def train(network_specs,
training_params,
image_path,
save_path,
ckpt_path,
epoch=10):
print('creating datapipe...')
# create images DataPipeline
datapipe = DataPipeline(image_path=image_path,
training_params=training_params)
print('creating network model...')
# create model VAE
model = UNet(network_specs=network_specs,
datapipe=datapipe,
training_params=training_params)
# train the model
# save_config is flexible
print('''
=============
HERE WE GO
=============
''')
model.train(save_path=save_path, ckpt_path=ckpt_path, epoch=epoch)
def lr_find(model: UNet,
data_loader,
optimizer: Optimizer,
criterion,
use_gpu,
min_lr=0.0001,
max_lr=0.1):
# Save model and optimizer states to revert
model_state = model.state_dict()
optimizer_state = optimizer.state_dict()
losses = []
lrs = []
scheduler = CyclicExpLR(optimizer,
min_lr,
max_lr,
step_size_up=100,
mode='triangular',
cycle_momentum=True)
model.train()
for i, (data, target, class_ids) in enumerate(data_loader):
data, target = data, target
if use_gpu:
data = data.cuda()
target = target.cuda()
optimizer.zero_grad()
output_raw = model(data)
# This step is specific for this project
output = torch.zeros(output_raw.shape[0], 1, output_raw.shape[2],
output_raw.shape[3])
if use_gpu:
output = output.cuda()
# This step is specific for this project
for idx, (raw_o, class_id) in enumerate(zip(output_raw, class_ids)):
output[idx] = raw_o[class_id - 1]
loss = criterion(output, target)
loss.backward()
current_lr = optimizer.param_groups[0]['lr']
# Stop if lr stopped increasing
if len(lrs) > 0 and current_lr < lrs[-1]:
break
lrs.append(current_lr)
losses.append(loss.item())
optimizer.step()
scheduler.step()
# Plot in log scale
plt.plot(lrs, losses)
plt.xscale('log')
plt.show()
model.load_state_dict(model_state)
optimizer.load_state_dict(optimizer_state)
def main():
FLAGS = parser.parse_args()
# Loading train and test data.
# load_data function exists in file utils.py
print("Loading dataset.")
X_train, y_train = load_data(FLAGS.dataset_dir + '/train', FLAGS.img_size,
FLAGS.augment_data)
X_test, y_test = load_data(FLAGS.dataset_dir + '/test', FLAGS.img_size,
FLAGS.augment_data)
# Making sure that the data was loaded successfully.
print("Train set image size : ", X_train.shape)
print("Train set label size : ", y_train.shape)
print("Test set image size : ", X_test.shape)
print("Test set label size : ", y_test.shape)
print("Dataset loaded successfully.")
# Creating a unet object.
# class UNet exists in file model.py
unet = UNet(FLAGS.img_size)
# Training the network, printing the loss value for every epoch
# , and the accuracy on the test set after the training is complete
train_loss_values, test_loss_values = unet.train(X_train, y_train, X_test,
y_test, FLAGS.num_epochs,
FLAGS.learning_rate,
FLAGS.model_save_dir)
# Plotting loss values on train and test set, and saving it as an image Loss.png
# plot_loss exists in file utils.py
plot_loss(train_loss_values, test_loss_values, 'Loss.png')
shuffle=True,
num_workers=8,
pin_memory=False,
drop_last=True)
val_dataloader = DataLoader(val_dataset,
batch_size=batchSize,
shuffle=False,
num_workers=8,
pin_memory=False)
trainSize = len(train_dataset)
valSize = len(val_dataset)
for epoch in range(epoch0, epoch_num):
train_loss, val_loss = 0.0, 0.0
G.train(), D.train()
for i, batch in enumerate(train_dataloader):
for d_iter in range(1):
optimizer_D.zero_grad()
img, label, onehot = batch[0].to(device), batch[1].to(
device), batch[2].to(device)
output = G(img)
real_imgs = torch.cat((img, onehot), 1)
fake_imgs = torch.cat((img, output.detach()), 1)
loss_D = -torch.mean(D(real_imgs)) + torch.mean(D(fake_imgs))
loss_D.backward()
optimizer_D.step()
def main():
weights = './weights'
logs = './logs'
makedirs(weights, logs)
#snapshot(logs)
device = torch.device("cpu" if not torch.cuda.is_available() else "cuda:0")
images = './kaggle_3m'
image_size = 256
scale = 0.05
angle = 15
batch_size = 16
workers = 4
loader_train, loader_valid = data_loaders(images, image_size, scale, angle,
batch_size, workers)
loaders = {"train": loader_train, "valid": loader_valid}
unet = UNet(in_channels=Dataset.in_channels,
out_channels=Dataset.out_channels)
unet.to(device)
dsc_loss = DiceLoss()
best_validation_dsc = 0.0
lr = 0.0001
optimizer = optim.Adam(unet.parameters(), lr)
logger = Logger(logs)
loss_train = []
loss_valid = []
step = 0
epochs = 100
vis_images = 200
vis_freq = 10
for epoch in tqdm(range(epochs), total=epochs):
for phase in ["train", "valid"]:
if phase == "train":
unet.train()
else:
unet.eval()
validation_pred = []
validation_true = []
for i, data in enumerate(loaders[phase]):
if phase == "train":
step += 1
x, y_true = data
x, y_true = x.to(device), y_true.to(device)
optimizer.zero_grad()
with torch.set_grad_enabled(phase == "train"):
y_pred = unet(x)
loss = dsc_loss(y_pred, y_true)
if phase == "valid":
loss_valid.append(loss.item())
y_pred_np = y_pred.detach().cpu().numpy()
validation_pred.extend(
[y_pred_np[s] for s in range(y_pred_np.shape[0])])
y_true_np = y_true.detach().cpu().numpy()
validation_true.extend(
[y_true_np[s] for s in range(y_true_np.shape[0])])
if (epoch % vis_freq == 0) or (epoch == epochs - 1):
if i * batch_size < vis_images:
tag = "image/{}".format(i)
num_images = vis_images - i * batch_size
logger.image_list_summary(
tag,
log_images(x, y_true, y_pred)[:num_images],
step,
)
if phase == "train":
loss_train.append(loss.item())
loss.backward()
optimizer.step()
if phase == "train" and (step + 1) % 10 == 0:
log_loss_summary(logger, loss_train, step)
loss_train = []
if phase == "valid":
log_loss_summary(logger, loss_valid, step, prefix="val_")
mean_dsc = np.mean(
dsc_per_volume(
validation_pred,
validation_true,
loader_valid.dataset.patient_slice_index,
))
logger.scalar_summary("val_dsc", mean_dsc, step)
if mean_dsc > best_validation_dsc:
best_validation_dsc = mean_dsc
torch.save(unet.state_dict(),
os.path.join(weights, "unet.pt"))
loss_valid = []
print("Best validation mean DSC: {:4f}".format(best_validation_dsc))
else:
loss = criterion(output, target)
out = output.cpu().detach().numpy()
out = np.array(out[0] > 0.5)
dice_val += dice(out, target_pt[0])
test_loss += loss.item() * data.size(0)
dice_val = dice_val / numpts
test_loss = test_loss / numpts
print("Dice: " + str(dice_val))
print("Loss: " + str(test_loss))
# Training
model.train()
np.random.shuffle(indices)
all_data = all_data[indices]
labels = labels[indices]
train_loss = 0.0
numpts = len(all_data)
for i in range(0, numpts, batch_size):
data = all_data[i:i+batch_size]
target = labels[i:i+batch_size]
if LOSS_NUM == 3 or LOSS_NUM == 4:
wts = weights[i:i+batch_size]
# Doesn't improve performance
if AUG > 0:
class Instructor:
''' Model training and evaluation '''
def __init__(self, opt):
self.opt = opt
if opt.inference:
self.testset = TestImageDataset(fdir=opt.impaths['test'],
imsize=opt.imsize)
else:
self.trainset = ImageDataset(fdir=opt.impaths['train'],
bdir=opt.impaths['btrain'],
imsize=opt.imsize,
mode='train',
aug_prob=opt.aug_prob,
prefetch=opt.prefetch)
self.valset = ImageDataset(fdir=opt.impaths['val'],
bdir=opt.impaths['bval'],
imsize=opt.imsize,
mode='val',
aug_prob=opt.aug_prob,
prefetch=opt.prefetch)
self.model = UNet(n_channels=3,
n_classes=1,
bilinear=self.opt.use_bilinear)
if opt.checkpoint:
self.model.load_state_dict(
torch.load('./state_dict/{:s}'.format(opt.checkpoint),
map_location=self.opt.device))
print('checkpoint {:s} has been loaded'.format(opt.checkpoint))
if opt.multi_gpu == 'on':
self.model = torch.nn.DataParallel(self.model)
self.model = self.model.to(opt.device)
self._print_args()
def _print_args(self):
n_trainable_params, n_nontrainable_params = 0, 0
for p in self.model.parameters():
n_params = torch.prod(torch.tensor(p.shape))
if p.requires_grad:
n_trainable_params += n_params
else:
n_nontrainable_params += n_params
self.info = 'n_trainable_params: {0}, n_nontrainable_params: {1}\n'.format(
n_trainable_params, n_nontrainable_params)
self.info += 'training arguments:\n' + '\n'.join([
'>>> {0}: {1}'.format(arg, getattr(self.opt, arg))
for arg in vars(self.opt)
])
if self.opt.device.type == 'cuda':
print('cuda memory allocated:',
torch.cuda.memory_allocated(opt.device.index))
print(self.info)
def _reset_records(self):
self.records = {
'best_epoch': 0,
'best_dice': 0,
'train_loss': list(),
'val_loss': list(),
'val_dice': list(),
'checkpoints': list()
}
def _update_records(self, epoch, train_loss, val_loss, val_dice):
if val_dice > self.records['best_dice']:
path = './state_dict/{:s}_dice{:.4f}_temp{:s}.pt'.format(
self.opt.model_name, val_dice,
str(time.time())[-6:])
if self.opt.multi_gpu == 'on':
