print('step_now: {}, step loss: {}'.format(step_now,loss_cpu))
if step_now % save_step == 0:
val_loss = 0
torch.save(model.state_dict(), './model-step%d.pth' % (step_now))
'''
## FP16 !
for i in range(num_epoches):
for id, sample in enumerate(train_dataloader):
step_now += 1
img = sample["image"]
mask = sample["mask"]
optimizer.zero_grad()
with autocast():
pred = model(img.to(device))
loss = criterion(pred, mask.to(device))
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
loss_cpu = loss.cpu().item()
print('step_now: {}, step loss: {}'.format(step_now, loss_cpu))
total_loss += loss_cpu
if step_now % save_step == 0:
val_loss = 0
torch.save(model.state_dict(), './model-step%d.pth' % (step_now))
torch.cuda.synchronize()
end = time.time()
print("training time per step(ms): ", (end - start) * 1000 / (num_epoches * 2))
torch.save(model.state_dict(), './model-all.pth')
def main():
global args, best_prec1
args = parser.parse_args()
print(args)
if args.saveTest == 'True':
args.saveTest = True
elif args.saveTest == 'False':
args.saveTest = False
# Check if the save directory exists or not
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
cudnn.benchmark = True
data_transforms = {
'train': transforms.Compose([
transforms.Resize((args.imageSize, args.imageSize), interpolation=Image.NEAREST),
transforms.TenCrop(args.resizedImageSize),
transforms.Lambda(lambda crops: torch.stack([transforms.ToTensor()(crop) for crop in crops])),
#transforms.Lambda(lambda normalized: torch.stack([transforms.Normalize([0.295, 0.204, 0.197], [0.221, 0.188, 0.182])(crop) for crop in normalized]))
#transforms.RandomResizedCrop(224, interpolation=Image.NEAREST),
#transforms.RandomHorizontalFlip(),
#transforms.RandomVerticalFlip(),
#transforms.ToTensor(),
]),
'test': transforms.Compose([
transforms.Resize((args.imageSize, args.imageSize), interpolation=Image.NEAREST),
transforms.ToTensor(),
#transforms.Normalize([0.295, 0.204, 0.197], [0.221, 0.188, 0.182])
]),
}
# Data Loading
data_dir = 'datasets/miccaiSegRefined'
# json path for class definitions
json_path = 'datasets/miccaiSegClasses.json'
image_datasets = {x: miccaiSegDataset(os.path.join(data_dir, x), data_transforms[x],
json_path) for x in ['train', 'test']}
dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x],
batch_size=args.batchSize,
shuffle=True,
num_workers=args.workers)
for x in ['train', 'test']}
dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'test']}
# Get the dictionary for the id and RGB value pairs for the dataset
classes = image_datasets['train'].classes
key = utils.disentangleKey(classes)
num_classes = len(key)
# Initialize the model
model = UNet(num_classes)
# # Optionally resume from a checkpoint
# if args.resume:
# if os.path.isfile(args.resume):
# print("=> loading checkpoint '{}'".format(args.resume))
# checkpoint = torch.load(args.resume)
# #args.start_epoch = checkpoint['epoch']
# pretrained_dict = checkpoint['state_dict']
# pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model.state_dict()}
# model.state_dict().update(pretrained_dict)
# model.load_state_dict(model.state_dict())
# print("=> loaded checkpoint (epoch {})".format(checkpoint['epoch']))
# else:
# print("=> no checkpoint found at '{}'".format(args.resume))
#
# # # Freeze the encoder weights
# # for param in model.encoder.parameters():
# # param.requires_grad = False
#
# optimizer = optim.Adam(model.parameters(), lr = args.lr, weight_decay = args.wd)
# else:
optimizer = optim.Adam(model.parameters(), lr = args.lr, weight_decay = args.wd)
# Load the saved model
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['state_dict'])
print("=> loaded checkpoint (epoch {})".format(checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
print(model)
# Define loss function (criterion)
criterion = nn.CrossEntropyLoss()
# Use a learning rate scheduler
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=10, gamma=0.5)
