def train(model, opt, train_loader, args, discriminators, writer):
model.train()
for disc in discriminators:
if discriminators[disc][
0] is not None and discriminators[disc][1] is None:
print(f"define an optimizer for discriminator:{disc}.. exiting",
file=sys.stderr)
sys.stderr.flush()
else:
discriminators[disc][0].train()
for idx, data in enumerate(train_loader):
# if there are any discriminators used for training then, graph may have to be retained
if len(discriminators) > 0:
retain_graph = True
else:
retain_graph = False
images = data[0]
images = images.to(args.device)
loss_dict, z = get_losses(model, images, args, discriminators)
# print(loss_dict, file=sys.stderr)
############1111 AE loss 11111############
opt.zero_grad()
loss_dict["recons_loss"].backward(retain_graph=retain_graph)
opt.step()
###########111111 ----------- 111111#########
# backprop all discriminators
for disc_idx, disc in enumerate(discriminators):
discriminators[disc][1].zero_grad()
if disc_idx >= len(discriminators) - 1:
retain_graph = False
#print(f"disc:{disc}\t loss:{loss_dict[f'{disc}_loss']}\t retain:{retain_graph}", file=sys.stderr)
loss_dict[f"{disc}_loss"].backward(retain_graph=retain_graph)
discriminators[disc][1].step()
# Logs
if idx % args.print_interval == 0:
print(
"iter:",
str(args.steps) + "/" + str(args.num_epochs *
(len(train_loader))), "iter loss:",
loss_dict["recons_loss"].item())
for disc in discriminators:
print(f"{disc} loss:{loss_dict[f'{disc}_loss']}")
sys.stderr.flush()
sys.stderr.flush()
train_util.log_losses("train", loss_dict, args.steps, writer)
train_util.log_latent_metrics("train", z, args.steps, writer)
def train(epoch, data_loader, model, optimizer, args, writer, discriminators):
for disc in discriminators:
if discriminators[disc][
0] is not None and discriminators[disc][1] is None:
print(f"define an optimizer for discriminator:{disc}.. exiting",
file=sys.stderr)
else:
discriminators[disc][0].train()
torch.cuda.empty_cache()
for idx, data in enumerate(data_loader):
# if there are any discriminators used for training then, graph may have to be retained
if len(discriminators) > 0:
retain_graph = True
else:
retain_graph = False
images = data[0]
images = images.to(args.device)
loss_dict, z = get_losses(model, images, args, discriminators)
optimizer.zero_grad()
loss_dict["recons_loss"].backward(retain_graph=retain_graph)
optimizer.step()
# backprop all discriminators
for disc_idx, disc in enumerate(discriminators):
discriminators[disc][1].zero_grad()
if disc_idx >= len(discriminators) - 1:
retain_graph = False
loss_dict[f"{disc}_loss"].backward(retain_graph=retain_graph)
discriminators[disc][1].step()
# Logs
if idx % 1000 == 0:
print(f"iter:{args.steps}\trecons loss:{loss_dict['recons_loss']}",
file=sys.stderr)
for disc in discriminators:
print(f"{disc} loss:{loss_dict[f'{disc}_loss']}",
file=sys.stderr)
sys.stderr.flush()
train_util.log_losses("train", loss_dict, args.steps, writer)
train_util.log_latent_metrics("train", z, args.steps, writer)
if idx == 0:
print(torch.cuda.max_memory_allocated(torch.device("cuda:0")),
file=sys.stderr)
#print(torch.cuda.max_memory_allocated(torch.device("cuda:1")), file=sys.stderr)
sys.stderr.flush()
args.steps += 1
def val_test(args):
writer = SummaryWriter('./logs/{0}'.format(args.output_folder))
save_filename = './models/{0}'.format(args.output_folder)
train_loader, valid_loader, test_loader = train_util.get_dataloaders(args)
recons_input_img = train_util.log_input_img_grid(test_loader, writer)
input_dim = 3
model = VectorQuantizedVAE(input_dim, args.hidden_size, args.k,
args.enc_type, args.dec_type)
# if torch.cuda.device_count() > 1 and args.device == "cuda":
# model = torch.nn.DataParallel(model)
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
discriminators = {}
if args.recons_loss == "gan":
recons_disc = Discriminator(input_dim, args.img_res,
args.input_type).to(args.device)
recons_disc_opt = torch.optim.Adam(recons_disc.parameters(),
lr=args.disc_lr,
amsgrad=True)
discriminators["recons_disc"] = [recons_disc, recons_disc_opt]
model.to(args.device)
for disc in discriminators:
discriminators[disc][0].to(args.device)
if args.weights == "load":
start_epoch = train_util.load_state(save_filename, model, optimizer,