torch.save(self.model.module.state_dict(), path)
else:
torch.save(self.model.state_dict(), path)
self.records['best_epoch'] = epoch
self.records['best_dice'] = val_dice
self.records['checkpoints'].append(path)
self.records['train_loss'].append(train_loss)
self.records['val_loss'].append(val_loss)
self.records['val_dice'].append(val_dice)
def _draw_records(self):
timestamp = str(int(time.time()))
print('best epoch: {:d}'.format(self.records['best_epoch']))
print('best train loss: {:.4f}, best val loss: {:.4f}'.format(
min(self.records['train_loss']), min(self.records['val_loss'])))
print('best val dice {:.4f}'.format(self.records['best_dice']))
os.rename(
self.records['checkpoints'][-1],
'./state_dict/{:s}_dice{:.4f}_save{:s}.pt'.format(
self.opt.model_name, self.records['best_dice'], timestamp))
for path in self.records['checkpoints'][0:-1]:
os.remove(path)
# Draw figures
plt.figure()
trainloss, = plt.plot(self.records['train_loss'])
valloss, = plt.plot(self.records['val_loss'])
plt.legend([trainloss, valloss], ['train', 'val'], loc='upper right')
plt.title('{:s} loss curve'.format(timestamp))
plt.savefig('./figs/{:s}_loss.png'.format(timestamp),
format='png',
transparent=True,
dpi=300)
plt.figure()
valdice, = plt.plot(self.records['val_dice'])
plt.title('{:s} dice curve'.format(timestamp))
plt.savefig('./figs/{:s}_dice.png'.format(timestamp),
format='png',
transparent=True,
dpi=300)
# Save report
report = '\t'.join(
['val_dice', 'train_loss', 'val_loss', 'best_epoch', 'timestamp'])
report += "\n{:.4f}\t{:.4f}\t{:.4f}\t{:d}\t{:s}\n{:s}".format(
self.records['best_dice'], min(self.records['train_loss']),
min(self.records['val_loss']), self.records['best_epoch'],
timestamp, self.info)
with open('./logs/{:s}_log.txt'.format(timestamp), 'w') as f:
f.write(report)
print('report saved:', './logs/{:s}_log.txt'.format(timestamp))
def _train(self, train_dataloader, criterion, optimizer):
self.model.train()
train_loss, n_total, n_batch = 0, 0, len(train_dataloader)
for i_batch, sample_batched in enumerate(train_dataloader):
inputs, target = sample_batched[0].to(
self.opt.device), sample_batched[1].to(self.opt.device)
predict = self.model(inputs)
optimizer.zero_grad()
loss = criterion(predict, target)
loss.backward()
optimizer.step()
train_loss += loss.item() * len(sample_batched)
n_total += len(sample_batched)
ratio = int((i_batch + 1) * 50 / n_batch)
sys.stdout.write("\r[" + ">" * ratio + " " * (50 - ratio) +
"] {}/{} {:.2f}%".format(i_batch + 1, n_batch,
(i_batch + 1) * 100 /
n_batch))
sys.stdout.flush()
print()
return train_loss / n_total
def _evaluation(self, val_dataloader, criterion):
self.model.eval()
val_loss, val_dice, n_total = 0, 0, 0
with torch.no_grad():
for sample_batched in val_dataloader:
inputs, target = sample_batched[0].to(
self.opt.device), sample_batched[1].to(self.opt.device)
predict = self.model(inputs)
loss = criterion(predict, target)
dice = dice_coeff(predict, target)
val_loss += loss.item() * len(sample_batched)
val_dice += dice.item() * len(sample_batched)
n_total += len(sample_batched)
return val_loss / n_total, val_dice / n_total
def run(self):
_params = filter(lambda p: p.requires_grad, self.model.parameters())
optimizer = torch.optim.Adam(_params,
lr=self.opt.lr,
weight_decay=self.opt.l2reg)
criterion = BCELoss2d()
train_dataloader = DataLoader(dataset=self.trainset,
batch_size=self.opt.batch_size,
shuffle=True)
val_dataloader = DataLoader(dataset=self.valset,
batch_size=self.opt.batch_size,
shuffle=False)
self._reset_records()
for epoch in range(self.opt.num_epoch):
train_loss = self._train(train_dataloader, criterion, optimizer)
val_loss, val_dice = self._evaluation(val_dataloader, criterion)
self._update_records(epoch, train_loss, val_loss, val_dice)
print(
'{:d}/{:d} > train loss: {:.4f}, val loss: {:.4f}, val dice: {:.4f}'
.format(epoch + 1, self.opt.num_epoch, train_loss, val_loss,
val_dice))
self._draw_records()
def inference(self):
test_dataloader = DataLoader(dataset=self.testset,
batch_size=1,
shuffle=False)
n_batch = len(test_dataloader)
with torch.no_grad():
for i_batch, sample_batched in enumerate(test_dataloader):
index, inputs = sample_batched[0], sample_batched[1].to(
self.opt.device)
predict = self.model(inputs)
self.testset.save_img(index.item(), predict, self.opt.use_crf)
ratio = int((i_batch + 1) * 50 / n_batch)
sys.stdout.write(
"\r[" + ">" * ratio + " " * (50 - ratio) +
"] {}/{} {:.2f}%".format(i_batch + 1, n_batch,
(i_batch + 1) * 100 / n_batch))
sys.stdout.flush()
print()
def train(n_channels, n_classes, bilinear, epochs, batch_size, lr, val_rate,
num_workers, pin_memory, roots, threshold):
data_root, model_root, log_root = roots
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
print(torch.cuda.get_device_properties(device))
print("id：{}".format(device))
print("name:{}".format(torch.cuda.get_device_name(0)))
logging.info(f'Using device {device}')
model = UNet(n_channels, n_classes, bilinear).to(device)
logging.info(f'Network:\n'
f'\t{n_channels} input channels\n'
f'\t{n_classes} output channels (classes)\n'
f'\t{"Bilinear" if bilinear else "Dilated conv"} upscaling')
dataset = SuperviselyDataset(data_root)
num_val = int(len(dataset) * val_rate)
num_train = len(dataset) - num_val
train_date, val_data = random_split(dataset, [num_train, num_val])
train_loader = DataLoader(train_date,
batch_size,
shuffle=True,
num_workers=num_workers,
pin_memory=pin_memory)
val_loader = DataLoader(val_data,
batch_size,
shuffle=False,
num_workers=num_workers,
pin_memory=pin_memory)
writer = SummaryWriter(comment=f'LR_{lr}_BS_{batch_size}')
logging.info(f'''Starting training:
Epochs: {epochs}
Batch size: {batch_size}
Learning rate (original): {lr}
Training item: {num_train}
Validation item: {num_val}
Device: {device.type}''')
# criterion = nn.CrossEntropyLoss() if n_classes > 1 else nn.BCEWithLogitsLoss()
criterion = nn.MSELoss()
optimizer = optim.RMSprop(model.parameters(),
lr=lr,
weight_decay=1e-8,
momentum=0.9)
# scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min' if n_classes > 1 else 'max', patience=2)
losses = []
global_step = 0
for epoch in range(epochs):
model.train()
with tqdm(total=num_train) as t:
t.set_description('epoch: {}/{}'.format(epoch + 1, epochs))
for img, mask in train_loader:
# import cv2
# print(mask.shape)
# cv2.imshow('figure1', img[0].permute(1, 2, 0).numpy())
# cv2.waitKey()
# cv2.imshow('figure2', mask[0].permute(1, 2, 0).numpy())
# cv2.waitKey()
mask_type = torch.float32 if n_classes == 1 else torch.long
mask = mask.to(device=device, dtype=mask_type)
img = img.to(device=device, dtype=torch.float32)
# update
pred = model(img)
loss = criterion(pred, mask)
optimizer.zero_grad()
nn.utils.clip_grad_value_(model.parameters(), 0.1)
loss.backward()
optimizer.step()
losses.append(loss.item())
# writer.add_scalar('Loss/train', loss.item(), global_step)
t.set_postfix(loss='{:.6f}'.format(loss),
lr='%.8f' % optimizer.param_groups[0]['lr'])
t.update(img.shape[0])
# value
# global_step += 1
# if global_step % (len(dataset) // (10 * batch_size)) == 0:
# score = evaluate(model, val_loader, n_classes, device, num_val)
# model.train()
# logging.info('Validation Dice Coeff: {}'.format(score))
# scheduler.step(score)
# if n_classes > 1:
# logging.info('Validation cross entropy: {}'.format(score))
# writer.add_scalar('Loss/test', score, global_step)
# else:
# logging.info('Validation Dice Coeff: {}'.format(score))
# writer.add_scalar('Dice/test', score, global_step)
#
# writer.add_images('images', img, global_step)
# if n_classes == 1:
# writer.add_images('masks/true', mask, global_step)
# writer.add_images('masks/pred', torch.sigmoid(pred) > threshold, global_step)
save_model_and_loss(model, model_root, losses, log_root)
writer.close()
def train(args):
"""
Train UNet from datasets
"""
# dataset
print('Reading dataset from {}...'.format(args.dataset_path))
train_dataset = SSDataset(dataset_path=args.dataset_path, is_train=True)
val_dataset = SSDataset(dataset_path=args.dataset_path, is_train=False)
train_dataloader = DataLoader(dataset=train_dataset,
batch_size=args.batch_size,
shuffle=True)
val_dataloader = DataLoader(dataset=val_dataset,
batch_size=args.batch_size,
shuffle=False)
# mask
with open(args.mask_json_path, 'w', encoding='utf-8') as mask:
colors = SSDataset.all_colors
mask.write(json.dumps(colors))
print('Mask colors list has been saved in {}'.format(
args.mask_json_path))
# model
net = UNet(in_channels=3, out_channels=5)
if args.cuda:
net = net.cuda()
# setting
lr = args.lr # 1e-3
optimizer = optim.Adam(net.parameters(), lr=lr)
criterion = loss_fn
# run
train_losses = []
val_losses = []
print('Start training...')
