def load_vgg_from_local(arch='vgg19',
cfg='E',
batch_norm=False,
pretrained=True,
vgg_dir=None,
parallel=True,
requires_grad=False,
**kwargs):
vgg = vgglib.VGG(vgglib.make_layers(cfgs[cfg], batch_norm=batch_norm),
**kwargs)
vgg.load_state_dict(
model_zoo.load_url(url=VGG_URL,
model_dir='/gpub/temp/imagenet2012/hdf5'))
vgg = (vgg.eval()).cuda()
if parallel:
print("Parallel VGG model...")
vgg = torch.nn.DataParallel(vgg)
toggle_grad(vgg, requires_grad)
return vgg
def run(config):
# Update the config dict as necessary
# This is for convenience, to add settings derived from the user-specified
# configuration into the config-dict (e.g. inferring the number of classes
# and size of the images from the dataset, passing in a pytorch object
# for the activation specified as a string)
config['resolution'] = utils.imsize_dict[config['dataset']]
config['n_classes'] = utils.nclass_dict[config['dataset']]
config['G_activation'] = utils.activation_dict[config['G_nl']]
config['D_activation'] = utils.activation_dict[config['D_nl']]
# By default, skip init if resuming training.
if config['resume']:
print('Skipping initialization for training resumption...')
config['skip_init'] = True
config = vae_utils.update_config_roots(config)
device = 'cuda'
# Seed RNG
utils.seed_rng(config['seed'])
# Prepare root folders if necessary
utils.prepare_root(config)
# Setup cudnn.benchmark for free speed
torch.backends.cudnn.benchmark = True
experiment_name = (config['experiment_name'] if config['experiment_name']
else utils.name_from_config(config))
print('Experiment name is %s' % experiment_name)
# Next, build the model
G = Generator(**{**config, 'skip_init': True, 'no_optim': True}).to(device)
print('Loading pretrained G for dir %s ...' % config['pretrained_G_dir'])
pretrained_dict = torch.load(config['pretrained_G_dir'])
G_dict = G.state_dict()
pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in G_dict}
G_dict.update(pretrained_dict)
G.load_state_dict(G_dict)
E = Encoder(**config).to(device)
utils.toggle_grad(G, False)
utils.toggle_grad(E, True)
class G_E(nn.Module):
def __init__(self):
super(G_E, self).__init__()
self.G = G
self.E = E
def forward(self, w, y):
with torch.no_grad():
net = self.G(w, self.G.shared(y))
net = self.E(net)
return net
GE = G_E()
# If using EMA, prepare it
if config['ema']:
print('Preparing EMA for E with decay of {}'.format(
config['ema_decay']))
E_ema = Encoder(**{
**config, 'skip_init': True,
'no_optim': True
}).to(device)
e_ema = utils.ema(E, E_ema, config['ema_decay'], config['ema_start'])
else:
E_ema, e_ema = None, None
print(G)
print(E)
print('Number of params in G: {} E: {}'.format(
*
[sum([p.data.nelement() for p in net.parameters()])
for net in [G, E]]))
# Prepare state dict, which holds things like epoch # and itr #
state_dict = {
'itr': 0,
'epoch': 0,
'save_num': 0,
'save_best_num': 0,
'best_IS': 0,
'best_FID': 999999,
'config': config
}
# If loading from a pre-trained model, load weights
if config['resume']:
print('Loading weights...')
vae_utils.load_weights(
[E], state_dict, config['weights_root'], experiment_name,
config['load_weights'] if config['load_weights'] else None,
[e_ema] if config['ema'] else None)
# If parallel, parallelize the GD module
if config['parallel']:
GE = nn.DataParallel(GE)
if config['cross_replica']:
patch_replication_callback(GE)
