load_dict = checkpoint_io.load(model_file)
except FileNotFoundError:
it = epoch_idx = -1
else:
it = load_dict.get('it', -1)
epoch_idx = load_dict.get('epoch_idx', -1)
fid_best = load_dict.get('fid_best', float('inf'))
logger.load_stats('stats.p')
# Additional losses to GAN loss
losses_g, losses_g_2d = build_g_losses(generator.module, config)
# Trainer
trainer = Trainer(
generator, discriminator, g_optimizer, d_optimizer,
'standard', 'real', 10.,
losses_g=losses_g, losses_g_2d=losses_g_2d,
n_labels=nlabels
)
# Training loop
print('Start training...')
while True:
epoch_idx += 1
print('Start epoch %d...' % epoch_idx)
for x_real, y in train_loader:
it += 1
x_real, y = x_real.to(device), y.to(device)
y.clamp_(None, nlabels - 1)
epoch_idx = -1
# Reinitialize model average if needed
if (config['training']['take_model_average']
and config['training']['model_average_reinit']):
update_average(generator_test, generator, 0.)
# Learning rate anneling
g_scheduler = build_lr_scheduler(g_optimizer, config, last_epoch=it)
d_scheduler = build_lr_scheduler(d_optimizer, config, last_epoch=it)
# Trainer
trainer = Trainer(generator,
discriminator,
g_optimizer,
d_optimizer,
gan_type=config['training']['gan_type'],
reg_type=config['training']['reg_type'],
reg_param=config['training']['reg_param'])
# Training loop
print('Start training...')
save_dir = config['training']['out_dir'] + '/models/'
if not os.path.isdir(save_dir):
os.makedirs(save_dir)
get_parameter_number(generator)
get_parameter_number(discriminator)
inception_mean_all = []
inception_std_all = []
def main():
pp = pprint.PrettyPrinter(indent=1)
pp.pprint({
'data': config['data'],
'generator': config['generator'],
'discriminator': config['discriminator'],
'clusterer': config['clusterer'],
'training': config['training']
})
is_cuda = torch.cuda.is_available()
# Short hands
batch_size = config['training']['batch_size']
log_every = config['training']['log_every']
inception_every = config['training']['inception_every']
backup_every = config['training']['backup_every']
sample_nlabels = config['training']['sample_nlabels']
nlabels = config['data']['nlabels']
sample_nlabels = min(nlabels, sample_nlabels)
checkpoint_dir = path.join(out_dir, 'chkpts')
nepochs = args.nepochs
# Create missing directories
if not path.exists(out_dir):
os.makedirs(out_dir)
if not path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
# Logger
checkpoint_io = CheckpointIO(checkpoint_dir=checkpoint_dir)
device = torch.device("cuda:0" if is_cuda else "cpu")
train_dataset, _ = get_dataset(
name=config['data']['type'],
data_dir=config['data']['train_dir'],
size=config['data']['img_size'],
deterministic=config['data']['deterministic'])
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=batch_size,
num_workers=config['training']['nworkers'],
shuffle=True,
pin_memory=True,
sampler=None,
drop_last=True)
# Create models
generator, discriminator = build_models(config)
# Put models on gpu if needed
generator = generator.to(device)
discriminator = discriminator.to(device)
for name, module in discriminator.named_modules():
if isinstance(module, nn.Sigmoid):
print('Found sigmoid layer in discriminator; not compatible with BCE with logits')
exit()
g_optimizer, d_optimizer = build_optimizers(generator, discriminator, config)
devices = [int(x) for x in args.devices]
generator = nn.DataParallel(generator, device_ids=devices)
discriminator = nn.DataParallel(discriminator, device_ids=devices)
# Register modules to checkpoint
checkpoint_io.register_modules(generator=generator,
discriminator=discriminator,
g_optimizer=g_optimizer,
d_optimizer=d_optimizer)
# Logger
logger = Logger(log_dir=path.join(out_dir, 'logs'),
img_dir=path.join(out_dir, 'imgs'),
monitoring=config['training']['monitoring'],