for epoch_idx in range(args.epochs):
# train
net.train()
train_loss = 0
for batch_idx, batch_data in enumerate(train_dataloader):
xs, ys = batch_data
if args.cuda:
xs = xs.cuda()
ys = ys.cuda()
ys_pred = net(xs)
loss = criterion(ys_pred, ys)
train_loss += loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
# val
net.eval()
val_loss = 0
for batch_idx, batch_data in enumerate(val_dataloader):
xs, ys = batch_data
if args.cuda:
xs = xs.cuda()
ys = ys.cuda()
ys_pred = net(xs)
loss = loss_fn(ys_pred, ys)
val_loss += loss
train_losses.append(train_loss)
val_losses.append(val_loss)
print('Epoch: {}, Train total loss: {}, Val total loss: {}'.format(
epoch_idx + 1, train_loss.item(), val_loss.item()))
# save
if (epoch_idx + 1) % args.save_epoch == 0:
checkpoint_path = os.path.join(
args.checkpoint_path,
'checkpoint_{}.pth'.format(epoch_idx + 1))
torch.save(net.state_dict(), checkpoint_path)
print('Saved Checkpoint at Epoch {} to {}'.format(
epoch_idx + 1, checkpoint_path))
# summary
if args.do_save_summary:
epoch_range = list(range(1, args.epochs + 1))
plt.plot(epoch_range, train_losses, 'r', label='Train loss')
plt.plot(epoch_range, val_loss, 'g', label='Val loss')
plt.imsave(args.summary_image)
print('Summary images have been saved in {}'.format(
args.summary_image))
# save
net.eval()
torch.save(net.state_dict(), args.model_state_dict)
print('Saved state_dict in {}'.format(args.model_state_dict))
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--bs', metavar='bs', type=int, default=2)
parser.add_argument('--path', type=str, default='../../data')
parser.add_argument('--results', type=str, default='../../results/model')
parser.add_argument('--nw', type=int, default=0)
parser.add_argument('--max_images', type=int, default=None)
parser.add_argument('--val_size', type=int, default=None)
parser.add_argument('--epochs', type=int, default=10)
parser.add_argument('--lr', type=float, default=0.003)
parser.add_argument('--lr_decay', type=float, default=0.99997)
parser.add_argument('--kernel_lvl', type=float, default=1)
parser.add_argument('--noise_lvl', type=float, default=1)
parser.add_argument('--motion_blur', type=bool, default=False)
parser.add_argument('--homo_align', type=bool, default=False)
parser.add_argument('--resume', type=bool, default=False)
args = parser.parse_args()
print()
print(args)
print()
if not os.path.isdir(args.results): os.makedirs(args.results)
PATH = args.results
if not args.resume:
f = open(PATH + "/param.txt", "a+")
f.write(str(args))
f.close()
writer = SummaryWriter(PATH + '/runs')
# CUDA for PyTorch
use_cuda = torch.cuda.is_available()
device = torch.device('cuda:0' if use_cuda else "cpu")
# Parameters
params = {'batch_size': args.bs, 'shuffle': True, 'num_workers': args.nw}
# Generators
print('Initializing training set')
training_set = Dataset(args.path + '/train/', args.max_images,
args.kernel_lvl, args.noise_lvl, args.motion_blur,
args.homo_align)
training_generator = data.DataLoader(training_set, **params)
print('Initializing validation set')
validation_set = Dataset(args.path + '/test/', args.val_size,
args.kernel_lvl, args.noise_lvl, args.motion_blur,
args.homo_align)
validation_generator = data.DataLoader(validation_set, **params)
# Model
model = UNet(in_channel=3, out_channel=3)
if args.resume:
models_path = get_newest_model(PATH)
print('loading model from ', models_path)
model.load_state_dict(torch.load(models_path))
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
model = torch.nn.DataParallel(model)
model.to(device)
# Loss + optimizer
criterion = BurstLoss()
optimizer = RAdam(model.parameters(), lr=args.lr)
scheduler = StepLR(optimizer, step_size=8 // args.bs, gamma=args.lr_decay)
if args.resume:
n_iter = np.loadtxt(PATH + '/train.txt', delimiter=',')[:, 0][-1]
else:
n_iter = 0
# Loop over epochs
for epoch in range(args.epochs):
train_loss = 0.0
# Training
model.train()
for i, (X_batch, y_labels) in enumerate(training_generator):
# Alter the burst length for each mini batch
burst_length = np.random.randint(2, 9)
X_batch = X_batch[:, :burst_length, :, :, :]
# Transfer to GPU
X_batch, y_labels = X_batch.to(device).type(
torch.float), y_labels.to(device).type(torch.float)
# zero the parameter gradients
optimizer.zero_grad()
# forward + backward + optimize
pred = model(X_batch)
loss = criterion(pred, y_labels)
loss.backward()
optimizer.step()
scheduler.step()
train_loss += loss.detach().cpu().numpy()
writer.add_scalar('training_loss', loss.item(), n_iter)
if i % 100 == 0 and i > 0:
loss_printable = str(np.round(train_loss, 2))
f = open(PATH + "/train.txt", "a+")
f.write(str(n_iter) + "," + loss_printable + "\n")
f.close()
print("training loss ", loss_printable)
train_loss = 0.0
if i % 1000 == 0:
if torch.cuda.device_count() > 1:
torch.save(
model.module.state_dict(),
os.path.join(PATH,
'model_' + str(int(n_iter)) + '.pt'))
else:
torch.save(
model.state_dict(),
os.path.join(PATH,
'model_' + str(int(n_iter)) + '.pt'))
if i % 1000 == 0:
# Validation
val_loss = 0.0
with torch.set_grad_enabled(False):
model.eval()
for v, (X_batch,
y_labels) in enumerate(validation_generator):
# Alter the burst length for each mini batch
burst_length = np.random.randint(2, 9)
X_batch = X_batch[:, :burst_length, :, :, :]
# Transfer to GPU
X_batch, y_labels = X_batch.to(device).type(
torch.float), y_labels.to(device).type(torch.float)
# forward + backward + optimize
pred = model(X_batch)
loss = criterion(pred, y_labels)
val_loss += loss.detach().cpu().numpy()
if v < 5:
im = make_im(pred, X_batch, y_labels)
writer.add_image('image_' + str(v), im, n_iter)
writer.add_scalar('validation_loss', val_loss, n_iter)
loss_printable = str(np.round(val_loss, 2))
print('validation loss ', loss_printable)
f = open(PATH + "/eval.txt", "a+")
f.write(str(n_iter) + "," + loss_printable + "\n")
f.close()
n_iter += args.bs
def train(args):
dataset = open("dataset.csv", "r").readlines()
train_set = dataset[:600]
val_set = dataset[600:]
root_dir = root_dir = "data/Lung_Segmentation/"
train_data = LungSegmentationDataGen(train_set, root_dir, args)
val_data = LungSegmentationDataGen(val_set, root_dir, args)
train_dataloader = DataLoader(train_data,
batch_size=5,
shuffle=True,
num_workers=4)
val_dataloader = DataLoader(val_data,
batch_size=5,
shuffle=True,
num_workers=4)
dataloaders = {"train": train_dataloader, "val": val_dataloader}
dataset_sizes = {"train": len(train_set), "val": len(val_set)}
print("dataset_sizes: {}".format(dataset_sizes))
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = UNet(in_channels=1)
model = model.to(device)
optimizer = optim.Adam(model.parameters())
loss_train = []
loss_valid = []
current_mean_dsc = 0.0
best_validation_dsc = 0.0
epochs = args.epochs
for epoch in range(epochs):
print('Epoch {}/{}'.format(epoch, epochs - 1))
print('-' * 10)
dice_score_list = []
for phase in ['train', 'val']:
if phase == 'train':
model.train() # Set model to training mode
else:
model.eval() # Set model to evaluate mode
# Iterate over data.
for i, data in enumerate(dataloaders[phase]):
inputs, y_true = data
inputs = inputs.to(device)
y_true = y_true.to(device)
# zero the parameter gradients
optimizer.zero_grad()
with torch.set_grad_enabled(phase == 'train'):
# forward pass with batch input
y_pred = model(inputs)
loss = dice_loss(y_true, y_pred)
# backward + optimize only if in training phase
if phase == 'train':
# print("step: {}, train_loss: {}".format(i, loss))
loss_train.append(loss.item())
# calculate the gradients based on loss
loss.backward()
# update the weights
optimizer.step()
if phase == "val":
loss_valid.append(loss.item())
dsc = dice_score(y_true, y_pred)
print("step: {}, val_loss: {}, val dice_score: {}".
format(i, loss, dsc))
dice_score_list.append(dsc.detach().numpy())
if phase == "train" and (i + 1) % 10 == 0:
print("step:{}, train_loss: {}".format(
i + 1, np.mean(loss_train)))
loss_train = []
if phase == "val":
print("mean val_loss: {}".format(np.mean(loss_valid)))
loss_valid = []
current_mean_dsc = np.mean(dice_score_list)
print("validation set dice_score: {}".format(current_mean_dsc))
if current_mean_dsc > best_validation_dsc:
best_validation_dsc = current_mean_dsc
print("best dice_score on val set: {}".format(
best_validation_dsc))
model_name = "unet_{0:.2f}.pt".format(best_validation_dsc)
torch.save(model.state_dict(),
os.path.join(args.weights, model_name))
def main(_):
# Make checkpoint directory
if not os.path.exists(args.checkpoint_dir):
os.makedirs(args.checkpoint_dir)
# Create a dataset object
if label_type == 'one_hot':
data=utils.DataOneHot(debug=args.debug, patch_overlap=args.patch_overlap, im_size=args.im_size, \
band_n=args.band_n, t_len=args.t_len, path=args.path, class_n=args.class_n, pc_mode=args.pc_mode, \
test_n_limit=args.test_n_limit,memory_mode=args.memory_mode, \
balance_samples_per_class=args.balance_samples_per_class, test_get_stride=args.test_get_stride, \
n_apriori=args.n_apriori,patch_length=args.patch_len,squeeze_classes=args.squeeze_classes,im_h=args.im_h,im_w=args.im_w, \
id_first=args.id_first, train_test_mask_name=args.train_test_mask_name, \
test_overlap_full=args.test_overlap_full,ram_store=args.ram_store,patches_save=args.patches_save)
elif label_type == 'semantic':
data=utils.DataSemantic(debug=args.debug, patch_overlap=args.patch_overlap, im_size=args.im_size, \
band_n=args.band_n, t_len=args.t_len, path=args.path, class_n=args.class_n, pc_mode=args.pc_mode, \
test_n_limit=args.test_n_limit,memory_mode=args.memory_mode, \
balance_samples_per_class=args.balance_samples_per_class, test_get_stride=args.test_get_stride, \
n_apriori=args.n_apriori,patch_length=args.patch_len,squeeze_classes=args.squeeze_classes,im_h=args.im_h,im_w=args.im_w, \
id_first=args.id_first, train_test_mask_name=args.train_test_mask_name, \
test_overlap_full=args.test_overlap_full,ram_store=args.ram_store,patches_save=args.patches_save)
# Load images and create dataset (Extract patches)
if args.memory_mode == "ram":
data.create()
deb.prints(data.ram_data["train"]["ims"].shape)
# Run tensorflow session
with tf.Session() as sess:
# Create a neural network object (Define model graph)
if args.model == 'convlstm':
model = conv_lstm(sess,
batch_size=args.batch_size,
epoch=args.epoch,
train_size=args.train_size,
timesteps=args.timesteps,
patch_len=args.patch_len,
kernel=args.kernel,
channels=args.channels,
filters=args.filters,
n_classes=args.n_classes,
checkpoint_dir=args.checkpoint_dir,
log_dir=args.log_dir,
data=data.ram_data,
conf=data.conf,
debug=args.debug)
elif args.model == 'conv3d':
model = Conv3DMultitemp(sess,
batch_size=args.batch_size,
epoch=args.epoch,
train_size=args.train_size,
timesteps=args.timesteps,
patch_len=args.patch_len,
kernel=args.kernel,
channels=args.channels,
filters=args.filters,
n_classes=args.n_classes,
checkpoint_dir=args.checkpoint_dir,
log_dir=args.log_dir,
data=data.ram_data,
conf=data.conf,
debug=args.debug)
elif args.model == 'unet':
model = UNet(sess,
batch_size=args.batch_size,
epoch=args.epoch,
train_size=args.train_size,
timesteps=args.timesteps,
patch_len=args.patch_len,
kernel=args.kernel,
channels=args.channels,
filters=args.filters,
n_classes=args.n_classes,
checkpoint_dir=args.checkpoint_dir,
log_dir=args.log_dir,
data=data.ram_data,
conf=data.conf,
debug=args.debug)
elif args.model == 'smcnn':
model = SMCNN(sess,
batch_size=args.batch_size,
epoch=args.epoch,
train_size=args.train_size,
timesteps=args.timesteps,
patch_len=args.patch_len,
kernel=args.kernel,
channels=args.channels,
filters=args.filters,
n_classes=args.n_classes,
checkpoint_dir=args.checkpoint_dir,
log_dir=args.log_dir,
data=data.ram_data,
conf=data.conf,
debug=args.debug)
elif args.model == 'smcnnlstm':
model = SMCNNlstm(sess,
batch_size=args.batch_size,
epoch=args.epoch,
train_size=args.train_size,
timesteps=args.timesteps,
patch_len=args.patch_len,
kernel=args.kernel,
channels=args.channels,
filters=args.filters,
n_classes=args.n_classes,
checkpoint_dir=args.checkpoint_dir,
log_dir=args.log_dir,
