def test_conditionalwgan(self):
bn = False
ns = 16
# Parameters for the generator
params_generator = dict()
params_generator['latent_dim'] = 126
params_generator['stride'] = [1, 2, 1, 1]
params_generator['nfilter'] = [32, 64, 32, 1]
params_generator['shape'] = [[5, 5], [5, 5], [5, 5], [5, 5]]
params_generator['batch_norm'] = [bn, bn, bn]
params_generator['full'] = [8 * 8 * 32]
params_generator['non_lin'] = tf.nn.relu
params_generator['in_conv_shape'] = [8, 8]
# Parameters for the discriminator
params_discriminator = dict()
params_discriminator['stride'] = [2, 2, 1]
params_discriminator['nfilter'] = [16, 64, 32]
params_discriminator['shape'] = [[5, 5], [5, 5], [3, 3]]
params_discriminator['batch_norm'] = [bn, bn, bn]
params_discriminator['full'] = [512, 128, 32]
params_discriminator['minibatch_reg'] = False
# Optimization parameters
d_opt = dict()
d_opt['optimizer'] = "rmsprop"
d_opt['learning_rate'] = 3e-5
params_optimization = dict()
params_optimization['n_critic'] = 5
params_optimization['batch_size'] = 8
params_optimization['epoch'] = 1
# all parameters
params = dict()
params['net'] = dict() # All the parameters for the model
params['net']['generator'] = params_generator
params['net']['discriminator'] = params_discriminator
params['net']['shape'] = [ns, ns, 1] # Shape of the image
params['net']['gamma_gp'] = 10 # Gradient penalty
# Conditional params
params['net']['prior_normalization'] = False
params['net']['cond_params'] = 2
params['net']['init_range'] = [[0, 1], [0, 1]]
params['net']['prior_distribution'] = "gaussian_length"
params['net']['final_range'] = [
0.1 * np.sqrt(params_generator['latent_dim']),
1 * np.sqrt(params_generator['latent_dim'])
]
params['optimization'] = params_optimization
params['summary_every'] = 4
params['save_every'] = 5
params['print_every'] = 3
X = np.random.rand(101, ns, ns)
parameters = np.random.rand(101, 2)
dataset = Dataset_parameters(X, parameters)
wgan = GANsystem(ConditionalParamWGAN, params)
wgan.train(dataset)
img = wgan.generate(N=16,
**{
'z':
wgan.net.sample_latent(bs=16,
params=np.random.rand(
16, 2))
})
assert (img.shape[0] == 16 and img.shape[1] == ns
and img.shape[2] == ns and img.shape[3] == 1)
def test_default_params_lapgan(self):
obj = GANsystem(LapWGAN)
def test_default_params_patchgan(self):
obj = GANsystem(UpscalePatchWGAN)
print("consistency_contribution", consistency_contribution)
self._R_Con = self.consistency((self.X_real[:, :, :, 0] - 1) * 5)
self._F_Con = self.consistency((self.X_fake[:, :, :, 0] - 1) * 5)
self._mean_R_Con, self._std_R_Con = tf.nn.moments(self._R_Con,
axes=[0])
self._mean_F_Con, self._std_F_Con = tf.nn.moments(self._F_Con,
axes=[0])
self._G_Reg = tf.abs(self._mean_R_Con - self._mean_F_Con)
self._G_loss += consistency_contribution * self._G_Reg
def _build_image_summary(self):
vmin = tf.reduce_min(self.X_real)
vmax = tf.reduce_max(self.X_real)
X_real = self.X_real[:, :, :, 0]
X_fake = self.X_fake[:, :, :, 0]
tf.summary.image("images/Real_Image",
colorize(X_real, vmin, vmax),
max_outputs=4,
collections=['model'])
tf.summary.image("images/Fake_Image",
colorize(X_fake, vmin, vmax),
max_outputs=4,
collections=['model'])
wgan = GANsystem(ModSpectrogramGAN, params)
wgan.train(dataset, resume=resume)
def test_default_params_wgan(self):
obj = GANsystem(WGAN)
params['net']['loss_type'] = 'wasserstein'
params['optimization'] = params_optimization
params[
'summary_every'] = 100 # Tensorboard summaries every ** iterations
params['print_every'] = 50 # Console summaries every ** iterations
params['save_every'] = 1000 # Save the model every ** iterations
params['summary_dir'] = os.path.join(global_path, name + '_summary/')
params['save_dir'] = os.path.join(global_path, name + '_checkpoints/')
params['Nstats'] = 0
resume, params = utils.test_resume(True, params)
# Build the model
print('Load the model')
wgan = GANsystem(InpaintingGAN, params)
# Generate new samples
print('Generate new samples')
real_signals = dataset.get_samples(N=N_f)
if model == 'extend':
border1 = real_signals[:, signal_split[0]:(signal_split[0] +
signal_split[1])]
border2 = real_signals[:, -(signal_split[3] +
signal_split[4]):-signal_split[4]]
border3 = real_signals[:, :(signal_split[0] + signal_split[1])]
border4 = real_signals[:, -(signal_split[3] + signal_split[4]):]
borders1 = np.stack([border1, border2], axis=2)
