def train(cfg: Namespace) -> None:
assert cfg.device == "cpu" or (cfg.device == "cuda"
and T.cuda.is_available())
root_dir = Path(__file__).resolve().parents[1]
logger.info("training: experiment %s" % (cfg.exp_name))
# make dir-tree
exp_dir = root_dir / "experiments" / cfg.exp_name
for d in ["out", "checkpoint", "logs"]:
os.makedirs(exp_dir / d, exist_ok=True)
cfg.to_file(exp_dir / "train_config.json")
# tb tb_writer
tb_writer = SummaryWriter(exp_dir / "logs")
logger.info("started tensorboard writer")
model = CAE()
model.train()
if cfg.device == "cuda":
model.cuda()
logger.info(f"loaded model on {cfg.device}")
dataloader = DataLoader(
dataset=ImageFolder720p(cfg.dataset_path),
batch_size=cfg.batch_size,
shuffle=cfg.shuffle,
num_workers=cfg.num_workers,
)
logger.info(f"loaded dataset from {cfg.dataset_path}")
optimizer = optim.Adam(model.parameters(),
lr=cfg.learning_rate,
weight_decay=1e-5)
loss_criterion = nn.MSELoss()
avg_loss, epoch_avg = 0.0, 0.0
ts = 0
# EPOCHS
for epoch_idx in range(cfg.start_epoch, cfg.num_epochs + 1):
# BATCHES
for batch_idx, data in enumerate(dataloader, start=1):
img, patches, _ = data
if cfg.device == "cuda":
patches = patches.cuda()
avg_loss_per_image = 0.0
for i in range(6):
for j in range(10):
optimizer.zero_grad()
x = patches[:, :, i, j, :, :]
y = model(x)
loss = loss_criterion(y, x)
avg_loss_per_image += (1 / 60) * loss.item()
loss.backward()
optimizer.step()
avg_loss += avg_loss_per_image
epoch_avg += avg_loss_per_image
if batch_idx % cfg.batch_every == 0:
tb_writer.add_scalar("train/avg_loss",
avg_loss / cfg.batch_every, ts)
for name, param in model.named_parameters():
tb_writer.add_histogram(name, param, ts)
logger.debug("[%3d/%3d][%5d/%5d] avg_loss: %.8f" % (
epoch_idx,
cfg.num_epochs,
batch_idx,
len(dataloader),
avg_loss / cfg.batch_every,
))
avg_loss = 0.0
ts += 1
# -- end batch every
if batch_idx % cfg.save_every == 0:
out = T.zeros(6, 10, 3, 128, 128)
for i in range(6):
for j in range(10):
x = patches[0, :, i, j, :, :].unsqueeze(0).cuda()
out[i, j] = model(x).cpu().data
out = np.transpose(out, (0, 3, 1, 4, 2))
out = np.reshape(out, (768, 1280, 3))
out = np.transpose(out, (2, 0, 1))
y = T.cat((img[0], out), dim=2).unsqueeze(0)
save_imgs(
imgs=y,
to_size=(3, 768, 2 * 1280),
name=exp_dir / f"out/{epoch_idx}_{batch_idx}.png",
)
# -- end save every
# -- end batches
if epoch_idx % cfg.epoch_every == 0:
epoch_avg /= len(dataloader) * cfg.epoch_every
tb_writer.add_scalar(
"train/epoch_avg_loss",
avg_loss / cfg.batch_every,
epoch_idx // cfg.epoch_every,
)
logger.info("Epoch avg = %.8f" % epoch_avg)
epoch_avg = 0.0
T.save(model.state_dict(),
exp_dir / f"checkpoint/model_{epoch_idx}.pth")
# -- end epoch every
# -- end epoch
# save final model
T.save(model.state_dict(), exp_dir / "model_final.pth")
# cleaning
tb_writer.close()
def train(cfg: Namespace) -> None:
print(cfg.device)
assert cfg.device == 'cpu' or (cfg.device == 'cuda'
and T.cuda.is_available())
logger.info('training: experiment %s' % (cfg.exp_name))
# make dir-tree
exp_dir = ROOT_DIR / 'experiments' / cfg.exp_name
for d in ['out', 'checkpoint', 'logs']:
os.makedirs(exp_dir / d, exist_ok=True)
cfg.to_file(exp_dir / 'train_config.txt')
# tb writer
writer = SummaryWriter(exp_dir / 'logs')
model = CAE()
model.train()
if cfg.device == 'cuda':
model.cuda()
logger.info(f'loaded model on {cfg.device}')
dataset = ImageFolder720p(cfg.dataset_path)
dataloader = DataLoader(dataset,
batch_size=cfg.batch_size,
shuffle=cfg.shuffle,
num_workers=cfg.num_workers)
logger.info('loaded dataset')
optimizer = optim.Adam(model.parameters(),
lr=cfg.learning_rate,
weight_decay=1e-5)
loss_criterion = nn.MSELoss()
avg_loss, epoch_avg = 0.0, 0.0
ts = 0
# EPOCHS
for epoch_idx in range(cfg.start_epoch, cfg.num_epochs + 1):
# BATCHES
for batch_idx, data in enumerate(dataloader, start=1):
img, patches, _ = data
