def train(self):
assert "training" in self.data_loaders
# set cumulative losses to zero
self.losses = self.training_process.losses['training'].to(
self.training_process.config.device)
self.cumulative_losses = {
loss_name: 0
for loss_name in self.losses.names
}
if self.training_process.interpolator is not None:
self.cumulative_losses.update(
{'itp_' + loss_name: 0
for loss_name in self.losses.names})
# set mode of network to train (parameters will be affected / updated)
self.training_process.model.train()
# set progress bar
progressBar = ProgressBar(self.num_loadings['training'],
displaySumCount=True)
# iterate over data loader
for i, batch in enumerate(self.data_loaders['training']):
# load data to device
(targets_device, inputs_device, mask_lr_device, mask_hr_device,
offset_lr, offset_hr) = self._prepare_data(batch)
# zero out optimizer weights
self.training_process.optimizer.zero_grad()
# apply model
predictions_device, interpolate = self._apply_model(
inputs_device, offset_lr, offset_hr)
# compute loss to current targets
loss = self._update_losses(predictions_device, targets_device,
interpolate, mask_hr_device, offset_hr)
# backpropagate loss
loss.backward()
# advance weights towards minimum
self.training_process.optimizer.step()
# update progress bar
progressBar.proceed(i + 1)
# update epoch count
self._update_summary('training')
# save epoch state
if self.training_process.config.saving_period is not None:
self._save_state()
# update scheduler
self._update_scheduler()
def validate(self):
assert "validation" in self.data_loaders
# set cumulative losses to 0
if "validation" in self.training_process.losses:
self.losses = self.training_process.losses['validation'].to(
self.training_process.config.device)
else:
self.losses = self.training_process.losses['training'].to(
self.training_process.config.device)
self.cumulative_losses = {
loss_name: 0
for loss_name in self.losses.names
}
if self.training_process.interpolator is not None:
self.cumulative_losses.update(
{'itp_' + loss_name: 0
for loss_name in self.losses.names})
# set mode of network to eval (parameters will not be affected / updated)
self.training_process.model.eval()
# set progress bar
progressBar = ProgressBar(self.num_loadings['validation'],
displaySumCount=True)
# iterate over data loader
with torch.no_grad():
for i, batch in enumerate(self.data_loaders['validation']):
# load data to device
(targets_device, inputs_device, mask_lr_device, mask_hr_device,
offset_lr, offset_hr) = self._prepare_data(batch)
# apply model
predictions_device, interpolate = self._apply_model(
inputs_device, offset_lr, offset_hr)
# compute loss to current targets
self._update_losses(predictions_device, targets_device,
interpolate, mask_hr_device, offset_hr)
# update progress bar
progressBar.proceed(i + 1)
# update epoch count
self._update_summary('validation')
