def _train_epoch(self, epoch_id, epochs):
losses = tools.AverageMeter()
accuracies = tools.AverageMeter()
# Total training files count / batch_size
batch_size = self.train_loader.batch_size
it_count = len(self.train_loader)
with tqdm(total=it_count,
desc="Epochs {}/{}".format(epoch_id + 1, epochs),
bar_format=
'{l_bar}{bar}| {n_fmt}/{total_fmt} [{remaining}{postfix}]'
) as pbar:
for ind, loader_dict in enumerate(self.train_loader):
#train
loss, acc = self._batch_train_validation(loader_dict,
volatile=False)
# backward + optimize
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
losses.update(loss.data[0], batch_size)
accuracies.update(acc, batch_size)
# Update pbar
pbar.set_postfix(
OrderedDict(loss='{0:1.5f}'.format(loss.data[0]),
acc='{0:1.5f}'.format(acc)))
pbar.update(1)
return losses.avg, accuracies.avg
def _validate_epoch(self, valid_loader, threshold):
losses = tools.AverageMeter()
dice_coeffs = tools.AverageMeter()
it_count = len(valid_loader)
batch_size = valid_loader.batch_size
images = None # To save the last images batch
targets = None # To save the last target batch
preds = None # To save the last prediction batch
with tqdm(total=it_count, desc="Validating", leave=False) as pbar:
for ind, (images, targets) in enumerate(valid_loader):
if self.use_cuda:
images = images.cuda()
targets = targets.cuda()
# Volatile because we are in pure inference mode
# http://pytorch.org/docs/master/notes/autograd.html#volatile
images = Variable(images, volatile=True)
targets = Variable(targets, volatile=True)
# forward
logits = self.net(images)
probs = F.sigmoid(logits)
preds = (probs > threshold).float()
loss = self._criterion(logits, targets)
acc = losses_utils.dice_coeff(preds, targets)
losses.update(loss.data[0], batch_size)
dice_coeffs.update(acc.data[0], batch_size)
pbar.update(1)
return losses.avg, dice_coeffs.avg, images, targets, preds
def _validate_epoch(self, valid_loader, threshold):
"""负责在一个epoch验证"""
losses = tools.AverageMeter()
dice_coeffs = tools.AverageMeter()
it_count = len(valid_loader)
batch_size = valid_loader.batch_size
images = None # To save the last images batch
targets = None # To save the last target batch
preds = None # To save the last prediction batch
with tqdm(total=it_count, desc="Validating", leave=False) as pbar:
for ind, (images, targets) in enumerate(valid_loader):
if self.use_cuda:
images = images.to(self.device)
targets = targets.to(self.device)
with torch.no_grad():
# forward
logits = self.net(images)
probs = torch.sigmoid(logits)
preds = (probs > threshold).float()
loss = self.criterion(logits, targets)
acc = losses_utils.dice_coeff(preds, targets)
losses.update(loss.data.item(), batch_size)
dice_coeffs.update(acc.data.item(), batch_size)
pbar.update(1)
batch_results = [ x.cpu() if self.use_cuda else x for x in [images[:2], targets[:2], preds[:2]]]
return losses.avg, dice_coeffs.avg, batch_results
def _train_epoch(self, train_loader, optimizer, threshold):
losses = tools.AverageMeter()
dice_coeffs = tools.AverageMeter()
# Total training files count / batch_size
batch_size = train_loader.batch_size
it_count = len(train_loader)
with tqdm(total=it_count,
desc="Epochs {}/{}".format(self.epoch_counter + 1, self.max_epochs),
bar_format='{l_bar}{bar}| {n_fmt}/{total_fmt} [{remaining}{postfix}]'
) as pbar:
for ind, (inputs, target, mask) in enumerate(train_loader):
if self.use_cuda:
inputs = inputs.cuda()
target = target.cuda()
mask = mask.cuda()
inputs, target, mask = Variable(inputs), Variable(target), Variable(mask)
# forward
logits = self.net.forward(inputs)
probs = F.sigmoid(logits)
pred = (probs > threshold).float()
# backward + optimize
loss = self._criterion(logits, target, mask)
optimizer.zero_grad()
loss.backward()
optimizer.step()
# print statistics
acc = losses_utils.dice_coeff(pred*mask, target*mask)
# No longer works for pytorch>=0.5
# losses.update(loss.data[0], batch_size)
losses.update(loss.data, batch_size)
# No longer works for pytorch>=0.5
# dice_coeffs.update(acc.data[0], batch_size)
dice_coeffs.update(acc.data, batch_size)
# Update pbar
# No longer works for pytorch>=0.5
# pbar.set_postfix(OrderedDict(loss='{0:1.5f}'.format(loss.data[0]),
# dice_coeff='{0:1.5f}'.format(acc.data[0])))
pbar.set_postfix(OrderedDict(loss='{0:1.5f}'.format(loss.data),
dice_coeff='{0:1.5f}'.format(acc.data)))
pbar.update(1)
return losses.avg, dice_coeffs.avg
def _validate_epoch(self):
losses = tools.AverageMeter()
accuracies = tools.AverageMeter()
it_count = len(self.valid_loader)
batch_size = self.train_loader.batch_size
with tqdm(total=it_count, desc="Validating", leave=False) as pbar:
for ind, loader_dict in enumerate(self.valid_loader):
#train
loss, acc = self._batch_train_validation(loader_dict,
volatile=True)
losses.update(loss.data[0], batch_size)
accuracies.update(acc, batch_size)
pbar.update(1)
return losses.avg, accuracies.avg
def _train_epoch(self, train_loader, threshold):
"""负责训练一个epoch"""
losses = tools.AverageMeter() # 每个epoch都一个meter
dice_coeffs = tools.AverageMeter()
# Total training files count / batch_size
batch_size = train_loader.batch_size
it_count = len(train_loader)
with tqdm(total=it_count,
desc=f"Epochs {self.epoch_counter + 1}/{self.max_epochs}",
bar_format='{l_bar}{bar}| {n_fmt}/{total_fmt} [{remaining}{postfix}]'
) as pbar:
for ind, (inputs, target) in enumerate(train_loader):
inputs, target = inputs.contiguous(), target.contiguous() # image=【8, 3, 572, 572】，mask=[8, 388, 388]了
if self.use_cuda:
inputs = inputs.to(self.device)
target = target.to(self.device)
# forward
logits = self.net.forward(inputs) # [8, 388, 388]
probs = torch.sigmoid(logits)
pred = (probs > threshold).float()
# backward + optimize
loss = self.criterion(logits, target)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
if self.scheduler:
self.scheduler.step()
# print statistics
acc = losses_utils.dice_coeff(pred, target)
losses.update(loss.data.item(), batch_size) # 将一个batch的平均loss输入
dice_coeffs.update(acc.data.item(), batch_size) # 一个batch的平均准确率
# Update pbar
pbar.set_postfix(OrderedDict(loss='{0:1.5f}'.format(loss.data.item()),
dice_coeff='{0:1.5f}'.format(acc.data.item()))) # 在最后显示该batch的损失和准确率
pbar.update(1) # 进度条涨1
return losses.avg, dice_coeffs.avg # 返回的是当前epoch所有样本(这里是像素点)的平均损失和精确度
