n = conv(layer_dict[l_] * self.filter_num,
kernel_size=1,
strides=1)(noise_crop)
o = keras.layers.Add()([o, o_n])
o = act_layer(o)
return o
def c(i, style, noise, l, l_):
o = Conv2DMod(layer_dict[l_] * self.filter_num,
kernel_size=3,
strides=1)([i, style])
noise_crop = keras.layers.Cropping2D(
(input_shape[0] - l_) // 2)(noise)
o_n = conv(layer_dict[l_] * self.filter_num,
kernel_size=1,
strides=1)(noise_crop)
o = keras.layers.Add()([o, o_n])
o = act_layer(o)
o = Conv2DMod(layer_dict[l_] * self.filter_num,
kernel_size=3,
strides=1)([o, style])
o_n = conv(layer_dict[l_] * self.filter_num,
kernel_size=1,
strides=1)(noise_crop)
o = keras.layers.Add()([o, o_n])
o = act_layer(o)
return o
style_list = []
inp = keras.layers.Input(shape=(self.latent_dim, ))
noise = keras.layers.Input(shape=input_shape)
outs = []
o_dict = {}
l_list = [min(layer_dict.keys())]
for l, s in layer_dict.items():
if l == min(layer_dict.keys()):
style = keras.layers.Input(shape=(self.latent_dim, ))
style_list.append(style)
o = keras.layers.Dense(self.latent_dim)(inp)
o = o[:, tf.newaxis, tf.newaxis, :]
o = up(o, style, noise, 1, l)
o_dict[min(layer_dict.keys())] = o
if l <= input_shape[0]:
l_list.append(l)
t_dict = dict()
for l in l_list[:-1]:
for l_ in l_list:
t_dict.setdefault(l_, [])
style = keras.layers.Input(shape=(self.latent_dim, ))
style_list.append(style)
if l == l_:
o = c(o_dict[l], style, noise, l, l_)
t_dict[l_].append(o)
elif l > l_:
o = down(o_dict[l], style, noise, l, l_)
t_dict[l_].append(o)
else:
o = up(o_dict[l], style, noise, l, l_)
t_dict[l_].append(o)
for l, os in t_dict.items():
if len(os) == 1:
o = os[0]
else:
o = keras.layers.Add()(os)
o = conv(filters=layer_dict[l] * self.filter_num,
kernel_size=1,
strides=1)(o)
o = norm_layer(o)
o = act_layer(o)
o_dict[l] = o
for l, o in o_dict.items():
style = keras.layers.Input(shape=(self.latent_dim, ))
style_list.append(style)
rgb = Conv2DMod(input_shape[-1],
kernel_size=1,
strides=1,
demod=False)([o, style])
rgb = keras.layers.UpSampling2D(size=input_shape[1] // l,
interpolation="bilinear")(rgb)
outs.append(rgb)
o = keras.layers.Add()(outs)
self._gen = keras.Model(inputs=[inp, noise] + style_list, outputs=o)
style_list_n = []
o_list = []
inp = keras.layers.Input(shape=(self.latent_dim, ))
noise = keras.layers.Input(shape=input_shape)
self.n = len(style_list)
self.m = self.n
for _ in style_list:
style = keras.layers.Input(shape=(self.latent_dim, ))
style_list_n.append(style)
o_list.append(self.M(style))
o = self._gen([inp, noise] + o_list)
self.gen = keras.Model(inputs=[inp, noise] + style_list_n, outputs=o)
def up(i, l, l_):
o = convt(filters=layer_dict[l_] * self.filter_num,
kernel_size=3,
strides=l_ // l)(i)
o = norm_layer(o)
o = act_layer(o)
return o
def down(i, l, l_):
o = conv(filters=layer_dict[l_] * self.filter_num,
kernel_size=3,
strides=l // l_)(i)
o = norm_layer(o)
o = act_layer(o)
return o
def c(i, l, l_):
o = conv(filters=layer_dict[l_] * self.filter_num,
kernel_size=3,
strides=1)(i)
o = norm_layer(o)
o = act_layer(o)
return o
images = keras.layers.Input(shape=input_shape)
o = AugmentLayer()(images)
o_dict = dict()
l_list = []
for l, s in reversed(layer_dict.items()):
if l == input_shape[0]:
o = conv(filters=layer_dict[l] * self.filter_num,
kernel_size=3,
strides=1)(o)
o = norm_layer(o)
o = act_layer(o)
o_dict[l] = o
l_list.append(l)
if l < input_shape[0]:
l_list.append(l)
t_dict = dict()
for l in l_list[:-1]:
for l_ in l_list:
t_dict.setdefault(l_, [])
if l == l_:
o = c(o_dict[l], l, l_)
# o = o_dict[l]
t_dict[l_].append(o)
elif l > l_:
o = down(o_dict[l], l, l_)
t_dict[l_].append(o)
else:
o = up(o_dict[l], l, l_)
t_dict[l_].append(o)
for l, os in t_dict.items():
if len(os) == 1:
o = os[0]
else:
o = keras.layers.Add()(os)
o = conv(filters=64, kernel_size=1, strides=1)(o)
o = norm_layer(o)
o = act_layer(o)
o = conv(filters=64, kernel_size=3, strides=1)(o)
o = norm_layer(o)
o = act_layer(o)
o = conv(filters=64, kernel_size=1, strides=1)(o)
o = norm_layer(o)
o = act_layer(o)
o_dict[l] = o
os = []
for l, o in o_dict.items():
if l >= 32:
o = conv(filters=input_shape[-1], kernel_size=1, strides=1)(o)
o = norm_layer(o)
o = act_layer(o)
os.append(o)
o = keras.layers.Flatten()(o_dict[min(layer_dict.keys())])
self.dis = keras.Model(inputs=images, outputs=[o] + os)
self.perform_gp = True
self.perform_pl = True
self.pl_mean = 0
self.pl_length = 0.