data=data.ram_data,
conf=data.conf,
debug=args.debug)
elif args.model == 'smcnn_unet':
model = SMCNN_UNet(sess,
batch_size=args.batch_size,
epoch=args.epoch,
train_size=args.train_size,
timesteps=args.timesteps,
patch_len=args.patch_len,
kernel=args.kernel,
channels=args.channels,
filters=args.filters,
n_classes=args.n_classes,
checkpoint_dir=args.checkpoint_dir,
log_dir=args.log_dir,
data=data.ram_data,
conf=data.conf,
debug=args.debug)
elif args.model == 'smcnn_conv3d':
model = SMCNN_conv3d(sess,
batch_size=args.batch_size,
epoch=args.epoch,
train_size=args.train_size,
timesteps=args.timesteps,
patch_len=args.patch_len,
kernel=args.kernel,
channels=args.channels,
filters=args.filters,
n_classes=args.n_classes,
checkpoint_dir=args.checkpoint_dir,
log_dir=args.log_dir,
data=data.ram_data,
conf=data.conf,
debug=args.debug)
elif args.model == 'lstm':
model = lstm(sess,
batch_size=args.batch_size,
epoch=args.epoch,
train_size=args.train_size,
timesteps=args.timesteps,
patch_len=args.patch_len,
kernel=args.kernel,
channels=args.channels,
filters=args.filters,
n_classes=args.n_classes,
checkpoint_dir=args.checkpoint_dir,
log_dir=args.log_dir,
data=data.ram_data,
conf=data.conf,
debug=args.debug)
elif args.model == 'convlstm_semantic':
model = conv_lstm_semantic(sess,
batch_size=args.batch_size,
epoch=args.epoch,
train_size=args.train_size,
timesteps=args.timesteps,
patch_len=args.patch_len,
kernel=args.kernel,
channels=args.channels,
filters=args.filters,
n_classes=args.n_classes,
checkpoint_dir=args.checkpoint_dir,
log_dir=args.log_dir,
data=data.ram_data,
conf=data.conf,
debug=args.debug)
elif args.model == 'smcnn_semantic':
model = SMCNN_semantic(sess,
batch_size=args.batch_size,
epoch=args.epoch,
train_size=args.train_size,
timesteps=args.timesteps,
patch_len=args.patch_len,
kernel=args.kernel,
channels=args.channels,
filters=args.filters,
n_classes=args.n_classes,
checkpoint_dir=args.checkpoint_dir,
log_dir=args.log_dir,
data=data.ram_data,
conf=data.conf,
debug=args.debug)
if args.phase == 'train':
# Train only once
model.train(args)
elif args.phase == 'repeat':
# Train for a specific number of repetitions
model.train_repeat(args)
elif args.phase == 'test':
# Test best model from experiment repetitions
model.test(args)
class FenceSegmentationNet():
def __init__(self, filePathTrain):
# Hyperparameters
self.batchSize = 1
self.numEpochs = 10
self.learningRate = 0.001
self.validPercent = 0.1
self.trainShuffle = True
self.testShuffle = False
self.momentum = 0.99
self.imageDim = 128
# Variables
self.imageDirectory = filePathTrain
self.labelDirectory = filePathTrain
self.numChannels = 3
self.numClasses = 1
# Device configuration
self.device = torch.device(
'cuda' if torch.cuda.is_available() else 'cpu')
# Load dataset
self.trainLoader = self.getTrainingLoader()
#self.testLoader = self.getTestLoader()
# Setup model
self.model = UNet(n_channels=self.numChannels,
n_classes=self.numClasses,
bilinear=True).to(self.device)
#self.optimizer = torch.optim.RMSprop(self.model.parameters(), lr=self.learningRate, weight_decay=self.weightDecay, momentum=self.momentum)
#self.optimizer = torch.optim.Adam(self.model.parameters(), lr=self.learningRate)
self.optimizer = torch.optim.SGD(self.model.parameters(),
lr=self.learningRate,
momentum=self.momentum)
#self.scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(self.optimizer, 'min' if self.numClasses > 1 else 'max', patience=2)
self.criterion = DiceLoss()
def loadImages(self, directory):
fileNames = os.listdir(directory + 'images/')
images = []
labels = []
for fileName in tqdm(fileNames, desc='Loading data'):
image = Image.open(directory + 'images/' + fileName)
label = Image.open(directory + 'labels/' + fileName).convert('L')
label = label.point(lambda x: 0 if x < 128 else 255, '1')
# Converting the data into the format [Channels Width Height]
label = np.asarray(label).reshape(1, self.imageDim, self.imageDim)
image = np.asarray(image).reshape(3, self.imageDim, self.imageDim)
labels.append(label)
images.append(image)
return images, labels
def getTrainingLoader(self):
print('Get trainings loader:')
data, labels = self.loadImages(self.imageDirectory)
# Convert image data and labels to tensor format
tensorData = torch.Tensor(data)
tensorLabels = torch.Tensor(labels)
# Convert the two tensors into one tensor dataset
tensorDataset = TensorDataset(tensorData, tensorLabels)
# Convert the tensor dataset into a dataloader with format [Batch Channels Width Height]
trainLoader = DataLoader(tensorDataset,
batch_size=self.batchSize,
shuffle=self.trainShuffle)
return trainLoader
def getTestLoader(self):
print('Get test loader:')
data, labels = self.loadImages(self.filePathTest)
# Convert image data and labels to tensor format
tensorData = torch.Tensor(data)
tensorLabels = torch.Tensor(labels)
# Convert the two tensors into one tensor dataset
tensorDataset = TensorDataset(tensorData, tensorLabels)
# Convert the tensor dataset into a dataloader with format [Batch Channels Width Height]
testLoader = DataLoader(tensorDataset,
batch_size=self.batchSize,
shuffle=self.testShuffle)
return testLoader
def trainModel(self, patience=3):
print('Train Model:')
self.model.train()
losses = []
n_total_steps = len(self.trainLoader)
for epoch in range(self.numEpochs):
n_correct, epoch_loss = 0, 0
tmpStr = 'Epoch [{:>3}/{:>3}]'.format(epoch + 1, self.numEpochs)
for (samples, labels) in tqdm(self.trainLoader, desc=tmpStr):
samples = samples.to(self.device, dtype=torch.float32)
labels = labels.to(self.device, dtype=torch.long)
# Forward pass
outputs = self.model(samples)
loss = self.criterion(outputs, labels)
epoch_loss += loss.item()
# Backward and optimize
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
tot_loss = epoch_loss / n_total_steps
print(tot_loss)
losses.append(tot_loss)
return losses
def testModel(self):
print('Testing Model:')
self.model.eval()
n_correct = 0
n_samples = 0
tmpStr = 'Testing'
for i, (samples,
labels) in enumerate(tqdm(self.testLoader, desc=tmpStr)):
samples = samples.to(self.device, dtype=torch.float32)
labels = labels.to(self.device, dtype=torch.float32)
outputs = self.model(samples)
_, predicted = torch.max(outputs, 1)
n_samples += labels.size(0)
pred = torch.sigmoid(outputs)
pred = (pred > 0.2).float()
n_correct += dice_coeff(pred, labels).item()
acc = 100.0 * n_correct / n_samples
return acc
def plot_history(self, train_losses):
print('Plotting epoch history')
plt.figure(figsize=(3, 3))
plt.xlabel('epoch')
plt.ylabel('loss')
plt.plot(train_losses, label='train')
plt.ylim(0, 1)
plt.legend()
plt.grid()
plt.tight_layout()
plt.savefig('plot.png')
def saveModel(self, fileName):
torch.save(self.model.state_dict(), fileName)
def loadModel(self, fileName):
self.model.load_state_dict(torch.load(fileName))
# Create context inputs (sequences of 3 slices)
print('Get contexts inputs ...')
ct_seqs = []
for idx, ct in enumerate(imgs):
if (idx != 0) and idx != (len(imgs) - 1):
spatial_context = np.stack((imgs[idx - 1], imgs[idx], imgs[idx + 1]))
ct_seqs.append(spatial_context)
ct_seqs_np = np.array(ct_seqs)
# Remove first and last output segmentation mask
cut_png_imgs = np.delete(png_imgs, [0, len(png_imgs) - 1], 0) # axis = 0
# Initializing u-net and train the network
fcn = UNet()
fcn.train(ct_seqs_np, cut_png_imgs) ##### TRAIN CT
# fcn.train(imgs_mri, png_imgs, (256,256,1)) ##### TRAIN MRI
# # TESTING
# test_data_path = ['./CT_data_batch1/10/DICOM_anon/i0081,0000b.dcm']
# test2 = read_dicoms(test_data_path)
# test1 = get_pixels_hu(test2)
# hi = fcn.predict(test1)
# # OUTPUT FILE plot
# plt.imshow(np.squeeze(hi[0], axis=2))
# plt.show()
print('OK')
def train():
if not os.path.exists('train_model/'):
os.makedirs('train_model/')
if not os.path.exists('result/'):
os.makedirs('result/')
train_data, dev_data, word2id, id2word, char2id, opts = load_data(
vars(args))
model = UNet(opts)
if args.use_cuda:
model = model.cuda()
dev_batches = get_batches(dev_data, args.batch_size, evaluation=True)
if args.eval:
print('load model...')
model.load_state_dict(torch.load(args.model_dir))
model.eval()
model.Evaluate(dev_batches,
args.data_path + 'dev_eval.json',
answer_file='result/' + args.model_dir.split('/')[-1] +
'.answers',
drop_file=args.data_path + 'drop.json',
dev=args.data_path + 'dev-v2.0.json')
exit()
if args.load_model:
print('load model...')
model.load_state_dict(torch.load(args.model_dir))
model.eval()
_, F1 = model.Evaluate(dev_batches,
args.data_path + 'dev_eval.json',
answer_file='result/' +
args.model_dir.split('/')[-1] + '.answers',
drop_file=args.data_path + 'drop.json',
dev=args.data_path + 'dev-v2.0.json')
best_score = F1
with open(args.model_dir + '_f1_scores.pkl', 'rb') as f:
f1_scores = pkl.load(f)
with open(args.model_dir + '_em_scores.pkl', 'rb') as f:
em_scores = pkl.load(f)
else:
best_score = 0.0
f1_scores = []
em_scores = []
parameters = filter(lambda p: p.requires_grad, model.parameters())
optimizer = torch.optim.Adamax(parameters, lr=args.lrate)
lrate = args.lrate
for epoch in range(1, args.epochs + 1):
train_batches = get_batches(train_data, args.batch_size)
dev_batches = get_batches(dev_data, args.batch_size, evaluation=True)
total_size = len(train_data) // args.batch_size
model.train()
for i, train_batch in enumerate(train_batches):
loss = model(train_batch)
model.zero_grad()
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(parameters, opts['grad_clipping'])
optimizer.step()
model.reset_parameters()
if i % 100 == 0:
print(
'Epoch = %d, step = %d / %d, loss = %.5f, lrate = %.5f best_score = %.3f'
% (epoch, i, total_size, model.train_loss.value, lrate,
best_score))
sys.stdout.flush()
model.eval()
exact_match_score, F1 = model.Evaluate(
dev_batches,
args.data_path + 'dev_eval.json',
answer_file='result/' + args.model_dir.split('/')[-1] + '.answers',
drop_file=args.data_path + 'drop.json',
dev=args.data_path + 'dev-v2.0.json')
f1_scores.append(F1)
em_scores.append(exact_match_score)
with open(args.model_dir + '_f1_scores.pkl', 'wb') as f:
pkl.dump(f1_scores, f)
with open(args.model_dir + '_em_scores.pkl', 'wb') as f:
pkl.dump(em_scores, f)
if best_score < F1:
best_score = F1
print('saving %s ...' % args.model_dir)
torch.save(model.state_dict(), args.model_dir)
if epoch > 0 and epoch % args.decay_period == 0:
lrate *= args.decay
for param_group in optimizer.param_groups:
param_group['lr'] = lrate
from model import UNet
from config import UnetConfig
config = UnetConfig()
unet = UNet(config=config)
unet.train("dataset-postdam", "logs")
class Trainer():
def __init__(self,config,trainLoader,validLoader):
self.config = config
self.trainLoader = trainLoader
self.validLoader = validLoader
self.numTrain = len(self.trainLoader.dataset)
self.numValid = len(self.validLoader.dataset)
self.saveModelDir = str(self.config.save_model_dir)+"/"
self.bestModel = config.bestModel
self.useGpu = self.config.use_gpu
self.net = UNet()
if(self.config.resume == True):
print("LOADING SAVED MODEL")
self.loadCheckpoint()
else:
print("INTIALIZING NEW MODEL")
self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
self.net = self.net.to(self.device)
self.totalEpochs = config.epochs
self.optimizer = optim.Adam(self.net.parameters(), lr=5e-4)
self.loss = DiceLoss()
self.num_params = sum([p.data.nelement() for p in self.net.parameters()])
self.trainPaitence = config.train_paitence
if not self.config.resume: # self.freezeLayers(6)
summary(self.net, input_size=(3,256,256))
print('[*] Number of model parameters: {:,}'.format(self.num_params))
self.writer = SummaryWriter(self.config.tensorboard_path+"/")
def train(self):
bestIOU = 0
print("\n[*] Train on {} sample pairs, validate on {} trials".format(
self.numTrain, self.numValid))
for epoch in range(0,self.totalEpochs):
print('\nEpoch: {}/{}'.format(epoch+1, self.totalEpochs))
self.trainOneEpoch(epoch)
validationIOU = self.validationTest(epoch)
print("VALIDATION IOU: ",validationIOU)
# check for improvement
if(validationIOU > bestIOU):
print("COUNT RESET !!!")
