def make_val_step(self, data):
report = {}
top, left, height, width = data['top'][0], data['left'][0], data['height'][0], data['width'][0]
image = variable(data['image'], volatile=True)
mask = variable(data['mask'], volatile=True)
prediction = self.model(image)
loss = self.criterion(prediction, mask)
report['loss'] = loss.data
for name, func in self.metrics:
report[name] = func(
prediction[:, :, top:height + top, left:width + left].contiguous(),
mask[:, :, top:height + top, left:width + left].contiguous()
).data
return report
def make_train_step(self, data):
report = {}
images = variable(data['image'])
masks = variable(data['mask'])
self.optimizer.zero_grad()
predictions = self.model(images)
loss = self.criterion(predictions, masks)
report['loss'] = loss.data
for name, func in self.metrics:
report[name] = func(predictions, masks).data
loss.backward()
torch.nn.utils.clip_grad_norm(self.model.parameters(), 1.0)
self.optimizer.step()
return report
def predict(model, ids, path_images, transforms, period, flips, num_workers):
dataset = TestDataset(ids, path_images, transforms, period)
loader = DataLoader(dataset=dataset,
shuffle=False,
drop_last=False,
num_workers=num_workers,
pin_memory=torch.cuda.is_available())
test_predictions = []
for data in tqdm(loader, desc='Predict', total=len(ids)):
images, tops, lefts, heights, widths = data['image'], data[
'top'], data['left'], data['height'], data['width']
inputs = variable(images, volatile=True)
predictions = batch_predict(model, inputs, flips=flips)
for i, prediction in enumerate(predictions):
prediction = np.moveaxis(prediction, 0, -1)
test_predictions.append(prediction[tops[i]:heights[i] + tops[i],
lefts[i]:widths[i] + lefts[i]])
return test_predictions
def flip_tensor_ud(batch):
invert_indices = torch.arange(batch.data.size()[-2] - 1, -1, -1).long()
return torch.index_select(batch, 2, variable(invert_indices))
def flip_tensor_lr(batch):
invert_indices = torch.arange(batch.data.size()[-1] - 1, -1, -1).long()
return batch.index_select(3, variable(invert_indices))