# Prepare loggers for stats; metrics holds test metrics,
# lmetrics holds any desired training metrics.
train_metrics_fname = '%s/%s' % (config['logs_root'], experiment_name)
print('Training Metrics will be saved to {}'.format(train_metrics_fname))
train_log = utils.MyLogger(train_metrics_fname,
reinitialize=(not config['resume']),
logstyle=config['logstyle'])
# Write metadata
utils.write_metadata(config['logs_root'], experiment_name, config,
state_dict)
G_batch_size = max(config['G_batch_size'], config['batch_size'])
z_, y_ = utils.prepare_z_y(G_batch_size,
G.dim_z,
config['n_classes'],
device=device,
fp16=config['G_fp16'])
def train():
E.optim.zero_grad()
z_.sample_()
y_.sample_()
net = GE(z_[:config['batch_size']], y_[:config['batch_size']])
loss = F.l1_loss(z_[:config['batch_size']], net)
loss.backward()
if config["E_ortho"] > 0.0:
print('using modified ortho reg in E')
utils.ortho(E, config['E_ortho'])
E.optim.step()
out = {'loss': float(loss.item())}
return out
print('Beginning training at epoch %d...' % state_dict['epoch'])
# Train for specified number of epochs, although we mostly track G iterations.
for epoch in range(state_dict['epoch'], config['num_epochs']):
for i in range(100000):
# Increment the iteration counter
state_dict['itr'] += 1
# Make sure G and D are in training mode, just in case they got set to eval
# For D, which typically doesn't have BN, this shouldn't matter much.
G.train()
E.train()
if config['ema']:
E_ema.train()
metrics = train()
train_log.log(itr=int(state_dict['itr']), **metrics)
# Every sv_log_interval, log singular values
if (config['sv_log_interval'] > 0) and (
not (state_dict['itr'] % config['sv_log_interval'])):
train_log.log(itr=int(state_dict['itr']),
**{
**utils.get_SVs(G, 'G'),
**utils.get_SVs(E, 'E')
})
# If using my progbar, print metrics.
if config['pbar'] == 'mine':
print(', '.join(
['itr: %d' % state_dict['itr']] +
['%s : %+4.3f' % (key, metrics[key]) for key in metrics]),
end=' ')
# Save weights and copies as configured at specified interval
if not (state_dict['itr'] % config['save_every']):
vae_utils.save_weights([E], state_dict, config['weights_root'],
experiment_name,
'copy%d' % state_dict['save_num'],
[E_ema if config['ema'] else None])
state_dict['save_num'] = (state_dict['save_num'] +
1) % config['num_save_copies']
# Increment epoch counter at end of epoch
state_dict['epoch'] += 1
def train(x, y):
G.optim.zero_grad()
D.optim.zero_grad()
# How many chunks to split x and y into?
x = torch.split(x, config['batch_size'])
y = torch.split(y, config['batch_size'])
counter = 0
# Optionally toggle D and G's "require_grad"
if config['toggle_grads']:
utils.toggle_grad(D, True)
utils.toggle_grad(G, False)
for step_index in range(config['num_D_steps']):
# If accumulating gradients, loop multiple times before an optimizer step
D.optim.zero_grad()
for accumulation_index in range(config['num_D_accumulations']):
z_.sample_()
D_fake, D_real = GD(z_[:config['batch_size']], y_[:config['batch_size']],
x[counter], y[counter], train_G=False,
split_D=config['split_D'])
# Compute components of D's loss, average them, and divide by
# the number of gradient accumulations
D_loss_real, D_loss_fake = losses.discriminator_loss(D_fake, D_real)
D_loss = (D_loss_real + D_loss_fake) / float(config['num_D_accumulations'])
D_loss.backward()
counter += 1
# Optionally apply ortho reg in D
if config['D_ortho'] > 0.0:
# Debug print to indicate we're using ortho reg in D.
print('using modified ortho reg in D')
utils.ortho(D, config['D_ortho'])
D.optim.step()
# Optionally toggle "requires_grad"
if config['toggle_grads']:
utils.toggle_grad(D, False)
utils.toggle_grad(G, True)
# Zero G's gradients by default before training G, for safety
G.optim.zero_grad()
# If accumulating gradients, loop multiple times
for accumulation_index in range(config['num_G_accumulations']):
z_.sample_()
D_fake = GD(z_, y_, train_G=True, split_D=config['split_D'])
G_loss = losses.generator_loss(D_fake) / float(config['num_G_accumulations'])
G_loss.backward()
# Optionally apply modified ortho reg in G
if config['G_ortho'] > 0.0:
print('using modified ortho reg in G') # Debug print to indicate we're using ortho reg in G
# Don't ortho reg shared, it makes no sense. Really we should blacklist any embeddings for this
utils.ortho(G, config['G_ortho'],
blacklist=[param for param in G.shared.parameters()])
G.optim.step()
# If we have an ema, update it, regardless of if we test with it or not
if config['ema']:
ema.update(state_dict['itr'])