monitoring_dir=path.join(out_dir, 'monitoring'))
# Distributions
ydist = get_ydist(nlabels, device=device)
zdist = get_zdist(config['z_dist']['type'], config['z_dist']['dim'], device=device)
ntest = config['training']['ntest']
x_test, y_test = utils.get_nsamples(train_loader, ntest)
x_cluster, y_cluster = utils.get_nsamples(train_loader, config['clusterer']['nimgs'])
x_test, y_test = x_test.to(device), y_test.to(device)
z_test = zdist.sample((ntest, ))
utils.save_images(x_test, path.join(out_dir, 'real.png'))
logger.add_imgs(x_test, 'gt', 0)
# Test generator
if config['training']['take_model_average']:
print('Taking model average')
bad_modules = [nn.BatchNorm1d, nn.BatchNorm2d, nn.BatchNorm3d]
for model in [generator, discriminator]:
for name, module in model.named_modules():
for bad_module in bad_modules:
if isinstance(module, bad_module):
print('Batch norm in discriminator not compatible with exponential moving average')
exit()
generator_test = copy.deepcopy(generator)
checkpoint_io.register_modules(generator_test=generator_test)
else:
generator_test = generator
clusterer = get_clusterer(config)(discriminator=discriminator,
x_cluster=x_cluster,
x_labels=y_cluster,
gt_nlabels=config['data']['nlabels'],
**config['clusterer']['kwargs'])
# Load checkpoint if it exists
it = utils.get_most_recent(checkpoint_dir, 'model') if args.model_it == -1 else args.model_it
it, epoch_idx, loaded_clusterer = checkpoint_io.load_models(it=it, load_samples='supervised' != config['clusterer']['name'])
if loaded_clusterer is None:
print('Initializing new clusterer. The first clustering can be quite slow.')
clusterer.recluster(discriminator=discriminator)
checkpoint_io.save_clusterer(clusterer, it=0)
np.savez(os.path.join(checkpoint_dir, 'cluster_samples.npz'), x=x_cluster)
else:
print('Using loaded clusterer')
clusterer = loaded_clusterer
# Evaluator
evaluator = Evaluator(
generator_test,
zdist,
ydist,
train_loader=train_loader,
clusterer=clusterer,
batch_size=batch_size,
device=device,
inception_nsamples=config['training']['inception_nsamples'])
# Trainer
trainer = Trainer(generator,
discriminator,
g_optimizer,
d_optimizer,
gan_type=config['training']['gan_type'],
reg_type=config['training']['reg_type'],
reg_param=config['training']['reg_param'])
# Training loop
print('Start training...')
while it < args.nepochs * len(train_loader):
epoch_idx += 1
for x_real, y in train_loader:
it += 1
x_real, y = x_real.to(device), y.to(device)
z = zdist.sample((batch_size, ))
y = clusterer.get_labels(x_real, y).to(device)
# Discriminator updates
dloss, reg = trainer.discriminator_trainstep(x_real, y, z)
logger.add('losses', 'discriminator', dloss, it=it)
logger.add('losses', 'regularizer', reg, it=it)
# Generators updates
gloss = trainer.generator_trainstep(y, z)
logger.add('losses', 'generator', gloss, it=it)
if config['training']['take_model_average']:
update_average(generator_test, generator, beta=config['training']['model_average_beta'])
# Print stats
if it % log_every == 0:
g_loss_last = logger.get_last('losses', 'generator')
d_loss_last = logger.get_last('losses', 'discriminator')
d_reg_last = logger.get_last('losses', 'regularizer')
print('[epoch %0d, it %4d] g_loss = %.4f, d_loss = %.4f, reg=%.4f'
% (epoch_idx, it, g_loss_last, d_loss_last, d_reg_last))
if it % config['training']['recluster_every'] == 0 and it > config['training']['burnin_time']:
# print cluster distribution for online methods
if it % 100 == 0 and config['training']['recluster_every'] <= 100:
print(f'[epoch {epoch_idx}, it {it}], distribution: {clusterer.get_label_distribution(x_real)}')
clusterer.recluster(discriminator=discriminator, x_batch=x_real)
# (i) Sample if necessary
if it % config['training']['sample_every'] == 0:
print('Creating samples...')
