def generator(nb_fake, nb_real, z_dim):
for real in real_generator("/home/leon/data/tags_plain_t6.hdf5", nb_real):
z = np.random.uniform(-1, 1, (nb_fake, z_dim)).astype(np.float32)
yield {
'real': 2*real - 1,
GAN.z_name: z
}
def train_g64_d64_fine_tune():
nb_units = 64
generator_input_dim = 100
nb_real = 64
nb_fake = 128 + nb_real
lr = 0.00002
beta_1 = 0.5
nb_batches_per_epoch = 100
nb_epoch = 60
output_dir = "models/dcgan_g64_d64_fine_tune"
hdf5_fname = "/home/leon/data/tags_plain_t6.hdf5"
g = dcgan_generator(nb_units, generator_input_dim)
d = dcgan_discriminator(nb_units)
g.load_weights("models/dcgan_g64_d64/generator.hdf5")
d.load_weights("models/dcgan_g64_d64/fix_discriminator.hdf5")
gan = sequential_to_gan(g,
d,
nb_real,
nb_fake,
nb_fake_for_gen=128,
nb_fake_for_dis=nb_real)
save = SaveModels(
{
"generator_{epoch:04d}.hdf5": g,
"discriminator_{epoch:04d}.hdf5": d
},
every_epoch=20,
output_dir=output_dir)
visual = VisualiseGAN(nb_samples=13**2,
output_dir=output_dir,
preprocess=lambda x: np.clip(x, -1, 1))
real_z_gen = zip_real_z(real_generator(hdf5_fname, nb_real, range=(-1, 1)),
z_generator((nb_fake, generator_input_dim)))
history = train_dcgan(gan,
Adam(lr, beta_1),
Adam(lr, beta_1),
real_z_gen,
nb_batches_per_epoch=nb_batches_per_epoch,
nb_epoch=nb_epoch,
callbacks=[save, visual])
with open(os.path.join(output_dir, "history.json"), 'w+') as f:
json.dump(history.history, f)
def train_data_generator(real_hdf5_fname, batch_size, z_dim, translation=2):
for data in real_generator(real_hdf5_fname, batch_size, range=(-1, 1)):
z = np.random.uniform(-1, 1, (3*batch_size, z_dim)).astype(np.float32)
size = 64
if data.shape[-2:] != (size, size):
translated_data = []
for sample in data:
random_translation = np.random.choice(np.arange(-translation, translation+1))
offset = (data.shape[-1] - size) // 2 + random_translation
crop = slice(offset, size+offset)
translated_data.append(sample[:, crop, crop])
data = np.stack(translated_data)
yield {
'data': data,
'z': z,
}
def train_data_generator(real_hdf5_fname, batch_size, z_dim, translation=2):
for data in real_generator(real_hdf5_fname, batch_size, range=(-1, 1)):
z = np.random.uniform(-1, 1,
(3 * batch_size, z_dim)).astype(np.float32)
size = 64
if data.shape[-2:] != (size, size):
translated_data = []
for sample in data:
random_translation = np.random.choice(
np.arange(-translation, translation + 1))
offset = (data.shape[-1] - size) // 2 + random_translation
crop = slice(offset, size + offset)
translated_data.append(sample[:, crop, crop])
data = np.stack(translated_data)
yield {
'data': data,
'z': z,
}
def train_g64_d64_dct():
nb_units = 64
generator_input_dim = 25
nb_real = 64
nb_fake = 96
lr = 0.0002
beta_1 = 0.5
nb_batches_per_epoch = 100
nb_epoch = 1000
output_dir = "models/dcgan_g64_d64_dct"
hdf5_fname = "/home/leon/data/tags_plain_t6.hdf5"
g = dcgan_generator(nb_units, generator_input_dim)
d = dcgan_discriminator(nb_units)
gan = sequential_to_gan(g, d, nb_real, nb_fake)
save = SaveModels({
"generator.hdf5": g,
"discriminator.hdf5": d
},
output_dir=output_dir)
visual = VisualiseGAN(nb_samples=13**2,
output_dir=output_dir,
preprocess=lambda x: np.clip(x, -1, 1))
real_z_gen = zip_real_z(real_generator(hdf5_fname, nb_real),
z_generator((nb_fake, generator_input_dim)))
history = train_dcgan(gan,
Adam(lr, beta_1),
Adam(lr, beta_1),
real_z_gen,
nb_batches_per_epoch=nb_batches_per_epoch,
nb_epoch=nb_epoch,
callbacks=[save, visual])
with open(os.path.join(output_dir, "history.json"), 'w+') as f:
json.dump(history.history, f)
def generator(nb_fake, nb_real, z_dim):
for real in real_generator("/home/leon/data/tags_plain_t6.hdf5", nb_real):
z = np.random.uniform(-1, 1, (nb_fake, z_dim)).astype(np.float32)
yield {'real': 2 * real - 1, GAN.z_name: z}