if use_gpu:
model.cuda()
criterion.cuda()
# Initialize an evaluation Object
evaluator = utils.Evaluate(key, use_gpu)
for epoch in range(args.start_epoch, args.epochs):
#adjust_learning_rate(optimizer, epoch)
# Train for one epoch
print('>>>>>>>>>>>>>>>>>>>>>>>Training<<<<<<<<<<<<<<<<<<<<<<<')
train(dataloaders['train'], model, criterion, optimizer, scheduler, epoch, key)
# Evaulate on validation set
print('>>>>>>>>>>>>>>>>>>>>>>>Testing<<<<<<<<<<<<<<<<<<<<<<<')
validate(dataloaders['test'], model, criterion, epoch, key, evaluator)
# Calculate the metrics
print('>>>>>>>>>>>>>>>>>> Evaluating the Metrics <<<<<<<<<<<<<<<<<')
IoU = evaluator.getIoU()
print('Mean IoU: {}, Class-wise IoU: {}'.format(torch.mean(IoU), IoU))
PRF1 = evaluator.getPRF1()
precision, recall, F1 = PRF1[0], PRF1[1], PRF1[2]
print('Mean Precision: {}, Class-wise Precision: {}'.format(torch.mean(precision), precision))
print('Mean Recall: {}, Class-wise Recall: {}'.format(torch.mean(recall), recall))
print('Mean F1: {}, Class-wise F1: {}'.format(torch.mean(F1), F1))
evaluator.reset()
save_checkpoint({
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
}, filename=os.path.join(args.save_dir, 'checkpoint_{}.tar'.format(epoch)))
pass
torch.save(net.state_dict(),
dir_checkpoint + f'CP_epoch{epoch + 1}.pth')
logging.info(f'Checkpoint {epoch + 1} saved !')
writer.close()
if __name__ == '__main__':
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
net = UNet(n_channels=3, n_classes=1, bilinear=True)
net.to(device=device)
# faster convolutions, but more memory
# cudnn.benchmark = True
try:
train_net(net=net,
epochs=5,
batch_size=1,
lr=0.001,
device=device,
val_percent=10.0 / 100)
except KeyboardInterrupt:
torch.save(net.state_dict(), 'INTERRUPTED.pth')
logging.info('Saved interrupt')
try:
sys.exit(0)
except SystemExit:
os._exit(0)
# resume
if (osp.isfile(resume_path) and resume_flag):
checkpoint = torch.load(resume_path)
model.load_state_dict(checkpoint["model_state"])
optimizer.load_state_dict(checkpoint["optimizer_state"])
best_iou = checkpoint['best_iou']
# scheduler.load_state_dict(checkpoint["scheduler_state"])
# start_epoch = checkpoint["epoch"]
print(
"=====>",
"Loaded checkpoint '{}' (iter {})".format(resume_path,
checkpoint["epoch"]))
else:
print("=====>", "No checkpoint found at '{}'".format(resume_path))
print("load unet weight and bias")
model_dict = model.state_dict()
pretrained_dict = torch.load("/home/cv_xfwang/Pytorch-UNet/MODEL.pth")
pretrained_dict = {
k: v
for k, v in pretrained_dict.items() if k in model_dict
}
model_dict.update(pretrained_dict)
model.load_state_dict(model_dict)
# Training
def train(epoch):
print("Epoch: ", epoch)
model.train()
total_loss = 0
# for index, (img, mask) in tqdm(enumerate(train_loader)):
class BaseModel:
losses = {'train': [], 'val': []}
acces = {'train': [], 'val': []}
scores = {'train': [], 'val': []}
pred = {'train': [], 'val': []}
true = {'train': [], 'val': []}
def __init__(self, args):
self.args = args
self.net = None
print(args.model_name)
if args.model_name == 'UNet':
self.net = UNet(args.in_channels, args.num_classes)
self.net.apply(weights_init)
elif args.model_name == 'UNetResNet34':
self.net = UNetResNet34(args.num_classes, dropout_2d=0.2)
elif args.model_name == 'UNetResNet152':
self.net = UNetResNet152(args.num_classes, dropout_2d=0.2)
elif args.model_name == 'UNet11':
self.net = UNet11(args.num_classes, pretrained=True)
elif args.model_name == 'UNetVGG16':
self.net = UNetVGG16(args.num_classes,
pretrained=True,
dropout_2d=0.0,
is_deconv=True)
elif args.model_name == 'deeplab50_v2':
if args.ms:
raise NotImplemented
else:
self.net = deeplab50_v2(args.num_classes,
pretrained=args.pretrained)
elif args.model_name == 'deeplab_v2':
if args.ms:
self.net = ms_deeplab_v2(args.num_classes,
pretrained=args.pretrained,
scales=args.ms_scales)
else:
self.net = deeplab_v2(args.num_classes,
pretrained=args.pretrained)