discriminators)
else:
start_epoch = 0
stop_patience = args.stop_patience
best_loss = torch.tensor(np.inf)
for epoch in tqdm(range(start_epoch, 4), file=sys.stdout):
val_loss_dict, z = train_util.test(get_losses, model, valid_loader,
args, discriminators, True)
# if args.weights == "init" and epoch==1:
# epoch+=1
# break
train_util.log_recons_img_grid(recons_input_img, model, epoch + 1,
args.device, writer)
train_util.log_interp_img_grid(recons_input_img, model, epoch + 1,
args.device, writer)
train_util.log_losses("val", val_loss_dict, epoch + 1, writer)
train_util.log_latent_metrics("val", z, epoch + 1, writer)
train_util.save_state(model, optimizer, discriminators,
val_loss_dict["recons_loss"], best_loss,
args.recons_loss, epoch, save_filename)
print(val_loss_dict)
def main(args):
writer = SummaryWriter('./logs/{0}'.format(args.output_folder))
save_filename = './models/{0}'.format(args.output_folder)
train_loader, val_loader, test_loader = train_util.get_dataloaders(args)
recons_input_img = train_util.log_input_img_grid(test_loader, writer)
input_dim = 3
model = ACAI(args.img_res, input_dim, args.hidden_size, args.enc_type,
args.dec_type).to(args.device)
disc = Discriminator(input_dim, args.img_res,
args.input_type).to(args.device)
disc_opt = torch.optim.Adam(disc.parameters(),
lr=args.disc_lr,
amsgrad=True)
# if torch.cuda.device_count() > 1 and args.device == "cuda":
# model = torch.nn.DataParallel(model)
opt = torch.optim.Adam(model.parameters(), lr=args.lr)
# ae_lr_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, "min", patience=args.lr_patience, factor=0.5,
# threshold=args.threshold, threshold_mode="abs", min_lr=1e-7)
# interp_disc_lr_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(disc_opt, "min", patience=args.lr_patience, factor=0.5,
# threshold=args.threshold, threshold_mode="abs", min_lr=1e-7)
discriminators = {"interp_disc": [disc, disc_opt]}
if args.recons_loss != "mse":
if args.recons_loss == "gan":
recons_disc = Discriminator(input_dim, args.img_res,
args.input_type).to(args.device)
elif args.recons_loss == "comp":
recons_disc = AnchorComparator(input_dim * 2, args.img_res,
args.input_type).to(args.device)
elif "comp_2" in args.recons_loss:
recons_disc = ClubbedPermutationComparator(
input_dim * 2, args.img_res, args.input_type).to(args.device)
elif "comp_6" in args.recons_loss:
recons_disc = FullPermutationComparator(
input_dim * 2, args.img_res, args.input_type).to(args.device)
recons_disc_opt = torch.optim.Adam(recons_disc.parameters(),
lr=args.disc_lr,
amsgrad=True)
discriminators["recons_disc"] = [recons_disc, recons_disc_opt]
# Generate the samples first once
train_util.save_recons_img_grid("test", recons_input_img, model, 0, args)
if args.weights == "load":
start_epoch = train_util.load_state(save_filename, model, opt,
discriminators)
else:
start_epoch = 0
best_loss = torch.tensor(np.inf)
for epoch in range(args.num_epochs):
print("Epoch {}:".format(epoch))
train(model, opt, train_loader, args, discriminators, writer)
# curr_loss = val(model, val_loader)
# print(f"epoch val loss:{curr_loss}")
val_loss_dict, z = train_util.test(get_losses, model, val_loader, args,
discriminators)
train_util.log_losses("val", val_loss_dict, epoch + 1, writer)
train_util.log_latent_metrics("val", z, epoch + 1, writer)
# train_util.log_recons_img_grid(recons_input_img, model, epoch+1, args.device, writer)
# train_util.log_interp_img_grid(recons_input_img, model, epoch+1, args.device, writer)
train_util.save_recons_img_grid("val", recons_input_img, model,
epoch + 1, args)
train_util.save_interp_img_grid("val", recons_input_img, model,
epoch + 1, args)
train_util.save_state(model, opt, discriminators,
val_loss_dict["recons_loss"], best_loss,
args.recons_loss, epoch, save_filename)
def main(args):
writer = SummaryWriter('./logs/{0}'.format(args.output_folder))
save_filename = './models/{0}'.format(args.output_folder)
train_loader, valid_loader, test_loader = train_util.get_dataloaders(args)
num_channels = 3
model = VectorQuantizedVAE(num_channels, args.hidden_size, args.k,
args.enc_type, args.dec_type)
model.to(args.device)
# Fixed images for Tensorboard
recons_input_img = train_util.log_input_img_grid(test_loader, writer)