x = evaluator.create_samples(z_test, y_test)
x = evaluator.create_samples(z_test, clusterer.get_labels(x_test, y_test).to(device))
logger.add_imgs(x, 'all', it)
for y_inst in range(sample_nlabels):
x = evaluator.create_samples(z_test, y_inst)
logger.add_imgs(x, '%04d' % y_inst, it)
# (ii) Compute inception if necessary
if it % inception_every == 0 and it > 0:
print('PyTorch Inception score...')
inception_mean, inception_std = evaluator.compute_inception_score()
logger.add('metrics', 'pt_inception_mean', inception_mean, it=it)
logger.add('metrics', 'pt_inception_stddev', inception_std, it=it)
print(f'[epoch {epoch_idx}, it {it}] pt_inception_mean: {inception_mean}, pt_inception_stddev: {inception_std}')
# (iii) Backup if necessary
if it % backup_every == 0:
print('Saving backup...')
checkpoint_io.save('model_%08d.pt' % it, it=it)
checkpoint_io.save_clusterer(clusterer, int(it))
logger.save_stats('stats_%08d.p' % it)
if it > 0: checkpoint_io.save('model.pt', it=it)
def main(outdir):
for subdir in ['all', 'snapshots', 'clusters']:
if not os.path.exists(os.path.join(outdir, subdir)):
os.makedirs(os.path.join(outdir, subdir), exist_ok=True)
if data_type == 'grid':
get_data = inputs.get_data_grid
percent_good = evaluation.percent_good_grid
elif data_type == 'ring':
get_data = inputs.get_data_ring
percent_good = evaluation.percent_good_ring
else:
raise NotImplementedError()
zdist = distributions.Normal(torch.zeros(z_dim, device=device),
torch.ones(z_dim, device=device))
z_test = zdist.sample((test_batch_size, ))
x_test, y_test = get_test(get_data=get_data,
batch_size=test_batch_size,
variance=variance,
k_value=k_value,
device=device)
x_cluster, _ = get_test(get_data=get_data,
batch_size=10000,
variance=variance,
k_value=k_value,
device=device)
train_loader = get_dataset(get_data=get_data,
batch_size=train_batch_size,
npts=npts,
variance=variance,
k_value=k_value)
def train(trainer, g, d, clusterer, exp_dir):
it = 0
if os.path.exists(os.path.join(exp_dir, 'log.txt')):
os.remove(os.path.join(exp_dir, 'log.txt'))
for epoch in range(nepochs):
for x_real, y in train_loader:
z = zdist.sample((train_batch_size, ))
x_real, y = x_real.to(device), y.to(device)
y = clusterer.get_labels(x_real, y)
dloss, _ = trainer.discriminator_trainstep(x_real, y, z)
gloss = trainer.generator_trainstep(y, z)
if it % args.recluster_every == 0 and args.clusterer != 'supervised':
if args.clusterer != 'burnin' or it >= args.burnin_time:
clusterer.recluster(discriminator, x_batch=x_real)
if it % 1000 == 0:
x_fake = g(z_test, clusterer.get_labels(
x_test, y_test)).detach().cpu().numpy()
visualize_generated(x_fake,
x_test.detach().cpu().numpy(), y, it,
exp_dir)
visualize_clusters(x_test.detach().cpu().numpy(),
clusterer.get_labels(x_test, y_test),
it, exp_dir)
torch.save(
{
'generator': g.state_dict(),
'discriminator': d.state_dict(),
'g_optimizer': g_optimizer.state_dict(),
'd_optimizer': d_optimizer.state_dict()
},
os.path.join(exp_dir, 'snapshots', 'model_%d.pt' % it))
if it % 1000 == 0:
g.eval()
d.eval()
x_fake = g(z_test, clusterer.get_labels(
x_test, y_test)).detach().cpu().numpy()
percent, modes, kl = percent_good(x_fake, var=variance)
