def main():
config = {
'name': '9L Conv Exponential Spline MNIST (lr1e-3)',
'eval_epochs': 1,
'sample_epochs': 1,
'log_interval': 100,
'lr': 1e-3,
'num_layers': 9,
'batch_size': 100,
'modified_grad': False,
'add_recon_grad': False,
'sym_recon_grad': False,
'activation': 'Spline',
'recon_loss_weight': 1.0,
'log_timing': True
}
train_loader, val_loader, test_loader = load_data(batch_size=config['batch_size'])
model = create_model(num_layers=config['num_layers'],
sym_recon_grad=config['sym_recon_grad'],
activation=config['activation'],
recon_loss_weight=config['recon_loss_weight']).to('cuda')
optimizer = optim.Adam(model.parameters(), lr=config['lr'], betas=(0.9, 0.999))
scheduler = StepLR(optimizer, step_size=1, gamma=1.0)
experiment = Experiment(model, train_loader, val_loader, test_loader,
optimizer, scheduler, **config)
experiment.run()
def main():
config = {
'name': 'Timing Experiment SNF FC',
'eval_epochs': 1,
'sample_epochs': 1,
'log_interval': 10000,
'lr': 1e-4,
'batch_size': 100,
'modified_grad': True,
'add_recon_grad': True,
'sym_recon_grad': False,
'only_R_recon': False,
'actnorm': False,
'split_prior': False,
'activation': 'None',
'log_timing': True,
'epochs': 100
}
train_loader, val_loader, test_loader = load_data(
batch_size=config['batch_size'])
model = create_model(data_size=(1, 28, 28), layer='fc').to('cuda')
optimizer = optim.Adam(model.parameters(),
lr=config['lr'],
betas=(0.9, 0.999))
scheduler = StepLR(optimizer, step_size=1, gamma=1.0)
experiment = Experiment(model, train_loader, val_loader, test_loader,
optimizer, scheduler, **config)
experiment.run()
def run_timing_experiment(name, snf_name, config, sz, m, results):
train_loader, val_loader, test_loader = load_data(
batch_size=config['batch_size'],
im_size=sz,
n_train=50_000,
n_val=100,
n_test=100)
model = create_model(data_size=sz, layer=m).to('cuda')
optimizer = optim.Adam(model.parameters(),
lr=config['lr'],
betas=(0.9, 0.999))
scheduler = StepLR(optimizer, step_size=1, gamma=1.0)
experiment = Experiment(model, train_loader, val_loader, test_loader,
optimizer, scheduler, **config)
experiment.run()
mean_time = experiment.summary['Batch Time Mean']
std_time = experiment.summary['Batch Time Std']
print(f"{name}: {mean_time} +/- {std_time}")
results[f'{m} {snf_name}']['n_params'].append(sz[0] * sz[1] * sz[2])
results[f'{m} {snf_name}']['mean'].append(mean_time)
results[f'{m} {snf_name}']['std'].append(std_time)
return results
def main():
config = {
'name': '2L-4K-16W Glow SNF (5x5 Kernel) MNIST w/ GECO',
'eval_epochs': 1,
'sample_epochs': 1,
'log_interval': 100,
'lr': 1e-3,
'num_blocks': 2,
'block_size': 4,
'width': 16,
'batch_size': 100,
'modified_grad': True,
'add_recon_grad': True,
'sym_recon_grad': False,
'actnorm': True,
'split_prior': True,
'activation': 'None',
'recon_loss_weight': 1.0,
'recon_loss_lr': 1e-3,
'recon_alpha': 0.9,
'sample_true_inv': True,
'plot_recon': True,
'vis_epochs': 1,
'log_timing': False,
'epochs': 1000
}
train_loader, val_loader, test_loader = load_data(
data_aug=False, batch_size=config['batch_size'])
model = create_model(num_blocks=config['num_blocks'],
block_size=config['block_size'],
width=config['width'],
sym_recon_grad=config['sym_recon_grad'],
actnorm=config['actnorm'],