train_util.log_recons_img_grid(recons_input_img, model, 0, args.device,
writer)
discriminators = {}
input_dim = 3
if args.recons_loss != "mse":
if args.recons_loss == "gan":
recons_disc = Discriminator(input_dim, args.img_res,
args.input_type).to(args.device)
elif args.recons_loss == "comp":
recons_disc = AnchorComparator(input_dim * 2, args.img_res,
args.input_type).to(args.device)
elif "comp_2" in args.recons_loss:
recons_disc = ClubbedPermutationComparator(
input_dim * 2, args.img_res, args.input_type).to(args.device)
elif "comp_6" in args.recons_loss:
recons_disc = FullPermutationComparator(
input_dim * 2, args.img_res, args.input_type).to(args.device)
recons_disc_opt = torch.optim.Adam(recons_disc.parameters(),
lr=args.disc_lr,
amsgrad=True)
recons_disc_lr_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
recons_disc_opt,
"min",
patience=args.lr_patience,
factor=0.5,
threshold=args.threshold,
threshold_mode="abs",
min_lr=1e-7)
discriminators["recons_disc"] = [recons_disc, recons_disc_opt]
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
ae_lr_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer,
"min",
patience=args.lr_patience,
factor=0.5,
threshold=args.threshold,
threshold_mode="abs",
min_lr=1e-7)
if torch.cuda.device_count() > 1:
model = train_util.ae_data_parallel(model)
for disc in discriminators:
discriminators[disc][0] = torch.nn.DataParallel(
discriminators[disc][0])
model.to(args.device)
for disc in discriminators:
discriminators[disc][0].to(args.device)
# Generate the samples first once
recons_input_img = train_util.log_input_img_grid(test_loader, writer)
train_util.log_recons_img_grid(recons_input_img, model, 0, args.device,
writer)
if args.weights == "load":
start_epoch = train_util.load_state(save_filename, model, optimizer,
discriminators)
else:
start_epoch = 0
stop_patience = args.stop_patience
best_loss = torch.tensor(np.inf)
for epoch in tqdm(range(start_epoch, args.num_epochs), file=sys.stdout):
try:
train(epoch, train_loader, model, optimizer, args, writer,
discriminators)
except RuntimeError as err:
print("".join(
traceback.TracebackException.from_exception(err).format()),
file=sys.stderr)
print("*******")
print(err, file=sys.stderr)
print(f"batch_size:{args.batch_size}", file=sys.stderr)
exit(0)
val_loss_dict, z = train_util.test(get_losses, model, valid_loader,
args, discriminators)
train_util.log_recons_img_grid(recons_input_img, model, epoch + 1,
args.device, writer)
train_util.log_interp_img_grid(recons_input_img, model, epoch + 1,
args.device, writer)
train_util.log_losses("val", val_loss_dict, epoch + 1, writer)
train_util.log_latent_metrics("val", z, epoch + 1, writer)
train_util.save_state(model, optimizer, discriminators,
val_loss_dict["recons_loss"], best_loss,
args.recons_loss, epoch, save_filename)
# early stop check
# if val_loss_dict["recons_loss"] - best_loss < args.threshold:
# stop_patience -= 1
# else:
# stop_patience = args.stop_patience
# if stop_patience == 0:
# print("training early stopped!")
# break
ae_lr_scheduler.step(val_loss_dict["recons_loss"])
if args.recons_loss != "mse":
recons_disc_lr_scheduler.step(val_loss_dict["recons_disc_loss"])
def train(epoch, data_loader, model, optimizer, args, writer, discriminators):
# use a dict or add interp dict and optim separately
model.train()
for disc in discriminators:
if discriminators[disc][
0] is not None and discriminators[disc][1] is None:
print(f"define an optimizer for discriminator:{disc}.. exiting",
file=sys.stderr)
sys.stderr.flush()
else:
discriminators[disc][0].train()
for idx, data in enumerate(data_loader):
retain_graph = True
# print(f"shape:{data.shape}")
loss_dict, z = get_losses(model, data, args, discriminators)
############1111 AE loss 11111############
optimizer.zero_grad()
loss_dict["recons_loss"].backward(retain_graph=retain_graph)
optimizer.step()
###########111111 ----------- 111111#########
# backprop all discriminators
for disc_idx, disc in enumerate(discriminators):
discriminators[disc][1].zero_grad()
if disc_idx >= len(discriminators) - 1:
retain_graph = False
#print(f"disc:{disc}\t loss:{loss_dict[f'{disc}_loss']}\t retain:{retain_graph}", file=sys.stderr)
loss_dict[f"{disc}_loss"].backward(retain_graph=retain_graph)