log_message = f'[epoch {epoch} it {it}] dloss = {dloss}, gloss = {gloss}, prop_real = {percent}, modes = {modes}, kl = {kl}'
with open(os.path.join(exp_dir, 'log.txt'), 'a+') as f:
f.write(log_message + '\n')
print(log_message)
it += 1
# train a G/D from scratch
generator, discriminator = get_models(args.model_type, 'conditional',
num_clusters, args.d_act_dim, device)
g_optimizer, d_optimizer = get_optimizers(generator, discriminator)
trainer = Trainer(generator,
discriminator,
g_optimizer,
d_optimizer,
gan_type='standard',
reg_type='none',
reg_param=0)
clusterer = clusterer_dict[args.clusterer](discriminator=discriminator,
k_value=num_clusters,
x_cluster=x_cluster)
clusterer.recluster(discriminator=discriminator)
train(trainer, generator, discriminator, clusterer, os.path.join(outdir))
epoch_idx = -1
# Reinitialize model average if needed
if (config['training']['take_model_average']
and config['training']['model_average_reinit']):
update_average(generator_test, generator, 0.)
# Learning rate anneling
g_scheduler = build_lr_scheduler(g_optimizer, config, last_epoch=it)
d_scheduler = build_lr_scheduler(d_optimizer, config, last_epoch=it)
# Trainer
trainer = Trainer(generator,
discriminator,
g_optimizer,
d_optimizer,
gan_type=config['training']['gan_type'],
reg_type=config['training']['reg_type'],
reg_param=config['training']['reg_param'],
D_fix_layer=config['discriminator']['layers'])
# Training loop
print('Start training...')
save_dir = config['training']['out_dir'] + '/models/'
if not os.path.isdir(save_dir):
os.makedirs(save_dir)
FLAG = 500
inception_mean_all = []
inception_std_all = []
fid_all = []
logger.load_stats('stats.p')
# Reinitialize model average if needed
if (config['training']['take_model_average']
and config['training']['model_average_reinit']):
update_average(generator_test, generator, 0.)
# Learning rate anneling
g_scheduler = build_lr_scheduler(g_optimizer, config, last_epoch=it)
d_scheduler = build_lr_scheduler(d_optimizer, config, last_epoch=it)
# Trainer
trainer = Trainer(dvae,
generator,
discriminator,
g_optimizer,
d_optimizer,
reg_param=config['training']['reg_param'],
w_info=config['training']['w_info'])
# Training loop
tqdm.write('Start training...')
pbar = tqdm(total=max_iter)
if it > 0:
pbar.update(it)
out = False
while not out:
epoch_idx += 1
tqdm.write('Start epoch %d...' % epoch_idx)
config['training']['reg_param'] = reg_param
# Reinitialize model average if needed
if (config['training']['take_model_average']
and config['training']['model_average_reinit']):
update_average(generator_test, generator, 0.)
# Learning rate anneling
g_scheduler = build_lr_scheduler(g_optimizer, config, last_epoch=it)
d_scheduler = build_lr_scheduler(d_optimizer, config, last_epoch=it)
# Trainer
trainer = Trainer(
generator, discriminator, g_optimizer, d_optimizer,
gan_type=config['training']['gan_type'],
reg_type=config['training']['reg_type'],
reg_param=config['training']['reg_param'],
adaptive_beta=adaptive_beta,
**config['training']['kwargs']
)
# Training loop
print('Start training...')
while True:
epoch_idx += 1
print('Start epoch %d...' % epoch_idx)
for x_real, y in train_loader:
it += 1
g_scheduler.step()
d_scheduler.step()