bestIOU=validationIOU
self.counter = 0
self.saveCheckPoint(
{
'epoch': epoch + 1,
'model_state': self.net.state_dict(),
'optim_state': self.optimizer.state_dict(),
'best_valid_acc': bestIOU,
},True)
else:
self.counter += 1
if self.counter > self.trainPaitence:
self.saveCheckPoint(
{
'epoch': epoch + 1,
'model_state': self.net.state_dict(),
'optim_state': self.optimizer.state_dict(),
'best_valid_acc': validationIOU,
},False)
print("[!] No improvement in a while, stopping training...")
print("BEST VALIDATION IOU: ",bestIOU)
return None
def trainOneEpoch(self,epoch):
self.net.train()
train_loss = 0
total_IOU = 0
for batch_idx, (images,targets) in enumerate(self.trainLoader):
images = images.to(self.device)
targets = targets.to(self.device)
self.optimizer.zero_grad()
outputMaps = self.net(images)
loss = self.loss(outputMaps,targets)
loss.backward()
self.optimizer.step()
train_loss += loss.item()
current_IOU = calc_IOU(outputMaps,targets)
total_IOU += current_IOU
del(images)
del(targets)
progress_bar(batch_idx, len(self.trainLoader), 'Loss: %.3f | IOU: %.3f'
% (train_loss/(batch_idx+1), current_IOU))
self.writer.add_scalar('Train/Loss', train_loss/batch_idx+1, epoch)
self.writer.add_scalar('Train/IOU', total_IOU/batch_idx+1, epoch)
def validationTest(self,epoch):
self.net.eval()
validationLoss = []
total_IOU = []
with torch.no_grad():
for batch_idx, (images,targets) in enumerate(self.validLoader):
images = images.to(self.device)
targets = targets.to(self.device)
outputMaps = self.net(images)
loss = self.loss(outputMaps,targets)
currentValidationLoss = loss.item()
validationLoss.append(currentValidationLoss)
current_IOU = calc_IOU(outputMaps,targets)
total_IOU.append(current_IOU)
# progress_bar(batch_idx, len(self.validLoader), 'Loss: %.3f | IOU: %.3f' % (currentValidationLoss), current_IOU)
del(images)
del(targets)
meanIOU = np.mean(total_IOU)
meanValidationLoss = np.mean(validationLoss)
self.writer.add_scalar('Validation/Loss', meanValidationLoss, epoch)
self.writer.add_scalar('Validation/IOU', meanIOU, epoch)
print("VALIDATION LOSS: ",meanValidationLoss)
return meanIOU
def test(self,dataLoader):
self.net.eval()
testLoss = []
total_IOU = []
total_outputs_maps = []
total_input_images = []
with torch.no_grad():
for batch_idx, (images,targets) in enumerate(dataLoader):
images = images.to(self.device)
targets = targets.to(self.device)
outputMaps = self.net(images)
loss = self.loss(outputMaps,targets)
testLoss.append(loss.item())
current_IOU = calc_IOU(outputMaps,targets)
total_IOU.append(current_IOU)
total_outputs_maps.append(outputMaps.cpu().detach().numpy())
# total_input_images.append(transforms.ToPILImage()(images))
total_input_images.append(images.cpu().detach().numpy())
del(images)
del(targets)
break
meanIOU = np.mean(total_IOU)
meanLoss = np.mean(testLoss)
print("TEST IOU: ",meanIOU)
print("TEST LOSS: ",meanLoss)
return total_input_images,total_outputs_maps
def saveCheckPoint(self,state,isBest):
filename = "model.pth"
ckpt_path = os.path.join(self.saveModelDir, filename)
torch.save(state, ckpt_path)
if isBest:
filename = "best_model.pth"
shutil.copyfile(ckpt_path, os.path.join(self.saveModelDir, filename))
def loadCheckpoint(self):
print("[*] Loading model from {}".format(self.saveModelDir))
if(self.bestModel):
print("LOADING BEST MODEL")
filename = "best_model.pth"
else:
filename = "model.pth"
ckpt_path = os.path.join(self.saveModelDir, filename)
print(ckpt_path)
if(self.useGpu==False):
self.net=torch.load(ckpt_path, map_location=lambda storage, loc: storage)
else:
print("*"*40+" LOADING MODEL FROM GPU "+"*"*40)
self.ckpt = torch.load(ckpt_path)
self.net.load_state_dict(self.ckpt['model_state'])
self.net.cuda()
# Returns the number of GPUs available.
torch.cuda.device_count()
# 1
# Gets the name of a device.
print(torch.cuda.get_device_name(0))
torch.cuda.device(0)
device = 0
# unet = UNet().to(device)
unet = UNet(n_class=1).to(device=cuda)
criterion = nn.BCEWithLogitsLoss()
learning_rates = 1e-5
optimizer = torch.optim.Adam(unet.parameters(), lr=args.learning_rate)
unet.train()
train_dataloader, val_dataloader = basic_dataloader(args.input_size, args.batch_size, args.num_workers)
curr_lr = args.learning_rate
print("Initializing Training!")
# save_path = 'C:/Users/USER/Desktop/hand/model_save/'
# dir_root = 'C:/Users/USER/Desktop/segmentation/'
# dir_img = dir_root+ 'data/x/train/'
# dir_mask = dir_root+ 'data/y/label/'
dir_checkpoint = 'C:/Users/USER/Desktop/segmentation/checkpoints/'
def trainUnet(dirP, name, setLen, epochs=20):
class WPCDEDataset(Dataset):
def __init__(self, lenG, root_dir, transform=None):
self.root_dir = root_dir
self.lenG = lenG
self.transform = transform
def __len__(self):
return self.lenG
def __getitem__(self, idx):
img_nameX = self.root_dir + '%dx.jpg' % (idx)
img_nameY = self.root_dir + '%dy.jpg' % (idx)
imageX = Image.open(img_nameX).convert('RGB')
imageY = Image.open(img_nameY).convert('RGB')
if self.transform:
imageX = self.transform(imageX)
imageY = self.transform(imageY)
return imageX, imageY
transf = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
trainDatas = WPCDEDataset(setLen, dirP + r'\\' + 'train', transform=transf)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
batch_size = 16
lr = 2e-4
weight_decay = 0
start_epoch = 0
outf = r"..\model"
unet = UNet(in_channels=3, out_channels=3)
unet.to(device)
optimizer = optim.Adam(list(unet.parameters()),
lr=lr,
weight_decay=weight_decay)
dataloaderT = torch.utils.data.DataLoader(trainDatas,
batch_size=batch_size,
shuffle=True,
num_workers=int(0))
# dataloaderV = torch.utils.data.DataLoader(valDatas, batch_size=batch_size,
# shuffle=True, num_workers=int(0))
dataSplit = None # reserved rate of data
for epoch in range(start_epoch, start_epoch + epochs):
unet.train()
for i, (x, y) in enumerate(dataloaderT):
if dataSplit is not None:
if i > len(dataloaderT) * dataSplit:
break
x = x.to(device)
y = y.to(device)
optimizer.zero_grad()
ypred = unet(x)
loss = F.mse_loss(y, ypred)
loss.backward()
optimizer.step()
# break
if (i) % int(len(dataloaderT) / 4) == 0:
print('[%d/%d][%d/%d]\tLoss: %.4f\t ' %
(epoch, start_epoch + epochs, i, len(dataloaderT), loss))
state = {
'model': unet.state_dict(),
'optimizer': optimizer.state_dict(),
'epoch': epoch
}
torch.save(state,
'%s/UnetS%d%sepoch%d.pth' % (outf, setLen, name, epoch))
## tensorboard 를 사용하기 위한 SummaryWriter 설정
writer_train = SummaryWriter(log_dir = os.path.join(log_dir, 'train'))
writer_val = SummaryWriter(log_dir = os.path.join(log_dir, 'val'))
## 네트워크 학습시키기
st_epoch = 0
# TRAIN MODE
if mode == 'train':
if train_continue == "off":
net, optim, st_epoch = load(ckpt_dir = ckpt_dir, net = net, optim = optim)
for epoch in range(st_epoch + 1, num_epoch + 1):
net.train()
loss_arr = []
for batch, data in enumerate(loader_train, start = 1):
# forward pass
label = data['label'].to(device)
input = data['input'].to(device)
output = net(input)
# backward pass
optim.zero_grad()
loss = fn_loss(output, label)
loss.backward()
class Trainer(object):
"""Trainer for training and testing the model"""
def __init__(self, data_loader, config):
"""Initialize configurations"""
# model configuration
self.in_dim = config.in_dim
self.out_dim = config.out_dim
self.num_filters = config.num_filters
self.patch_size = config.patch_size
# training configuration
self.batch_size = config.batch_size
self.num_iters = config.num_iters
self.lr = config.lr
self.beta1 = config.beta1
self.beta2 = config.beta2
self.weight_decay = config.weight_decay
self.resume_iters = config.resume_iters
self.mode = config.mode
# miscellaneous.
self.use_tensorboard = config.use_tensorboard
self.use_cuda = torch.cuda.is_available()
self.device = torch.device('cuda:{}'.format(config.device_id) \
if self.use_cuda else 'cpu')
# training result configuration
self.log_dir = config.log_dir
self.log_step = config.log_step
self.model_save_dir = config.model_save_dir
self.model_save_step = config.model_save_step
# Build the model and tensorboard.
self.build_model()
if self.use_tensorboard:
self.build_tensorboard()
# data loader
if self.mode == 'train' or self.mode == 'test':
self.data_loader = data_loader
else:
self.train_data_loader, self.test_data_loader = data_loader
def build_model(self):
"""Create a model"""
self.model = UNet(self.in_dim, self.out_dim, self.num_filters)
self.model = self.model.float()
self.optimizer = torch.optim.Adam(self.model.parameters(),
self.lr, [self.beta1, self.beta2],
weight_decay=self.weight_decay)
self.print_network(self.model, 'unet')
self.model.to(self.device)
def _load(self, checkpoint_path):
if self.use_cuda:
checkpoint = torch.load(checkpoint_path)
else:
checkpoint = torch.load(checkpoint_path,
map_location=lambda storage, loc: storage)
return checkpoint
def restore_model(self, resume_iters):
"""Restore the trained model"""
print(
'Loading the trained models from step {}...'.format(resume_iters))
model_path = os.path.join(self.model_save_dir,
'{}-unet'.format(resume_iters) + '.ckpt')
checkpoint = self._load(model_path)
self.model.load_state_dict(checkpoint['model'])
self.optimizer.load_state_dict(checkpoint['optimizer'])
def print_network(self, model, name):
"""Print out the network information"""
num_params = 0
for p in model.parameters():
num_params += p.numel()
#print(model)
print(name)
print("The number of parameters: {}".format(num_params))
def print_optimizer(self, opt, name):
"""Print out optimizer information"""
print(opt)
print(name)
def build_tensorboard(self):
"""Build tensorboard for visualization"""
from logger import Logger
self.logger = Logger(self.log_dir)
def reset_grad(self):
"""Reset the gradient buffers."""
