return (keras.losses.BinaryCrossentropy(from_logits=True)
(tf.zeros_like(fake), real - tf.reduce_mean(fake)) +
keras.losses.BinaryCrossentropy(from_logits=True)
(tf.ones_like(real), fake - tf.reduce_mean(real))) / 2.0
def _dis_loss(self, real, fake):
return (keras.losses.BinaryCrossentropy(from_logits=True)
(tf.ones_like(real), real - tf.reduce_mean(fake)) +
keras.losses.BinaryCrossentropy(from_logits=True)
(tf.zeros_like(fake), fake - tf.reduce_mean(real))) / 2.0
if __name__ == "__main__":
k = 64
anime_path = f"./cat{k}.npy"
anime_imgs = np.load(anime_path)
anime_dataset = process_numpy(anime_imgs, batch_size=32)
ema = EMACallback()
show = ShowCallback()
model = RaGAN()
model.build(anime_imgs.shape[1:])
model.compile(d_optimizer=tfa.optimizers.AdamW(learning_rate=1e-4,
weight_decay=5e-5),
g_optimizer=tfa.optimizers.AdamW(learning_rate=1e-4,
weight_decay=5e-5),
loss=keras.losses.BinaryCrossentropy(from_logits=True))
model.fit(anime_dataset, epochs=10, callbacks=[ema, show])
show.make_gif("ragan.gif")
o = keras.layers.LeakyReLU(0.2)(o)
for l, s in layer_dict.items():
if l == img_shape[0] // 2:
break
o = keras.layers.Conv2DTranspose(filters=filter_num *
layer_dict[o.shape[1]],
kernel_size=1,
strides=2,
padding="SAME")(o)
o = keras.layers.BatchNormalization()(o)
o = keras.layers.LeakyReLU(0.2)(o)
o = keras.layers.Conv2DTranspose(filters=256,
kernel_size=1,
strides=2,
padding="SAME")(o)
return keras.Model(inputs=inputs, outputs=o)
dataset = process_numpy(img_data, seg_data, batch_size=8)
model = get_model()
model.summary()
keras.utils.plot_model(model, show_shapes=True)
model.compile(
optimizer=keras.optimizers.Adam(learning_rate=1e-4),
loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
metrics=CustomIoU(num_classes=256))
model.fit(dataset, epochs=20)
show(seg_data[:16], model(img_data[:16]))
import tensorflow as tf
import tensorflow_addons as tfa
from utils import EMACallback, process_directory, ShowCallback, process_numpy, StyleShowCallback, PathCallback
from hrgan import HrGAN
from pergan import PerGAN
from dcgan import DCGAN
from stylegan2 import StyleGAN2
from stylegan import StyleGAN
# anime_path = "E:/project/data/process/picture/seeprettyface_anime_face/anime_face"
# anime_dataset = process_directory(anime_path, batch_size=16)
anime_path = "E:/project/data/process/picture/anime64.npy"
anime_dataset = process_numpy(np.load(anime_path), batch_size=8)
ema = EMACallback()
# show = ShowCallback()
show = StyleShowCallback()
path = PathCallback()
model = HrGAN(latent_dim=128, filter_num=8)
model.build((64, 64, 3))
model.compile(d_optimizer=tfa.optimizers.AdamW(learning_rate=1e-4, weight_decay=5e-5, beta_1=0., beta_2=0.99),
g_optimizer=tfa.optimizers.AdamW(learning_rate=1e-4, weight_decay=5e-5, beta_1=0., beta_2=0.99),
loss=keras.losses.MeanSquaredError())
model.summary()
keras.utils.plot_model(model.gen, show_shapes=True, to_file="gen.png", expand_nested=True)
keras.utils.plot_model(model.dis, show_shapes=True, to_file="dis.png", expand_nested=True)
model.fit(anime_dataset, epochs=50, callbacks=[ema, show, path])