def training_step(self, train_loader):
self.train()
train_loss = []
train_iou = []
train_mAP = []
for i, (index, im, mask, z) in enumerate(train_loader):
self.optimizer.zero_grad()
im = im.cuda()
mask = mask.cuda()
z = z.cuda()
logit = self.forward(im, z)
pred = torch.sigmoid(logit)
loss = self.criterion(logit, mask)
iou = dice_accuracy(pred, mask, is_average=False)
mAP = do_mAP(pred.data.cpu().numpy(), mask.cpu().numpy(), is_average=False)
train_loss.append(loss.item())
train_iou.extend(iou)
train_mAP.extend(mAP)
loss.backward()
self.optimizer.step()
for metric in ['loss', 'iou', 'mAP']:
self.train_log[metric].append(np.mean(eval('train_{}'.format(metric))))
return train_loss, train_iou, train_mAP
def perform_validation(self, val_loader):
# self.set_mode('valid')
self.model.eval()
val_loss = []
val_iou = []
val_mAP = []
for images, targets in val_loader:
images = images.cuda()
targets = targets.long().cuda()
with torch.no_grad():
logit = self.model(images)
pred = torch.sigmoid(logit)
# pred = torch.softmax(logit, dim=1)
pred, pred_ind = pred.data.max(1)
loss = self.criterion(logit, targets)
iou = dice_accuracy(pred, targets,
is_average=False).cpu().numpy()
mAP = do_mAP(pred.cpu().numpy(),
targets.cpu().numpy(),
is_average=False)
val_loss.append(loss.item())
val_iou.extend(iou)
val_mAP.extend(mAP)
# Append epoch data to metrics dict
for metric in ['loss', 'iou', 'mAP']:
self.val_log[metric].append(np.mean(eval('val_{}'.format(metric))))
return val_loss, val_iou, val_mAP
def perform_validation(self, val_loader):
self.set_mode('valid')
val_loss = []
val_iou = []
val_mAP = []
for index, im, mask, z in val_loader:
im = im.cuda()
mask = mask.cuda()
z = z.cuda()
with torch.no_grad():
logit = self.forward(im, z)
pred = torch.sigmoid(logit)
loss = self.criterion(logit, mask)
iou = dice_accuracy(pred, mask, is_average=False)
mAP = do_mAP(pred.cpu().numpy(),
mask.cpu().numpy(),
is_average=False)
val_loss.append(loss.item())
val_iou.extend(iou)
val_mAP.extend(mAP)
# Append epoch data to metrics dict
for metric in ['loss', 'iou', 'mAP']:
self.val_log[metric].append(np.mean(eval('val_{}'.format(metric))))
return val_loss, val_iou, val_mAP
def training_step(self, train_loader):
# self.set_mode('train')
self.model.train()
train_loss = []
train_iou = []
train_mAP = []
for i, (images, targets) in enumerate(train_loader):
self.optimizer.zero_grad()
images = images.cuda()
targets = targets.long().cuda()
# logit = self.forward(im, z) self.model
logit = self.model(images)
pred = torch.sigmoid(logit)
# pred = torch.softmax(logit, dim=1)
pred, pred_ind = pred.data.max(1)
loss = self.criterion(logit, targets)
iou = dice_accuracy(pred, targets, is_average=False).cpu().numpy()
mAP = do_mAP(pred.data.cpu().numpy(),
targets.cpu().numpy(),
is_average=False)
train_loss.append(loss.item())
train_iou.extend(iou)
train_mAP.extend(mAP)
loss.backward()
self.optimizer.step()
if self.debug and not self.epoch % 5 and not i % 30:
# if self.debug and self.epoch == 1 and i == 2:
show_image_mask_pred(images.cpu().data.numpy(),
targets.cpu().data.numpy(),
logit.cpu().data.numpy())
# Append epoch data to metrics dict
for metric in ['loss', 'iou', 'mAP']:
self.train_log[metric].append(
np.mean(eval('train_{}'.format(metric))))
return train_loss, train_iou, train_mAP