self.optimizer.zero_grad()
def train(self):
"""Train model"""
if self.mode != 'train_test':
data_loader = self.data_loader
else:
data_loader = self.train_data_loader
print("current dataset size: ", len(data_loader))
data_iter = iter(data_loader)
if not os.path.exists(self.model_save_dir):
os.makedirs(self.model_save_dir)
# start training from scratch or resume training.
start_iters = 0
if self.resume_iters:
print('Resuming ...')
start_iters = self.resume_iters
self.restore_model(self.resume_iters)
self.print_optimizer(self.optimizer, 'optimizer')
# print learning rate information
lr = self.lr
print('Current learning rates, g_lr: {}.'.format(lr))
# start training.
print('Start training...')
start_time = time.time()
for i in range(start_iters, self.num_iters):
# fetch batch data
try:
in_data, label = next(data_iter)
except:
data_iter = iter(data_loader)
in_data, label, _, _, _ = next(data_iter)
in_data = in_data.float().to(self.device)
label = label.to(self.device)
# train the model
self.model = self.model.train()
y_out = self.model(in_data)
loss = nn.BCEWithLogitsLoss()
output = loss(y_out, label)
self.reset_grad()
output.backward()
self.optimizer.step()
# logging
if (i + 1) % self.log_step == 0:
et = time.time() - start_time
et = str(datetime.timedelta(seconds=et))[:-7]
log = "Elapsed [{}], Iteration [{}/{}]".format(
et, i + 1, self.num_iters)
log += ", {}: {:.4f}".format("loss", output.mean().item())
print(log)
if self.use_tensorboard:
self.logger.scalar_summary("loss",
output.mean().item(), i + 1)
# save model checkpoints
if (i + 1) % self.model_save_step == 0:
path = os.path.join(self.model_save_dir,
'{}-unet'.format(i + 1) + '.ckpt')
torch.save(
{
'model': self.model.state_dict(),
'optimizer': self.optimizer.state_dict()
}, path)
print('Saved model checkpoints into {}...'.format(
self.model_save_dir))
def test(self):
"""Test model"""
if self.mode != 'train_test':
data_loader = self.data_loader
else:
data_loader = self.test_data_loader
print("current dataset size: ", len(data_loader))
data_iter = iter(data_loader)
# start testing on trained model
if self.resume_iters and self.mode != 'train_test':
print('Resuming ...')
self.restore_model(self.resume_iters)
# start testing.
result, trace = np.zeros((78, 110, 24)), np.zeros((78, 110, 24))
print('Start testing...')
correct, total, bcorrect = 0, 0, 0
while (True):
# fetch batch data
try:
data_in, label, i, j, k = next(data_iter)
except:
break
data_in = data_in.float().to(self.device)
label = label.float().to(self.device)
# test the model
self.model = self.model.eval()
y_hat = self.model(data_in)
m = nn.Sigmoid()
y_hat = m(y_hat)
y_hat = y_hat.squeeze().detach().cpu().numpy()
label = label.cpu().numpy().astype(int)
y_hat_th = (y_hat > 0.2)
label = (label > 0.5)
test = (label == y_hat_th)
correct += np.sum(test)
btest = (label == 0)
bcorrect += np.sum(btest)
total += y_hat_th.size
radius = int(self.patch_size / 2)
for step in range(self.batch_size):
x, y, z, pred = i[step], j[step], k[step], np.squeeze(
y_hat_th[step, :, :, :])
result[x - radius:x + radius, y - radius:y + radius,
z - radius:z + radius] += pred
trace[x - radius:x + radius, y - radius:y + radius,
z - radius:z + radius] += np.ones(
(self.patch_size, self.patch_size, self.patch_size))
print('Accuracy: %.3f%%' % (correct / total * 100))
print('Baseline Accuracy: %.3f%%' % (bcorrect / total * 100))
trace += (trace == 0)
result = result / trace
scipy.io.savemat('prediction.mat', {'result': result})
def train_test(self):
"""Train and test model"""
self.train()
self.test()
model.load_state_dict(
torch.load(os.path.join(args.load_model_dir, args.load_model_name)))
initial_epoch = findLastCheckpoint(
save_dir=save_dir) # load the last model in matconvnet style
# initial_epoch = 150
if initial_epoch > 0:
print('resuming by loading epoch %03d' % initial_epoch)
u_model.load_state_dict(
torch.load(os.path.join(save_dir,
'model_%03d.pth' % initial_epoch)))
# model = torch.load(os.path.join(save_dir, 'model_%03d.pth' % initial_epoch))
model.eval()
u_model.train()
criterion = nn.MSELoss()
if cuda:
model = model.cuda()
u_model = u_model.cuda()
optimizer = optim.Adam(model.parameters(), lr=args.lr)
scheduler = MultiStepLR(optimizer, milestones=[30, 60, 90],
gamma=0.2) # learning rates
for epoch in range(initial_epoch, n_epoch):
scheduler.step(epoch) # step to the learning rate in this epcoh
xs = dg.datagenerator(data_dir=args.train_data)
xs = xs.astype('float32') / 255.0
xs = torch.from_numpy(xs.transpose(
# Tensorboard
train_writer = SummaryWriter(log_dir=TRAIN_LOG_DIR)
val_writer = SummaryWriter(log_dir=VAL_LOG_DIR)
# Training
start_epoch = 0
# Load Checkpoint File
if os.listdir(CKPT_DIR):
net, optim, start_epoch = load_net(ckpt_dir=CKPT_DIR, net=net, optim=optim)
else:
print('* Training from scratch')
num_epochs = cfg.NUM_EPOCHS
for epoch in range(start_epoch + 1, num_epochs + 1):
net.train() # Train Mode
train_loss_arr = list()
for batch_idx, data in enumerate(train_loader, 1):
# Forward Propagation
img = data['img'].to(device)
label = data['label'].to(device)
output = net(img)
# Backward Propagation
optim.zero_grad()
loss = loss_fn(output, label)
loss.backward()
def UNet(self):
if self.mode == 'train':
transform = transforms.Compose([Normalization(mean=0.5, std=0.5, mode='train'), ToTensor()])
dataset_train = Dataset(mode = self.mode, data_dir=self.data_dir, image_type = self.image_type, transform=transform)
loader_train = DataLoader(dataset_train, batch_size=self.batch_size, shuffle=True, num_workers=8)
# dataset_val = Dataset(data_dir=os.path.join(data_dir, 'val'), transform=transform)
# loader_val = DataLoader(dataset_val, batch_size=batch_size, shuffle=False, num_workers=8)
# 그밖에 부수적인 variables 설정하기
num_data_train = len(dataset_train)
# num_data_val = len(dataset_val)
num_batch_train = np.ceil(num_data_train / self.batch_size)
# num_batch_val = np.ceil(num_data_val / batch_size)
elif self.mode == 'test':
transform = transforms.Compose([Normalization(mean=0.5, std=0.5, mode='test'), ToTensor()])
dataset_test = Dataset(mode = self.mode, data_dir=self.data_dir, image_type = self.image_type, transform=transform)
loader_test = DataLoader(dataset_test, batch_size=self.batch_size, shuffle=False, num_workers=8)
# 그밖에 부수적인 variables 설정하기
num_data_test = len(dataset_test)
num_batch_test = np.ceil(num_data_test / self.batch_size)
fn_tonumpy = lambda x: x.to('cpu').detach().numpy().transpose(0, 2, 3, 1)
fn_denorm = lambda x, mean, std: (x * std) + mean
fn_class = lambda x: 1.0 * (x > 0.5)
net = UNet().to(self.device)
criterion = torch.nn.MSELoss().to(self.device)
optimizer = torch.optim.Adam(net.parameters(), lr=self.lr)
writer_train = SummaryWriter(log_dir=os.path.join(self.log_dir, 'train'))
if self.mode == 'train':
if self.train_continue == "on":
net, optimizer = load_model(ckpt_dir=self.ckpt_dir, net=net, optim=optimizer)
for epoch in range(1, self.num_epoch + 1):
net.train()
loss_arr = []
for batch, data in enumerate(loader_train, 1):
# forward pass
label = data['label'].to(self.device)
input = data['input'].to(self.device)
output = net(input)
# backward pass
optimizer.zero_grad()
loss = criterion(output, label)
loss.backward()
optimizer.step()
# 손실함수 계산
loss_arr += [loss.item()]
# Tensorboard 저장하기
label = fn_tonumpy(label)
input = fn_tonumpy(fn_denorm(input, mean=0.5, std=0.5))
output = fn_tonumpy(fn_class(output))
writer_train.add_image('label', label, num_batch_train * (epoch - 1) + batch, dataformats='NHWC')
writer_train.add_image('input', input, num_batch_train * (epoch - 1) + batch, dataformats='NHWC')
writer_train.add_image('output', output, num_batch_train * (epoch - 1) + batch, dataformats='NHWC')
writer_train.add_scalar('loss', np.mean(loss_arr), epoch)
print("TRAIN: EPOCH %04d / %04d | LOSS %.4f" %(epoch, self.num_epoch, np.mean(loss_arr)))
if epoch % 20 == 0:
save_model(ckpt_dir=self.ckpt_dir, net=net, optim=optimizer, epoch=0)
writer_train.close()
# TEST MODE
elif self.mode == 'test':
net, optimizer = load_model(ckpt_dir=self.ckpt_dir, net=net, optim=optimizer)
with torch.no_grad():
net.eval()
loss_arr = []
id = 1
for batch, data in enumerate(loader_test, 1):
# forward pass
input = data['input'].to(self.device)
output = net(input)
# 손실함수 계산하기
#loss = criterion(output, label)
#loss_arr += [loss.item()]
#print("TEST: BATCH %04d / %04d | " %
# (batch, num_batch_test))
# Tensorboard 저장하기
output = fn_tonumpy(fn_class(output))
for j in range(input.shape[0]):
if id == 800:
id = 2350
print(id)
#plt.imsave(os.path.join(self.result_dir, 'png', 'label_%04d.png' % id), label[j].squeeze(), cmap='gray')
#plt.imsave(os.path.join(self.result_dir, 'png', 'input_%04d.png' % id), input[j].squeeze(), cmap='gray')
plt.imsave(os.path.join(self.result_dir, 'png', 'gt%06d.png' % id), output[j].squeeze(), cmap='gray')
id+=1
# np.save(os.path.join(result_dir, 'numpy', 'label_%04d.npy' % id), label[j].squeeze())
# np.save(os.path.join(result_dir, 'numpy', 'input_%04d.npy' % id), input[j].squeeze())
# np.save(os.path.join(result_dir, 'numpy', 'output_%04d.npy' % id), output[j].squeeze())
print("AVERAGE TEST: BATCH %04d / %04d | LOSS %.4f" %
(batch, num_batch_test, np.mean(loss_arr)))
def main(args):
writer = SummaryWriter(os.path.join('./logs'))
# torch.backends.cudnn.benchmark = False
# if not os.path.isdir(args.checkpoint_dir):
# os.mkdir(args.checkpoint_dir)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print('[MODEL] CUDA DEVICE : {}'.format(device))
# TODO DEFINE TRAIN AND TEST TRANSFORMS
train_tf = None
test_tf = None
# Channel wise mean calculated on adobe240-fps training dataset
mean = [0.429, 0.431, 0.397]
std = [1, 1, 1]
normalize = transforms.Normalize(mean=mean,
std=std)
transform = transforms.Compose([transforms.ToTensor(), normalize])
test_valid = 'validation' if args.valid else 'test'
# train_data = BlurDataset(os.path.join(args.dataset_root, 'train'),
# seq_len=args.sequence_length, tau=args.num_frame_blur, delta=5, transform=train_tf)
test_data = BlurDataset(os.path.join(args.dataset_root, test_valid),
seq_len=args.sequence_length, tau=args.num_frame_blur, delta=5, transform=train_tf, return_path=True)
# train_loader = DataLoader(train_data, batch_size=args.train_batch_size, shuffle=True)
test_loader = DataLoader(test_data, batch_size=args.test_batch_size, shuffle=False)
# TODO IMPORT YOUR CUSTOM MODEL
model = UNet(3, 3, device, decode_mode=args.decode_mode)
if args.checkpoint:
store_dict = torch.load(args.checkpoint)
try:
print('Loading checkpoint...')