if cfg.device == 'cuda':
patches = patches.cuda()
avg_loss_per_image = 0.0
for i in range(6):
for j in range(10):
optimizer.zero_grad()
x = patches[:, :, i, j, :, :]
y = model(x)
loss = loss_criterion(y, x)
avg_loss_per_image += (1 / 60) * loss.item()
loss.backward()
optimizer.step()
avg_loss += avg_loss_per_image
epoch_avg += avg_loss_per_image
if batch_idx % cfg.batch_every == 0:
writer.add_scalar('train/avg_loss', avg_loss / cfg.batch_every,
ts)
for name, param in model.named_parameters():
writer.add_histogram(name, param, ts)
logger.debug('[%3d/%3d][%5d/%5d] avg_loss: %.8f' %
(epoch_idx, cfg.num_epochs, batch_idx,
len(dataloader), avg_loss / cfg.batch_every))
avg_loss = 0.0
ts += 1
# -- end batch every
if batch_idx % cfg.save_every == 0:
out = T.zeros(6, 10, 3, 128, 128)
for i in range(6):
for j in range(10):
x = patches[0, :, i, j, :, :].unsqueeze(0).cuda()
out[i, j] = model(x).cpu().data
out = np.transpose(out, (0, 3, 1, 4, 2))
out = np.reshape(out, (768, 1280, 3))
out = np.transpose(out, (2, 0, 1))
y = T.cat((img[0], out), dim=2).unsqueeze(0)
save_imgs(imgs=y,
to_size=(3, 768, 2 * 1280),
name=exp_dir / f'out/{epoch_idx}_{batch_idx}.png')
# -- end save every
# -- end batches
if epoch_idx % cfg.epoch_every == 0:
epoch_avg /= (len(dataloader) * cfg.epoch_every)
writer.add_scalar('train/epoch_avg_loss',
avg_loss / cfg.batch_every,
epoch_idx // cfg.epoch_every)
logger.info('Epoch avg = %.8f' % epoch_avg)
epoch_avg = 0.0
T.save(model.state_dict(),
exp_dir / f'checkpoint/model_{epoch_idx}.state')
# -- end epoch every
# -- end epoch
# save final model
T.save(model.state_dict(), exp_dir / 'model_final.state')
# cleaning
writer.close()
def train(cfg):
os.makedirs(f"out/{cfg['exp_name']}", exist_ok=True)
os.makedirs(f"checkpoints/{cfg['exp_name']}", exist_ok=True)
# dump config for current experiment
with open(f"checkpoints/{cfg['exp_name']}/setup.cfg", "wt") as f:
for k, v in cfg.items():
f.write("%15s: %s\n" % (k, v))
model = CAE().cuda()
if cfg['load']:
model.load_state_dict(torch.load(cfg['chkpt']))
logger.info("Loaded model from", cfg['chkpt'])
model.train()
logger.info("Done setup model")
dataset = ImageFolder720p(cfg['dataset_path'])
dataloader = DataLoader(dataset,
batch_size=cfg['batch_size'],
shuffle=cfg['shuffle'],
num_workers=cfg['num_workers'])
logger.info(
f"Done setup dataloader: {len(dataloader)} batches of size {cfg['batch_size']}"
)
mse_loss = nn.MSELoss()
adam = torch.optim.Adam(model.parameters(),
lr=cfg['learning_rate'],
weight_decay=1e-5)
sgd = torch.optim.SGD(model.parameters(), lr=cfg['learning_rate'])
optimizer = adam
ra = 0
for ei in range(cfg['resume_epoch'], cfg['num_epochs']):
for bi, (img, patches, _) in enumerate(dataloader):
avg_loss = 0
for i in range(6):
for j in range(10):
x = Variable(patches[:, :, i, j, :, :]).cuda()
y = model(x)
loss = mse_loss(y, x)
avg_loss += (1 / 60) * loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
ra = avg_loss if bi == 0 else ra * bi / (bi + 1) + avg_loss / (bi +
1)
logger.debug('[%3d/%3d][%5d/%5d] avg_loss: %f, ra: %f' %
(ei + 1, cfg['num_epochs'], bi + 1, len(dataloader),
avg_loss, ra))
# save img
if bi % cfg['out_every'] == 0:
out = torch.zeros(6, 10, 3, 128, 128)
for i in range(6):
for j in range(10):
x = Variable(patches[0, :, i,
j, :, :].unsqueeze(0)).cuda()
out[i, j] = model(x).cpu().data
out = np.transpose(out, (0, 3, 1, 4, 2))
out = np.reshape(out, (768, 1280, 3))
out = np.transpose(out, (2, 0, 1))
y = torch.cat((img[0], out), dim=2).unsqueeze(0)
save_imgs(imgs=y,
to_size=(3, 768, 2 * 1280),
name=f"out/{cfg['exp_name']}/out_{ei}_{bi}.png")
# save model
if bi % cfg['save_every'] == cfg['save_every'] - 1:
torch.save(
model.state_dict(),
f"checkpoints/{cfg['exp_name']}/model_{ei}_{bi}.state")
# save final model
torch.save(model.state_dict(),
f"checkpoints/{cfg['exp_name']}/model_final.state")