elif args.model_name == 'deeplab_v3':
if args.ms:
self.net = ms_deeplab_v3(args.num_classes,
out_stride=args.out_stride,
pretrained=args.pretrained,
scales=args.ms_scales)
else:
self.net = deeplab_v3(args.num_classes,
out_stride=args.out_stride,
pretrained=args.pretrained)
elif args.model_name == 'deeplab_v3_plus':
if args.ms:
self.net = ms_deeplab_v3_plus(args.num_classes,
out_stride=args.out_stride,
pretrained=args.pretrained,
scales=args.ms_scales)
else:
self.net = deeplab_v3_plus(args.num_classes,
out_stride=args.out_stride,
pretrained=args.pretrained)
self.interp = nn.Upsample(size=args.size, mode='bilinear')
self.iterations = args.epochs
self.lr_current = args.lr
self.cuda = args.cuda
self.phase = args.phase
self.lr_policy = args.lr_policy
self.cyclic_m = args.cyclic_m
if self.lr_policy == 'cyclic':
print('using cyclic')
assert self.iterations % self.cyclic_m == 0
if args.loss == 'CELoss':
self.criterion = nn.CrossEntropyLoss(size_average=True)
elif args.loss == 'DiceLoss':
self.criterion = DiceLoss(num_classes=args.num_classes)
elif args.loss == 'MixLoss':
self.criterion = MixLoss(args.num_classes,
weights=args.loss_weights)
elif args.loss == 'LovaszLoss':
self.criterion = LovaszSoftmax(per_image=args.loss_per_img)
elif args.loss == 'FocalLoss':
self.criterion = FocalLoss(args.num_classes, alpha=None, gamma=2)
else:
raise RuntimeError('must define loss')
if 'deeplab' in args.model_name:
self.optimizer = optim.SGD(
[{
'params': get_1x_lr_params_NOscale(self.net),
'lr': args.lr
}, {
'params': get_10x_lr_params(self.net),
'lr': 10 * args.lr
}],
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
else:
self.optimizer = optim.SGD(filter(lambda p: p.requires_grad,
self.net.parameters()),
lr=args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
self.iters = 0
self.best_val = 0.0
self.count = 0
def init_model(self):
if self.args.resume_model:
saved_state_dict = torch.load(
self.args.resume_model,
map_location=lambda storage, loc: storage)
if self.args.ms:
new_params = self.net.Scale.state_dict().copy()
for i in saved_state_dict:
# Scale.layer5.conv2d_list.3.weight
i_parts = i.split('.')
# print i_parts
if not (not i_parts[0] == 'layer5') and (not i_parts[0]
== 'decoder'):
new_params[i] = saved_state_dict[i]
self.net.Scale.load_state_dict(new_params)
else:
new_params = self.net.state_dict().copy()
for i in saved_state_dict:
# Scale.layer5.conv2d_list.3.weight
i_parts = i.split('.')
# print i_parts
if (not i_parts[0] == 'layer5') and (not i_parts[0]
== 'decoder'):
# if not i_parts[0] == 'layer5':
new_params[i] = saved_state_dict[i]
self.net.load_state_dict(new_params)
print('Resuming training, image net loading {}...'.format(
self.args.resume_model))
# self.load_weights(self.net, self.args.resume_model)
if self.args.mGPUs:
self.net = nn.DataParallel(self.net)
if self.args.cuda:
self.net = self.net.cuda()
cudnn.benchmark = True
def _adjust_learning_rate(self, epoch):
"""Sets the learning rate to the initial LR decayed by 10 at every specified step
# Adapted from PyTorch Imagenet example:
# https://github.com/pytorch/examples/blob/master/imagenet/main.py
"""
if epoch < int(self.iterations * 0.5):
self.lr_current = max(self.lr_current * self.args.gamma, 1e-4)
elif epoch < int(self.iterations * 0.85):
self.lr_current = max(self.lr_current * self.args.gamma, 1e-5)
else:
self.lr_current = max(self.lr_current * self.args.gamma, 1e-6)
self.optimizer.param_groups[0]['lr'] = self.lr_current
self.optimizer.param_groups[1]['lr'] = self.lr_current * 10
def save_network(self, net, net_name, epoch, label=''):
save_fname = '%s_%s_%s.pth' % (epoch, net_name, label)
save_path = os.path.join(self.args.save_folder, self.args.exp_name,
save_fname)
torch.save(net.state_dict(), save_path)
def load_weights(self, net, base_file):
other, ext = os.path.splitext(base_file)
if ext == '.pkl' or '.pth':
print('Loading weights into state dict...')