model.load_state_dict(store_dict['state_dict'])
print('Done.')
except KeyError:
print('Loading checkpoint...')
model.load_state_dict(store_dict)
print('Done.')
model.to(device)
model.train(False)
# model = nn.DataParallel(model)
# TODO DEFINE MORE CRITERIA
# input(True if device == torch.device('cuda:0') else False)
criterion = {
'MSE': nn.MSELoss(),
'L1' : nn.L1Loss(),
# 'Perceptual': PerceptualLoss(model='net-lin', net='vgg', dataparallel=False,
# use_gpu=True if device == torch.device('cuda:0') else False)
}
# Validation
running_loss_test = 0.0
psnr_test = 0.0
dssim_test = 0.0
tqdm_loader_test = tqdm(range(len(test_loader)), ncols=150)
loss_tracker_test = {}
for loss_fn in criterion.keys():
loss_tracker_test[loss_fn] = 0.0
with torch.no_grad():
model.eval()
total_steps_test = 0.0
interp_idx = int(math.ceil((args.num_frame_blur / 2) - 0.49))
for test_idx, data in enumerate(test_loader, 1):
loss = 0.0
blur_data, sharpe_data, sharp_names = data
import pdb; pdb.set_trace()
interp_idx = int(math.ceil((args.num_frame_blur / 2) - 0.49))
# input(interp_idx)
if args.decode_mode == 'interp':
sharpe_data = sharpe_data[:, :, 1::2, :, :]
elif args.decode_mode == 'deblur':
sharpe_data = sharpe_data[:, :, 0::2, :, :]
else:
# print('\nBoth\n')
sharpe_data = sharpe_data
# print(sharpe_data.shape)
# input(blur_data.shape)
blur_data = blur_data.to(device)[:, :, :, :352, :].permute(0, 1, 2, 4, 3)
try:
sharpe_data = sharpe_data.squeeze().to(device)[:, :, :, :352, :].permute(0, 1, 2, 4, 3)
except:
sharpe_data = sharpe_data.squeeze(3).to(device)[:, :, :, :352, :].permute(0, 1, 2, 4, 3)
# forward pass
sharpe_out = model(blur_data).float()
# compute losses
sharpe_out = sharpe_out.permute(0, 2, 1, 3, 4)
B, C, S, Fx, Fy = sharpe_out.shape
for loss_fn in criterion.keys():
loss_tmp = 0.0
if loss_fn == 'Perceptual':
for bidx in range(B):
loss_tmp += criterion_w[loss_fn] * \
criterion[loss_fn](sharpe_out[bidx].permute(1, 0, 2, 3),
sharpe_data[bidx].permute(1, 0, 2, 3)).sum()
# loss_tmp /= B
else:
loss_tmp = criterion_w[loss_fn] * \
criterion[loss_fn](sharpe_out, sharpe_data)
loss += loss_tmp
try:
loss_tracker_test[loss_fn] += loss_tmp.item()
except KeyError:
loss_tracker_test[loss_fn] = loss_tmp.item()
# statistics
#sharpe_out = sharpe_out.detach().cpu().numpy()
#sharpe_data = sharpe_data.cpu().numpy()
# import pdb; pdb.set_trace()
# t_grid = torchvision.utils.make_grid(torch.stack([blur_data[0], sharpe_out[0], sharpe_data[0]], dim=0),
# nrow=3)
# tsave(t_grid, './imgs/{}/combined.jpg'.format(test_idx))
for sidx in range(S):
for bidx in range(B):
if not os.path.exists('./imgs/{}'.format(sharp_names[1])):
os.makedirs('./imgs/{}'.format(test_idx))
blur_path = './imgs/{}/blur_input_{}.jpg'.format(test_idx, sidx)
# import pdb; pdb.set_trace()
# torchvision.utils.save_image(sharpe_out[bidx, :, sidx, :, :],blur_path, normalize=True, range=(0,255));
imsave(blur_data, blur_path, bidx, sidx)
sharp_path = './imgs/{}/sharpe_gt_{}{}.jpg'.format(test_idx, sidx, sidx)
imsave(sharpe_data, sharp_path, bidx, sidx)
deblur_path = './imgs/{}/out_{}{}.jpg'.format(test_idx, sidx, sidx)
imsave(sharpe_out, deblur_path, bidx, sidx)
if sidx > 0 and sidx < S:
interp_path = './imgs/{}/out_{}{}.jpg'.format(test_idx, sidx-1, sidx)
imsave(sharpe_out, interp_path, bidx, sidx)
sharp_path = './imgs/{}/sharpe_gt_{}{}.jpg'.format(test_idx, sidx-1, sidx)
imsave(sharpe_data, sharp_path, bidx, sidx)
psnr_local = psnr(im_nm * sharpe_out[bidx, :, sidx, :, :].detach().cpu().numpy(),
im_nm * sharpe_data[bidx, :, sidx, :, :].cpu().numpy())
dssim_local = dssim(np.moveaxis(im_nm * sharpe_out[bidx, :, sidx, :, :].cpu().numpy(), 0, 2),
np.moveaxis(im_nm * sharpe_data[bidx, :, sidx, :, :].cpu().numpy(), 0, 2)
)
psnr_test += psnr_local
dssim_test += dssim_local
f = open('./imgs/{0}/psnr-{1:.4f}-dssim-{2:.4f}.txt'.format(test_idx, psnr_local/(B), dssim_local/(B)),'w')
f.close()
running_loss_test += loss.item()
total_steps_test += B*S
loss_str = ''
for key in loss_tracker_test.keys():
loss_str += ' {0} : {1:6.4f} '.format(key, 1.0 * loss_tracker_test[key] / total_steps_test)
# set display info
tqdm_loader_test.set_description(
('\r[Test ] loss: {0:6.4f} PSNR: {1:6.4f} SSIM: {2:6.4f} '.format
( running_loss_test / total_steps_test,
psnr_test / total_steps_test,
dssim_test / total_steps_test
) + loss_str
)
)
tqdm_loader_test.update(1)
tqdm_loader_test.close()
return None
def main(args):
writer = SummaryWriter(os.path.join('./logs'))
# torch.backends.cudnn.benchmark = True
if not os.path.isdir(args.checkpoint_dir):
os.mkdir(args.checkpoint_dir)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print('[MODEL] CUDA DEVICE : {}'.format(device))
# TODO DEFINE TRAIN AND TEST TRANSFORMS
train_tf = None
test_tf = None
# Channel wise mean calculated on adobe240-fps training dataset
mean = [0.429, 0.431, 0.397]
std = [1, 1, 1]
normalize = transforms.Normalize(mean=mean, std=std)
transform = transforms.Compose([transforms.ToTensor(), normalize])
test_valid = 'validation' if args.valid else 'test'
train_data = BlurDataset(os.path.join(args.dataset_root, 'train'),
seq_len=args.sequence_length,
tau=args.num_frame_blur,
delta=5,
transform=train_tf)
test_data = BlurDataset(os.path.join(args.dataset_root, test_valid),
seq_len=args.sequence_length,
tau=args.num_frame_blur,
delta=5,
transform=train_tf)
train_loader = DataLoader(train_data,
batch_size=args.train_batch_size,
shuffle=True)
test_loader = DataLoader(test_data,
batch_size=args.test_batch_size,
shuffle=False)
# TODO IMPORT YOUR CUSTOM MODEL
model = UNet(3, 3, device, decode_mode=args.decode_mode)
if args.checkpoint:
store_dict = torch.load(args.checkpoint)
try:
model.load_state_dict(store_dict['state_dict'])
except KeyError:
model.load_state_dict(store_dict)
if args.train_continue:
store_dict = torch.load(args.checkpoint)
model.load_state_dict(store_dict['state_dict'])
else:
store_dict = {'loss': [], 'valLoss': [], 'valPSNR': [], 'epoch': -1}
model.to(device)
model.train(True)
# model = nn.DataParallel(model)
# TODO DEFINE MORE CRITERIA
# input(True if device == torch.device('cuda:0') else False)
criterion = {
'MSE': nn.MSELoss(),
'L1': nn.L1Loss(),
# 'Perceptual': PerceptualLoss(model='net-lin', net='vgg', dataparallel=True,
# use_gpu=True if device == torch.device('cuda:0') else False)
}
criterion_w = {'MSE': 1.0, 'L1': 10.0, 'Perceptual': 10.0}
# Define optimizers
# optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.9,weight_decay=5e-4)
optimizer = optim.Adam(model.parameters(), lr=args.init_learning_rate)
# Define lr scheduler
scheduler = optim.lr_scheduler.MultiStepLR(optimizer,
milestones=args.milestones,
gamma=0.1)
# best_acc = 0.0
# start = time.time()
cLoss = store_dict['loss']
valLoss = store_dict['valLoss']
valPSNR = store_dict['valPSNR']
checkpoint_counter = 0
loss_tracker = {}
loss_tracker_test = {}
psnr_old = 0.0
dssim_old = 0.0
for epoch in range(1, 10 *
args.epochs): # loop over the dataset multiple times
# Append and reset
cLoss.append([])
valLoss.append([])
valPSNR.append([])
running_loss = 0
# Increment scheduler count
scheduler.step()
tqdm_loader = tqdm(range(len(train_loader)), ncols=150)
loss = 0.0
psnr_ = 0.0
dssim_ = 0.0
loss_tracker = {}
for loss_fn in criterion.keys():
loss_tracker[loss_fn] = 0.0
# Train
model.train(True)
total_steps = 0.01
total_steps_test = 0.01
'''for train_idx, data in enumerate(train_loader, 1):
loss = 0.0
blur_data, sharpe_data = data
#import pdb; pdb.set_trace()
# input(sharpe_data.shape)
#import pdb; pdb.set_trace()
interp_idx = int(math.ceil((args.num_frame_blur/2) - 0.49))
#input(interp_idx)
if args.decode_mode == 'interp':
sharpe_data = sharpe_data[:, :, 1::2, :, :]
elif args.decode_mode == 'deblur':
sharpe_data = sharpe_data[:, :, 0::2, :, :]
else:
#print('\nBoth\n')
sharpe_data = sharpe_data
#print(sharpe_data.shape)
#input(blur_data.shape)
blur_data = blur_data.to(device)[:, :, :, :352, :].permute(0, 1, 2, 4, 3)
try:
sharpe_data = sharpe_data.squeeze().to(device)[:, :, :, :352, :].permute(0, 1, 2, 4, 3)
except:
sharpe_data = sharpe_data.squeeze(3).to(device)[:, :, :, :352, :].permute(0, 1, 2, 4, 3)
# clear gradient
optimizer.zero_grad()
# forward pass
sharpe_out = model(blur_data)
# import pdb; pdb.set_trace()
# input(sharpe_out.shape)
# compute losses
# import pdb;
# pdb.set_trace()
sharpe_out = sharpe_out.permute(0, 2, 1, 3, 4)
B, C, S, Fx, Fy = sharpe_out.shape
for loss_fn in criterion.keys():
loss_tmp = 0.0
if loss_fn == 'Perceptual':
for bidx in range(B):
loss_tmp += criterion_w[loss_fn] * \
criterion[loss_fn](sharpe_out[bidx].permute(1, 0, 2, 3),
sharpe_data[bidx].permute(1, 0, 2, 3)).sum()
# loss_tmp /= B
else:
loss_tmp = criterion_w[loss_fn] * \