split_prior=config['split_prior'],
recon_loss_weight=config['recon_loss_weight'],
recon_loss_lr=config['recon_loss_lr'],
recon_alpha=config['recon_alpha']).to('cuda')
optimizer = optim.Adam(model.parameters(),
lr=config['lr'],
betas=(0.9, 0.999))
scheduler = StepLR(optimizer, step_size=1, gamma=1.0)
experiment = Experiment(model, train_loader, val_loader, test_loader,
optimizer, scheduler, **config)
experiment.run()
def main():
config = {
'name': '3L-48K Glow Exact Imagenet32',
'eval_epochs': 1,
'sample_epochs': 1,
'log_interval': 100,
'lr': 1e-3,
'num_blocks': 3,
'block_size': 48,
'batch_size': 64,
'modified_grad': False,
'add_recon_grad': False,
'sym_recon_grad': False,
'actnorm': True,
'split_prior': True,
'activation': 'None',
'recon_loss_weight': 0.0,
'sample_true_inv': False,
'plot_recon': False,
'grad_clip_norm': 10_000,
'warmup_epochs': 0
}
train_loader, val_loader, test_loader = load_data(
data_aug=False,
resolution=32,
data_dir='data/imagenet',
batch_size=config['batch_size'])
model = create_model(
num_blocks=config['num_blocks'],
block_size=config['block_size'],
sym_recon_grad=config['sym_recon_grad'],
actnorm=config['actnorm'],
split_prior=config['split_prior'],
recon_loss_weight=config['recon_loss_weight']).to('cuda')
optimizer = optim.Adam(model.parameters(),
lr=config['lr'],
betas=(0.9, 0.999))
scheduler = StepLR(optimizer, step_size=1, gamma=1.0)
experiment = Experiment(model, train_loader, val_loader, test_loader,
optimizer, scheduler, **config)
experiment.run()
def main():
config = {
'name': '2L-16K Glow SNF(1x1) recon 100x MNIST (lr1e-3)',
'eval_epochs': 1,
'sample_epochs': 1,
'log_interval': 100,
'lr': 1e-3,
'num_blocks': 2,
'block_size': 16,
'batch_size': 100,
'modified_grad': True,
'add_recon_grad': True,
'sym_recon_grad': False,
'actnorm': True,
'split_prior': True,
'activation': 'None',
'recon_loss_weight': 100.0,
'sample_true_inv': True,
'plot_recon': True
}
train_loader, val_loader, test_loader = load_data(
batch_size=config['batch_size'])
model = create_model(
num_blocks=config['num_blocks'],
block_size=config['block_size'],
sym_recon_grad=config['sym_recon_grad'],
actnorm=config['actnorm'],
split_prior=config['split_prior'],
recon_loss_weight=config['recon_loss_weight']).to('cuda')
optimizer = optim.Adam(model.parameters(),
lr=config['lr'],
betas=(0.9, 0.999))
scheduler = StepLR(optimizer, step_size=1, gamma=1.0)
experiment = Experiment(model, train_loader, val_loader, test_loader,
optimizer, scheduler, **config)
experiment.run()
'warmup_epochs': 0,
'step_epochs': 1,
'step_gamma': 1.0,
}
train_loader, val_loader, test_loader = load_data(
data_aug=False,
resolution=32,
data_dir='data/imagenet',
batch_size=config['batch_size'])
model = create_model(
num_blocks=config['num_blocks'],
block_size=config['block_size'],
sym_recon_grad=config['sym_recon_grad'],
actnorm=config['actnorm'],
split_prior=config['split_prior'],
recon_loss_weight=config['recon_loss_weight']).to('cuda')
optimizer = optim.Adam(model.parameters(),
lr=config['lr'],
betas=(0.9, 0.999))
scheduler = StepLR(optimizer,
step_size=config['step_epochs'],
gamma=config['step_gamma'])
experiment = Experiment(model, train_loader, val_loader, test_loader,
optimizer, scheduler, **config)
experiment.run()