net.load_state_dict(
torch.load(base_file,
map_location=lambda storage, loc: storage))
print('Finished!')
else:
print('Sorry only .pth and .pkl files supported.')
def load_trained_model(self):
path = os.path.join(self.args.save_folder, self.args.exp_name,
self.args.trained_model)
print('eval cls, image net loading {}...'.format(path))
if self.args.ms:
self.load_weights(self.net.Scale, path)
else:
self.load_weights(self.net, path)
def eval(self, dataloader):
assert self.phase == 'test', "Command arg phase should be 'test'. "
from tqdm import tqdm
self.net.eval()
output = []
for i, image in tqdm(enumerate(dataloader)):
if self.cuda:
image = Variable(image.cuda(), volatile=True)
else:
image = Variable(image, volatile=True)
# cls forward
out = self.net(image)
if isinstance(out, list):
out_max = out[-1]
if out_max.size(2) != image.size(2):
out = self.interp(out_max)
else:
if out.size(2) != image.size(2):
out = self.interp(out)
# out [bs * num_tta, c, h, w]
if self.args.use_tta:
num_tta = len(tta_config)
# out = F.softmax(out, dim=1)
out = detta_score(
out.view(num_tta, -1, self.args.num_classes, out.size(2),
out.size(3))) # [num_tta, bs, nclass, H, W]
out = out.mean(dim=0) # [bs, nclass, H, W]
out = F.softmax(out)
output.extend([
resize(pred[1].data.cpu().numpy(), (101, 101)) for pred in out
])
return np.array(output)
def tta(self, dataloaders):
results = np.zeros(shape=(len(dataloaders[0].dataset),
self.args.num_classes))
for dataloader in dataloaders:
output = self.eval(dataloader)
results += output
return np.argmax(results, 1)
def tta_output(self, dataloaders):
results = np.zeros(shape=(len(dataloaders[0].dataset),
self.args.num_classes))
for dataloader in dataloaders:
output = self.eval(dataloader)
results += output
return results
def test_val(self, dataloader):
assert self.phase == 'test', "Command arg phase should be 'test'. "
from tqdm import tqdm
self.net.eval()
predict = []
true = []
t1 = time.time()
for i, (image, mask) in tqdm(enumerate(dataloader)):
if self.cuda:
image = Variable(image.cuda(), volatile=True)
label_image = Variable(mask.cuda(), volatile=True)
else:
image = Variable(image, volatile=True)
label_image = Variable(mask, volatile=True)
# cls forward
out = self.net(image)
if isinstance(out, list):
out_max = out[-1]
if out_max.size(2) != label_image.size(2):
out = self.interp(out_max)
else:
if out.size(2) != image.size(2):
out = self.interp(out)
# out [bs * num_tta, c, h, w]
if self.args.use_tta:
num_tta = len(tta_config)
# out = F.softmax(out, dim=1)
out = detta_score(
out.view(num_tta, -1, self.args.num_classes, out.size(2),
out.size(3))) # [num_tta, bs, nclass, H, W]
out = out.mean(dim=0) # [bs, nclass, H, W]
out = F.softmax(out)
if self.args.aug == 'heng':
out = out[:, :, 11:11 + 202, 11:11 + 202]
predict.extend([
resize(pred[1].data.cpu().numpy(), (101, 101)) for pred in out
])
# predict.extend([pred[1, :101, :101].data.cpu().numpy() for pred in out])
# pred.extend(out.data.cpu().numpy())
true.extend(label_image.data.cpu().numpy())
# pred_all = np.argmax(np.array(pred), 1)
for t in np.arange(0.05, 0.51, 0.01):
pred_all = np.array(predict) > t
true_all = np.array(true).astype(np.int)
# new_iou = intersection_over_union(true_all, pred_all)
# new_iou_t = intersection_over_union_thresholds(true_all, pred_all)
mean_iou, iou_t = mIoU(true_all, pred_all)
print('threshold : {:.4f}'.format(t))
print('mean IoU : {:.4f}, IoU threshold : {:.4f}'.format(
mean_iou, iou_t))
return predict, true