criterion[loss_fn](sharpe_out, sharpe_data)
# try:
# import pdb; pdb.set_trace()
loss += loss_tmp # if
# except :
try:
loss_tracker[loss_fn] += loss_tmp.item()
except KeyError:
loss_tracker[loss_fn] = loss_tmp.item()
# Backpropagate
loss.backward()
optimizer.step()
# statistics
# import pdb; pdb.set_trace()
sharpe_out = sharpe_out.detach().cpu().numpy()
sharpe_data = sharpe_data.cpu().numpy()
for sidx in range(S):
for bidx in range(B):
psnr_ += psnr(sharpe_out[bidx, :, sidx, :, :], sharpe_data[bidx, :, sidx, :, :]) #, peak=1.0)
"""dssim_ += dssim(np.moveaxis(sharpe_out[bidx, :, sidx, :, :], 0, 2),
np.moveaxis(sharpe_data[bidx, :, sidx, :, :], 0, 2)
)"""
"""sharpe_out = sharpe_out.reshape(-1,3, sx, sy).detach().cpu().numpy()
sharpe_data = sharpe_data.reshape(-1, 3, sx, sy).cpu().numpy()
for idx in range(sharpe_out.shape[0]):
# import pdb; pdb.set_trace()
psnr_ += psnr(sharpe_data[idx], sharpe_out[idx])
dssim_ += dssim(np.swapaxes(sharpe_data[idx], 2, 0), np.swapaxes(sharpe_out[idx], 2, 0))"""
# psnr_ /= sharpe_out.shape[0]
# dssim_ /= sharpe_out.shape[0]
running_loss += loss.item()
loss_str = ''
total_steps += B*S
for key in loss_tracker.keys():
loss_str += ' {0} : {1:6.4f} '.format(key, 1.0*loss_tracker[key] / total_steps)
# set display info
if train_idx % 5 == 0:
tqdm_loader.set_description(('\r[Training] [Ep {0:6d}] loss: {1:6.4f} PSNR: {2:6.4f} SSIM: {3:6.4f} '.format
(epoch, running_loss / total_steps,
psnr_ / total_steps,
dssim_ / total_steps) + loss_str
))
tqdm_loader.update(5)
tqdm_loader.close()'''
# Validation
running_loss_test = 0.0
psnr_test = 0.0
dssim_test = 0.0
# print('len', len(test_loader))
tqdm_loader_test = tqdm(range(len(test_loader)), ncols=150)
# import pdb; pdb.set_trace()
loss_tracker_test = {}
for loss_fn in criterion.keys():
loss_tracker_test[loss_fn] = 0.0
with torch.no_grad():
model.eval()
total_steps_test = 0.0
for test_idx, data in enumerate(test_loader, 1):
loss = 0.0
blur_data, sharpe_data = data
interp_idx = int(math.ceil((args.num_frame_blur / 2) - 0.49))
# input(interp_idx)
if args.decode_mode == 'interp':
sharpe_data = sharpe_data[:, :, 1::2, :, :]
elif args.decode_mode == 'deblur':
sharpe_data = sharpe_data[:, :, 0::2, :, :]
else:
# print('\nBoth\n')
sharpe_data = sharpe_data
# print(sharpe_data.shape)
# input(blur_data.shape)
blur_data = blur_data.to(device)[:, :, :, :352, :].permute(
0, 1, 2, 4, 3)
try:
sharpe_data = sharpe_data.squeeze().to(
device)[:, :, :, :352, :].permute(0, 1, 2, 4, 3)
except:
sharpe_data = sharpe_data.squeeze(3).to(
device)[:, :, :, :352, :].permute(0, 1, 2, 4, 3)
# clear gradient
optimizer.zero_grad()
# forward pass
sharpe_out = model(blur_data)
# import pdb; pdb.set_trace()
# input(sharpe_out.shape)
# compute losses
sharpe_out = sharpe_out.permute(0, 2, 1, 3, 4)
B, C, S, Fx, Fy = sharpe_out.shape
for loss_fn in criterion.keys():
loss_tmp = 0.0
if loss_fn == 'Perceptual':
for bidx in range(B):
loss_tmp += criterion_w[loss_fn] * \
criterion[loss_fn](sharpe_out[bidx].permute(1, 0, 2, 3),
sharpe_data[bidx].permute(1, 0, 2, 3)).sum()
# loss_tmp /= B
else:
loss_tmp = criterion_w[loss_fn] * \
criterion[loss_fn](sharpe_out, sharpe_data)
loss += loss_tmp
try:
loss_tracker_test[loss_fn] += loss_tmp.item()
except KeyError:
loss_tracker_test[loss_fn] = loss_tmp.item()
if ((test_idx % args.progress_iter) == args.progress_iter - 1):
itr = test_idx + epoch * len(test_loader)
# itr_train
writer.add_scalars(
'Loss', {
'trainLoss': running_loss / total_steps,
'validationLoss':
running_loss_test / total_steps_test
}, itr)
writer.add_scalar('Train PSNR', psnr_ / total_steps, itr)
writer.add_scalar('Test PSNR',
psnr_test / total_steps_test, itr)
# import pdb; pdb.set_trace()
# writer.add_image('Validation', sharpe_out.permute(0, 2, 3, 1), itr)
# statistics
sharpe_out = sharpe_out.detach().cpu().numpy()
sharpe_data = sharpe_data.cpu().numpy()
for sidx in range(S):
for bidx in range(B):
psnr_test += psnr(
sharpe_out[bidx, :, sidx, :, :],
sharpe_data[bidx, :, sidx, :, :]) #, peak=1.0)
dssim_test += dssim(
np.moveaxis(sharpe_out[bidx, :, sidx, :, :], 0, 2),
np.moveaxis(sharpe_data[bidx, :, sidx, :, :], 0,
2)) #,range=1.0 )
running_loss_test += loss.item()
total_steps_test += B * S
loss_str = ''
for key in loss_tracker.keys():
loss_str += ' {0} : {1:6.4f} '.format(
key, 1.0 * loss_tracker_test[key] / total_steps_test)
# set display info
tqdm_loader_test.set_description((
'\r[Test ] [Ep {0:6d}] loss: {1:6.4f} PSNR: {2:6.4f} SSIM: {3:6.4f} '
.format(epoch, running_loss_test / total_steps_test,
psnr_test / total_steps_test,
dssim_test / total_steps_test) + loss_str))
tqdm_loader_test.update(1)
tqdm_loader_test.close()
# save model
if psnr_old < (psnr_test / total_steps_test):
if epoch != 1:
os.remove(
os.path.join(
args.checkpoint_dir,
'epoch-{}-test-psnr-{}-ssim-{}.ckpt'.format(
epoch_old,
str(round(psnr_old, 4)).replace('.', 'pt'),
str(round(dssim_old, 4)).replace('.', 'pt'))))
epoch_old = epoch
psnr_old = psnr_test / total_steps_test
dssim_old = dssim_test / total_steps_test
checkpoint_dict = {
'epoch': epoch_old,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
'train_psnr': psnr_ / total_steps,
'train_dssim': dssim_ / total_steps,
'train_mse': loss_tracker['MSE'] / total_steps,
'train_l1': loss_tracker['L1'] / total_steps,
# 'train_percp': loss_tracker['Perceptual'] / total_steps,
'test_psnr': psnr_old,
'test_dssim': dssim_old,
'test_mse': loss_tracker_test['MSE'] / total_steps_test,
'test_l1': loss_tracker_test['L1'] / total_steps_test,
# 'test_percp': loss_tracker_test['Perceptual'] / total_steps_test,
}
torch.save(
checkpoint_dict,
os.path.join(
args.checkpoint_dir,
'epoch-{}-test-psnr-{}-ssim-{}.ckpt'.format(
epoch_old,
str(round(psnr_old, 4)).replace('.', 'pt'),
str(round(dssim_old, 4)).replace('.', 'pt'))))
# if epoch % args.checkpoint_epoch == 0:
# torch.save(model.state_dict(),args.checkpoint_dir + str(int(epoch/100))+".ckpt")
return None
def train():
if not os.path.exists("train_model/"):
os.makedirs("train_model/")
if not os.path.exists("result/"):
os.makedirs("result/")
train_data, dev_data, word2id, id2word, char2id, opts = load_data(
vars(args))
model = UNet(opts)
if args.use_cuda:
model = model.cuda()
dev_batches = get_batches(dev_data, args.batch_size, evaluation=True)
if args.eval:
print("load model...")
model.load_state_dict(torch.load(args.model_dir))
model.eval()
model.Evaluate(
dev_batches,
os.path.join(args.prepro_dir, "dev_eval.json"),
answer_file="result/" + args.model_dir.split("/")[-1] + ".answers",
drop_file=os.path.join(args.prepro_dir, "drop.json"),
dev=args.dev_file,
)
exit()
if args.load_model:
print("load model...")
model.load_state_dict(torch.load(args.model_dir))
model.eval()
_, F1 = model.Evaluate(
dev_batches,
os.path.join(args.prepro_dir, "dev_eval.json"),
answer_file=os.path.join("result/",
args.model_dir.split("/")[-1],
".answers"),
drop_file=os.path.join(args.prepro_dir, "drop.json"),
dev=args.dev_file,
)
best_score = F1
with open(args.model_dir + "_f1_scores.pkl", "rb") as f:
f1_scores = pkl.load(f)
with open(args.model_dir + "_em_scores.pkl", "rb") as f:
em_scores = pkl.load(f)
else:
best_score = 0.0
f1_scores = []
em_scores = []
parameters = filter(lambda p: p.requires_grad, model.parameters())
optimizer = torch.optim.Adamax(parameters, lr=args.lrate)
lrate = args.lrate
for epoch in range(1, args.epochs + 1):
train_batches = get_batches(train_data, args.batch_size)
dev_batches = get_batches(dev_data, args.batch_size, evaluation=True)
total_size = len(train_data) // args.batch_size
model.train()
for i, train_batch in enumerate(train_batches):
loss = model(train_batch)
model.zero_grad()
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(parameters, opts["grad_clipping"])
optimizer.step()
model.reset_parameters()
if i % 100 == 0:
print(
"Epoch = %d, step = %d / %d, loss = %.5f, lrate = %.5f best_score = %.3f"
% (epoch, i, total_size, model.train_loss.value, lrate,
best_score))
sys.stdout.flush()
model.eval()
exact_match_score, F1 = model.Evaluate(
dev_batches,
os.path.join(args.prepro_dir, "dev_eval.json"),
answer_file=os.path.join("result/",
args.model_dir.split("/")[-1],
".answers"),
drop_file=os.path.join(args.prepro_dir, "drop.json"),
dev=args.dev_file,
)
f1_scores.append(F1)
em_scores.append(exact_match_score)
with open(args.model_dir + "_f1_scores.pkl", "wb") as f:
pkl.dump(f1_scores, f)
with open(args.model_dir + "_em_scores.pkl", "wb") as f:
pkl.dump(em_scores, f)
if best_score < F1:
best_score = F1
print("saving %s ..." % args.model_dir)
torch.save(model.state_dict(), args.model_dir)
if epoch > 0 and epoch % args.decay_period == 0:
lrate *= args.decay
for param_group in optimizer.param_groups:
param_group["lr"] = lrate