borders2 = np.stack([border3, border4], axis=2)
fake_signals = np.squeeze(wgan.generate(N=N_f,
borders1=borders1,
params['net']['loss_type'] = 'wasserstein'
params['optimization'] = params_optimization
params['summary_every'] = 100 # Tensorboard summaries every ** iterations
params['print_every'] = 50 # Console summaries every ** iterations
params['save_every'] = 1000 # Save the model every ** iterations
params['summary_dir'] = os.path.join(global_path, name + '_summary/')
params['save_dir'] = os.path.join(global_path, name + '_checkpoints/')
params['Nstats'] = 0
resume, params = utils.test_resume(False, params)
#%%
# # Build the model
wgan = GANsystem(InpaintingGAN, params)
# # Train the model
wgan.train(dataset, resume=resume)
end = time.time()
print('Elapse time: {} minutes'.format((end - start) / 60))
# =============================================================================
# #%%
# # # Generate new samples
# # To have meaningful statistics, be sure to generate enough samples
# # * 2000 : 32 x 32
# # * 500 : 64 x 64
# # * 200 : 128 x 128
params['optimization'] = params_optimization
params['summary_every'] = 100 # Tensorboard summaries every ** iterations
params['print_every'] = 50 # Console summaries every ** iterations
params['save_every'] = 1000 # Save the model every ** iterations
params['summary_dir'] = os.path.join(global_path, name + '_summary/')
params['save_dir'] = os.path.join(global_path, name + '_checkpoints/')
params['Nstats'] = 500
resume, params = utils.test_resume(True, params)
params['optimization']['epoch'] = 10000
params['summary_every'] = 250
params['print_every'] = 250
params['save_dir'] = os.path.join(global_path, name + '_checkpoints/')
wgan = GANsystem(SpectrogramGAN, params)
nsamples = 256
nlatent = 100
def clip_dist2(nsamples, nlatent, m=2.5):
shape = [nsamples, nlatent]
z = np.random.randn(*shape)
support = np.logical_or(z < -m, z > m)
while np.sum(support):
z[support] = np.random.randn(*shape)[support]
support = np.logical_or(z < -m, z > m)
return z
def test_default_params(self):
wgan = GANsystem(WGAN)
wgan = DualGANsystem(WGAN)
params['net']['gamma_gp'] = 10 # Gradient penalty
params['optimization'] = params_optimization
params['summary_every'] = 500 # Tensorboard summaries every ** iterations
params['print_every'] = 50 # Console summaries every ** iterations
params['save_every'] = 1000 # Save the model every ** iterations
params['summary_dir'] = os.path.join(global_path, name + '_summary/')
params['save_dir'] = os.path.join(global_path, name + '_checkpoints/')
params['Nstats'] = 100
params['Nstats_cubes'] = 10
resume, params = utils.test_resume(try_resume, params)
class CosmoUpscalePatchWGAN(UpscalePatchWGAN, CosmoWGAN):
pass
wgan = GANsystem(CosmoUpscalePatchWGAN, params)
dataset = load.load_nbody_dataset(spix=ns,
scaling=1,
resolution=256,
Mpch=350,
patch=True,
augmentation=True,
forward_map=forward,
is_3d=True)
wgan.train(dataset, resume=resume)
def test_default_params(self):
wgan = GANsystem(WGAN)
wgan = PaulinaGANsystem(WGAN)
params_cosmology['forward_map'] = forward
params_cosmology['backward_map'] = backward
# all parameters
params = dict()
params['net'] = dict() # All the parameters for the model
params['net']['generator'] = params_generator
params['net']['discriminator'] = params_discriminator
params['net'][
'cosmology'] = params_cosmology # Parameters for the cosmological summaries
params['net']['prior_distribution'] = 'gaussian'
params['net']['shape'] = [ns, ns, 1] # Shape of the image
params['net']['loss_type'] = 'wasserstein' # loss ('hinge' or 'wasserstein')
params['net']['gamma_gp'] = 10 # Gradient penalty
params['optimization'] = params_optimization
params['summary_every'] = 500 # Tensorboard summaries every ** iterations
params['print_every'] = 50 # Console summaries every ** iterations
params['save_every'] = 2000 # Save the model every ** iterations
params['summary_dir'] = os.path.join(global_path, name + '_summary/')
params['save_dir'] = os.path.join(global_path, name + '_checkpoints/')
params['Nstats'] = (64 * 32 * 32) // ns
resume, params = utils.test_resume(try_resume, params)
# If a model is reloaded and some parameters have to be changed, then it should be done here.
# For example, setting the number of epoch to 5 would be:
wgan = GANsystem(CosmoWGAN, params)
wgan.train(dataset, resume=resume)