def run_epoch(self, dataloader, writer, epoch, train=True, metrics=True):
if train:
self.net.train()
flag = 'train'
else:
self.net.eval()
flag = 'val'
t2 = time.time()
for image, mask in dataloader:
if train and self.lr_policy != 'step':
adjust_learning_rate(self.args.lr, self.optimizer, self.iters,
self.iterations * len(dataloader), 0.9,
self.cyclic_m, self.lr_policy)
self.iters += 1
if self.cuda:
image = Variable(image.cuda(), volatile=(not train))
label_image = Variable(mask.cuda(), volatile=(not train))
else:
image = Variable(image, volatile=(not train))
label_image = Variable(mask, volatile=(not train))
# cls forward
out = self.net(image)
if isinstance(out, list):
out_max = None
loss = 0.0
for i, out_scale in enumerate(out):
if out_scale.size(2) != label_image.size(2):
out_scale = self.interp(out_scale)
if i == (len(out) - 1):
out_max = out_scale
loss += self.criterion(out_scale, label_image)
label_image_np = label_image.data.cpu().numpy()
sig_out_np = out_max.data.cpu().numpy()
acc = accuracy(label_image_np, np.argmax(sig_out_np, 1))
self.pred[flag].extend(sig_out_np)
self.true[flag].extend(label_image_np)
self.losses[flag].append(loss.data[0])
self.acces[flag].append(acc)
else:
if out.size(-1) != label_image.size(-1):
out = self.interp(out)
loss = self.criterion(out, label_image)
label_image_np = label_image.data.cpu().numpy()
sig_out_np = out.data.cpu().numpy()
acc = accuracy(label_image_np, np.argmax(sig_out_np, 1))
self.pred[flag].extend(sig_out_np)
self.true[flag].extend(label_image_np)
self.losses[flag].append(loss.data[0])
self.acces[flag].append(acc)
if train:
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
if metrics:
n = len(self.losses[flag])
loss = sum(self.losses[flag]) / n
scalars = [
loss,
]
names = [
'loss',
]
write_scalars(writer, scalars, names, epoch, tag=flag + '_loss')
all_acc = sum(self.acces[flag]) / n
scalars = [
all_acc,
]
names = [
'all_acc',
]
write_scalars(writer, scalars, names, epoch, tag=flag + '_acc')
# all_score = sum(self.scores[flag]) / n
# scalars = [all_score, ]
# names = ['all_score', ]
# write_scalars(writer, scalars, names, epoch, tag=flag + '_score')
pred_all = np.argmax(np.array(self.pred[flag]), 1)
true_all = np.array(self.true[flag]).astype(np.int)
mean_iou, iou_t = mIoU(true_all, pred_all)
# new_iou = intersection_over_union(true_all, pred_all)
# new_iou_t = intersection_over_union_thresholds(true_all, pred_all)
scalars = [
mean_iou,
iou_t,
]
names = [
'mIoU',
'mIoU_threshold',
]
write_scalars(writer, scalars, names, epoch, tag=flag + '_IoU')
scalars = [
self.optimizer.param_groups[0]['lr'],
]
names = [
'learning_rate',
]
write_scalars(writer, scalars, names, epoch, tag=flag + '_lr')
print(
'{} loss: {:.4f} | acc: {:.4f} | mIoU: {:.4f} | mIoU_threshold: {:.4f} | n_iter: {} | learning_rate: {} | time: {:.2f}'
.format(flag, loss, all_acc, mean_iou, iou_t, epoch,
self.optimizer.param_groups[0]['lr'],
time.time() - t2))
self.losses[flag] = []
self.pred[flag] = []
self.true[flag] = []
self.acces[flag] = []
self.scores[flag] = []
if (not train) and (iou_t >= self.best_val):
if self.args.ms:
if self.args.mGPUs:
self.save_network(self.net.module.Scale,
self.args.model_name,
epoch=epoch,
label='best')
else:
self.save_network(self.net.Scale,
self.args.model_name,
epoch=epoch,
label='best')
else:
if self.args.mGPUs:
self.save_network(self.net.module,
self.args.model_name,
epoch=epoch,
label='best')
else:
self.save_network(self.net,
self.args.model_name,
epoch=epoch,
label='best')
print(