discriminators[disc][1].step()
# Logs
if idx % args.print_interval == 0:
print("iter:",
str(args.steps) + "/" + str(args.num_epochs *
(len(data_loader))),
"iter loss:",
loss_dict["recons_loss"].item(),
file=sys.stderr)
print(f"memory:\n{subprocess.run('nvidia-smi')}")
for disc in discriminators:
print(f"{disc} loss:{loss_dict[f'{disc}_loss']}",
file=sys.stderr)
sys.stderr.flush()
sys.stderr.flush()
train_util.log_losses("train", loss_dict, args.steps, writer)
train_util.log_latent_metrics("train", z, args.steps, writer)
if idx == 0:
print(torch.cuda.max_memory_allocated(torch.device(args.device)),
file=sys.stderr)
sys.stderr.flush()
args.steps += 1
def main(args):
input_dim = 3
model = VAE(input_dim, args.hidden_size, args.enc_type, args.dec_type)
opt = torch.optim.Adam(model.parameters(), lr=args.lr, amsgrad=True)
# scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(opt, "min", patience=args.lr_patience, factor=0.5,
# threshold=args.threshold, threshold_mode="abs", min_lr=1e-6)
# ae_lr_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(opt, "min", patience=args.lr_patience, factor=0.5,
# threshold=args.threshold, threshold_mode="abs", min_lr=1e-7)
discriminators = {}
if args.recons_loss != "mse":
if args.recons_loss == "gan":
recons_disc = Discriminator(input_dim, args.img_res,
args.input_type).to(args.device)
elif args.recons_loss == "comp":
recons_disc = AnchorComparator(input_dim * 2, args.img_res,
args.input_type).to(args.device)
elif "comp_2" in args.recons_loss:
recons_disc = ClubbedPermutationComparator(
input_dim * 2, args.img_res, args.input_type).to(args.device)
elif "comp_6" in args.recons_loss:
recons_disc = FullPermutationComparator(
input_dim * 2, args.img_res, args.input_type).to(args.device)
recons_disc_opt = torch.optim.Adam(recons_disc.parameters(),
lr=args.disc_lr,
amsgrad=True)
discriminators["recons_disc"] = [recons_disc, recons_disc_opt]
if torch.cuda.device_count() > 1:
model = train_util.ae_data_parallel(model)
for disc in discriminators:
discriminators[disc][0] = torch.nn.DataParallel(
discriminators[disc][0])
model.to(args.device)
for disc in discriminators:
discriminators[disc][0].to(args.device)
print("model built", file=sys.stderr)
#print("model created")
train_loader, val_loader, test_loader = train_util.get_dataloaders(args)
print("loaders acquired", file=sys.stderr)
#print("loaders acquired")
model_name = f"vae_{args.recons_loss}"
if args.output_folder is None:
args.output_folder = os.path.join(
model_name, args.dataset,
f"depth_{args.enc_type}_{args.dec_type}_hs_{args.img_res}_{args.hidden_size}"
)
log_save_path = os.path.join("./logs", args.output_folder)
model_save_path = os.path.join("./models", args.output_folder)
if not os.path.exists(log_save_path):
os.makedirs(log_save_path)
print(f"log:{log_save_path}", file=sys.stderr)
sys.stderr.flush()
if not os.path.exists(model_save_path):
os.makedirs(model_save_path)
writer = SummaryWriter(log_save_path)
print(f"train loader length:{len(train_loader)}", file=sys.stderr)
best_loss = torch.tensor(np.inf)
if args.weights == "load":
start_epoch = train_util.load_state(model_save_path, model, opt,
discriminators)
else:
start_epoch = 0
recons_input_img = train_util.log_input_img_grid(test_loader, writer)
train_util.log_recons_img_grid(recons_input_img, model, 0, args.device,
writer)
stop_patience = args.stop_patience
for epoch in range(start_epoch, args.num_epochs):
try:
train(model, train_loader, opt, epoch, writer, args,
discriminators)
except RuntimeError as err:
print("".join(
traceback.TracebackException.from_exception(err).format()),
file=sys.stderr)
print("*******", file=sys.stderr)
print(err, file=sys.stderr)
exit(0)
val_loss_dict, z = train_util.test(get_losses, model, val_loader, args,
discriminators)
print(f"epoch loss:{val_loss_dict['recons_loss'].item()}")
train_util.save_recons_img_grid("test", recons_input_img, model,
epoch + 1, args)
train_util.save_interp_img_grid("test", recons_input_img, model,
epoch + 1, args)
train_util.log_losses("val", val_loss_dict, epoch + 1, writer)
train_util.log_latent_metrics("val", z, epoch + 1, writer)
train_util.save_state(model, opt, discriminators,
val_loss_dict["recons_loss"], best_loss,
args.recons_loss, epoch, model_save_path)