out = {'G_loss': float(G_loss.item()),
'D_loss_real': float(D_loss_real.item()),
'D_loss_fake': float(D_loss_fake.item())}
# Return G's loss and the components of D's loss.
return out
def train(x):
G.optim.zero_grad()
D.optim.zero_grad()
I.optim.zero_grad()
E.optim.zero_grad()
L.optim.zero_grad()
# How many chunks to split x and y into?
x = torch.split(x, config['batch_size'])
counter = 0
# Optionally toggle D and G's "require_grad"
if config['toggle_grads']:
utils.toggle_grad(D, True)
utils.toggle_grad(L, True)
utils.toggle_grad(G, False)
utils.toggle_grad(I, False)
utils.toggle_grad(E, False)
for step_index in range(config['num_D_steps']):
# If accumulating gradients, loop multiple times before an optimizer step
D.optim.zero_grad()
L.optim.zero_grad()
for accumulation_index in range(config['num_D_accumulations']):
z_.sample_()
y_.sample_()
ey_.sample_()
D_fake, D_real, D_inv, D_en, _, _ = Decoder(
z_[:config['batch_size']],
y_[:config['batch_size']],
x[counter],
ey_[:config['batch_size']],
train_G=False,
split_D=config['split_D'])
# Compute components of D's loss, average them, and divide by
# the number of gradient accumulations
D_loss_real, D_loss_fake = losses.discriminator_loss(
D_fake, D_real)
Latent_loss = losses.latent_loss_dis(D_inv, D_en)
D_loss = (D_loss_real + D_loss_fake + Latent_loss) / float(
config['num_D_accumulations'])
D_loss.backward()
counter += 1
# Optionally apply ortho reg in D
if config['D_ortho'] > 0.0:
# Debug print to indicate we're using ortho reg in D.
print('using modified ortho reg in D and Latent_Binder')
utils.ortho(D, config['D_ortho'])
utils.ortho(L, config['L_ortho'])
D.optim.step()
L.optim.step()
# Optionally toggle "requires_grad"
if config['toggle_grads']:
utils.toggle_grad(D, False)
utils.toggle_grad(L, False)
utils.toggle_grad(G, True)
utils.toggle_grad(I, True)
utils.toggle_grad(E, True)
# Zero G's gradients by default before training G, for safety
G.optim.zero_grad()
I.optim.zero_grad()
E.optim.zero_grad()
counter = 0
# If accumulating gradients, loop multiple times
for accumulation_index in range(config['num_G_accumulations']):
z_.sample_()
y_.sample_()
ey_.sample_()
D_fake, _, D_inv, D_en, G_en, reals = Decoder(
z_,
y_,
x[counter],
ey_,
train_G=True,
split_D=config['split_D'])
G_loss_fake = losses.generator_loss(
D_fake) * config['adv_loss_scale']
Latent_loss = losses.latent_loss_gen(D_inv, D_en)
Recon_loss = losses.recon_loss(G_en, reals)
G_loss = (G_loss_fake + Latent_loss + Recon_loss) / float(
config['num_G_accumulations'])
G_loss.backward()
counter += 1
# Optionally apply modified ortho reg in G
if config['G_ortho'] > 0.0:
print('using modified ortho reg in G, Invert, and Encoder')
# Debug print to indicate we're using ortho reg in G
# Don't ortho reg shared, it makes no sense. Really we should blacklist any embeddings for this
utils.ortho(G,
config['G_ortho'],
blacklist=[param for param in G.shared.parameters()])
utils.ortho(E, config['E_ortho'])
utils.ortho(I, config['I_ortho'])
G.optim.step()
I.optim.step()
E.optim.step()
# If we have an ema, update it, regardless of if we test with it or not
if config['ema']:
for ema in ema_list:
ema.update(state_dict['itr'])
out = {
'G_loss': float(G_loss.item()),
'D_loss_real': float(D_loss_real.item()),
'D_loss_fake': float(D_loss_fake.item()),
'Latent_loss': float(Latent_loss.item()),
'Recon_loss': float(Recon_loss.item())
}
# Release GPU memory:
del G_loss, D_loss_real, D_loss_fake, Latent_loss, Recon_loss
del D_fake, D_real, D_inv, D_en, G_en, reals
del x
# Return G's loss and the components of D's loss.
return out