'val improve from {:.4f} to {:.4f} saving in best val_iteration {}'
.format(self.best_val, iou_t, epoch))
self.best_val = iou_t
self.count = 0
if (not train) and (self.best_val - iou_t > 0.003) and (
self.count < 10) and (self.lr_policy == 'step'):
self.count += 1
if (not train) and (self.count >= 10) and (self.lr_policy
== 'step'):
self._adjust_learning_rate(epoch)
self.count = 0
def train_val(self, dataloader_train, dataloader_val, writer):
val_epoch = 0
for epoch in range(self.iterations):
if (self.lr_policy == 'cyclic') and (
epoch % int(self.iterations / self.cyclic_m) == 0):
print('-------start cycle {}------------'.format(
epoch // int(self.iterations / self.cyclic_m)))
self.best_val = 0.0
self.run_epoch(dataloader_train,
writer,
epoch,
train=True,
metrics=True)
self.run_epoch(dataloader_val,
writer,
val_epoch,
train=False,
metrics=True)
val_epoch += 1
if (epoch + 1) % self.args.save_freq == 0:
if self.args.ms:
if self.args.mGPUs:
self.save_network(
self.net.module.Scale,
self.args.model_name,
epoch=val_epoch,
)
else:
self.save_network(
self.net.Scale,
self.args.model_name,
epoch=val_epoch,
)
else:
if self.args.mGPUs:
self.save_network(
self.net.module,
self.args.model_name,
epoch=val_epoch,
)
else:
self.save_network(
self.net,
self.args.model_name,
epoch=val_epoch,
)
print('saving in val_iteration {}'.format(val_epoch))
def train(train_sources, eval_source):
path = sys.argv[1]
dr = DataReader(path, train_sources)
dr.read()
print(len(dr.train.x))
batch_size = 8
device = torch.device('cpu')
if torch.cuda.is_available():
device = torch.device('cuda')
dataset_s_train = MultiDomainDataset(dr.train.x, dr.train.y, dr.train.vendor, device, DomainAugmentation())
dataset_s_dev = MultiDomainDataset(dr.dev.x, dr.dev.y, dr.dev.vendor, device)
dataset_s_test = MultiDomainDataset(dr.test.x, dr.test.y, dr.test.vendor, device)
loader_s_train = DataLoader(dataset_s_train, batch_size, shuffle=True)
dr_eval = DataReader(path, [eval_source])
dr_eval.read()
dataset_eval_dev = MultiDomainDataset(dr_eval.dev.x, dr_eval.dev.y, dr_eval.dev.vendor, device)
dataset_eval_test = MultiDomainDataset(dr_eval.test.x, dr_eval.test.y, dr_eval.test.vendor, device)
dataset_da_train = MultiDomainDataset(dr.train.x+dr_eval.train.x, dr.train.y+dr_eval.train.y, dr.train.vendor+dr_eval.train.vendor, device, DomainAugmentation())
loader_da_train = DataLoader(dataset_da_train, batch_size, shuffle=True)
segmentator = UNet()
discriminator = Discriminator(n_domains=len(train_sources))
discriminator.to(device)
segmentator.to(device)
sigmoid = nn.Sigmoid()
selector = Selector()
s_criterion = nn.BCELoss()
d_criterion = nn.CrossEntropyLoss()
s_optimizer = optim.AdamW(segmentator.parameters(), lr=0.0001, weight_decay=0.01)
d_optimizer = optim.AdamW(discriminator.parameters(), lr=0.001, weight_decay=0.01)
a_optimizer = optim.AdamW(segmentator.encoder.parameters(), lr=0.001, weight_decay=0.01)
lmbd = 1/150
s_train_losses = []
s_dev_losses = []
d_train_losses = []
eval_domain_losses = []
train_dices = []
dev_dices = []
eval_dices = []
epochs = 3
da_loader_iter = iter(loader_da_train)
for epoch in tqdm(range(epochs)):
s_train_loss = 0.0
d_train_loss = 0.0
for index, sample in enumerate(loader_s_train):
img = sample['image']
target_mask = sample['target']
da_sample = next(da_loader_iter, None)
if epoch == 100:
s_optimizer.defaults['lr'] = 0.001
d_optimizer.defaults['lr'] = 0.0001
if da_sample is None:
da_loader_iter = iter(loader_da_train)
da_sample = next(da_loader_iter, None)
if epoch < 50 or epoch >= 100:
# Training step of segmentator
predicted_activations, inner_repr = segmentator(img)
predicted_mask = sigmoid(predicted_activations)
s_loss = s_criterion(predicted_mask, target_mask)
s_optimizer.zero_grad()
s_loss.backward()
s_optimizer.step()
s_train_loss += s_loss.cpu().detach().numpy()
if epoch >= 50:
# Training step of discriminator
predicted_activations, inner_repr = segmentator(da_sample['image'])
predicted_activations = predicted_activations.clone().detach()
inner_repr = inner_repr.clone().detach()
predicted_vendor = discriminator(predicted_activations, inner_repr)
d_loss = d_criterion(predicted_vendor, da_sample['vendor'])
d_optimizer.zero_grad()
d_loss.backward()
d_optimizer.step()
d_train_loss += d_loss.cpu().detach().numpy()
if epoch >= 100:
# adversarial training step
predicted_mask, inner_repr = segmentator(da_sample['image'])
predicted_vendor = discriminator(predicted_mask, inner_repr)
a_loss = -1 * lmbd * d_criterion(predicted_vendor, da_sample['vendor'])
a_optimizer.zero_grad()
a_loss.backward()
a_optimizer.step()
lmbd += 1/150
inference_model = nn.Sequential(segmentator, selector, sigmoid)
inference_model.to(device)
inference_model.eval()
d_train_losses.append(d_train_loss / len(loader_s_train))
s_train_losses.append(s_train_loss / len(loader_s_train))
s_dev_losses.append(calculate_loss(dataset_s_dev, inference_model, s_criterion, batch_size))
eval_domain_losses.append(calculate_loss(dataset_eval_dev, inference_model, s_criterion, batch_size))
train_dices.append(calculate_dice(inference_model, dataset_s_train))
dev_dices.append(calculate_dice(inference_model, dataset_s_dev))
eval_dices.append(calculate_dice(inference_model, dataset_eval_dev))
segmentator.train()
date_time = datetime.now().strftime("%m%d%Y_%H%M%S")
model_path = os.path.join(pathlib.Path(__file__).parent.absolute(), "model", "weights", "segmentator"+str(date_time)+".pth")
torch.save(segmentator.state_dict(), model_path)
util.plot_data([(s_train_losses, 'train_losses'), (s_dev_losses, 'dev_losses'), (d_train_losses, 'discriminator_losses'),
(eval_domain_losses, 'eval_domain_losses')],
'losses.png')
util.plot_dice([(train_dices, 'train_dice'), (dev_dices, 'dev_dice'), (eval_dices, 'eval_dice')],
'dices.png')
inference_model = nn.Sequential(segmentator, selector, sigmoid)
inference_model.to(device)
inference_model.eval()
print('Dice on annotated: ', calculate_dice(inference_model, dataset_s_test))
print('Dice on unannotated: ', calculate_dice(inference_model, dataset_eval_test))
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
n_classes = utils.params.n_classes
n_channels = utils.params.n_channels
print(
"Number of input channels = {}. Number of classes = {}.".format(
n_channels, n_classes
)
)
if n_classes == 3:
class_weights = np.array([0.3, 0.5, 1]).astype(np.float)
elif n_classes == 2:
class_weights = np.array([0.047619, 1]).astype(np.float)
print("Current class weights = {}".format(class_weights))
assert (
len(class_weights) == n_classes
), "Should be a 1D Tensor assigning weight to each of the classes. Lenght of the weights-vector should be equal to the number of classes"
net = UNet(n_channels=n_channels, n_classes=n_classes)
net.to(device=device)
try:
print("Training starting...")
train_net(net, n_channels, n_classes, class_weights)
print("Training done")
except KeyboardInterrupt:
torch.save(net.state_dict(), model_path + "INTERRUPTED.pth")
try:
sys.exit(0)
except SystemExit:
os._exit(0)
