def predict(self):
"""The testing process.
"""
self.net.eval()
dataloader = self.test_dataloader
pbar = tqdm(dataloader, desc='test', ascii=True)
epoch_log = EpochLog()
for i, batch in enumerate(pbar):
with torch.no_grad():
test_dict = self._test_step(batch)
loss = test_dict['loss']
losses = test_dict.get('losses')
metrics = test_dict.get('metrics')
if (i + 1) == len(dataloader) and not dataloader.drop_last:
batch_size = len(dataloader.dataset) % dataloader.batch_size
else:
batch_size = dataloader.batch_size
epoch_log.update(batch_size, loss, losses, metrics)
pbar.set_postfix(**epoch_log.on_step_end_log)
test_log = epoch_log.on_epoch_end_log
LOGGER.info(f'Test log: {test_log}.')
return test_log
def _run_epoch(self, mode):
"""Run an epoch for training.
Args:
mode (str): The mode of running an epoch ('train' or 'valid').
Returns:
log (dict): The log information.
batch (dict or sequence): The last batch of the data.
outputs (torch.Tensor or sequence of torch.Tensor): The corresponding model outputs.
"""
if mode == 'train':
self.net.train()
dataloader = self.train_dataloader
dataloader_iterator = iter(dataloader)
outter_pbar = tqdm(total=(
(len(dataloader) + dataloader.grad_accumulation_steps() - 1) //
dataloader.grad_accumulation_steps()),
desc='train',
ascii=True)
inner_pbar = tqdm(total=math.ceil(
dataloader.grad_accumulation_steps(0)),
desc='grad_accumulation',
leave=False,
ascii=True)
epoch_log = EpochLog()
for i in range(len(dataloader)):
batch = next(dataloader_iterator)
train_dict = self._train_step(batch)
loss = train_dict.get('loss')
if loss is None:
raise KeyError(
f"The train_dict must have the key named 'loss'. "
'Please check the returned keys as defined in MyTrainer._train_step().'
)
losses = train_dict.get('losses')
metrics = train_dict.get('metrics')
outputs = train_dict.get('outputs')
if self.use_amp:
with amp.scale_loss(loss, self.optimizer) as scaled_loss:
(scaled_loss /
dataloader.grad_accumulation_steps(i)).backward()
else:
(loss / dataloader.grad_accumulation_steps(i)).backward()
if (i + 1) % dataloader.grad_accumulation_steps() == 0 or (
i + 1) == len(dataloader):
self.optimizer.step()
self.optimizer.zero_grad()
if isinstance(self.lr_scheduler, CyclicLR):
self.lr_scheduler.step()
elif isinstance(self.lr_scheduler,
CosineAnnealingWarmRestarts):
self.lr_scheduler.step((self.epoch - 1) +
i / len(dataloader))
if (i + 1) == len(dataloader) and not dataloader.drop_last:
batch_size = len(
dataloader.dataset) % dataloader.batch_size
else:
batch_size = dataloader.batch_size
epoch_log.update(batch_size, loss, losses, metrics)
inner_pbar.update()
if (i + 1) % dataloader.grad_accumulation_steps() == 0 or (
i + 1) == len(dataloader):
outter_pbar.update()
outter_pbar.set_postfix(**epoch_log.on_step_end_log)
inner_pbar.close()
inner_pbar = tqdm(total=math.ceil(
dataloader.grad_accumulation_steps(i + 1)),
desc='grad_accumulation',
leave=False,
ascii=True)
outter_pbar.close()
inner_pbar.close()
else:
self.net.eval()
dataloader = self.valid_dataloader
pbar = tqdm(dataloader, desc='valid', ascii=True)
epoch_log = EpochLog()
for i, batch in enumerate(pbar):
with torch.no_grad():
valid_dict = self._valid_step(batch)
loss = valid_dict.get('loss')
if loss is None:
raise KeyError(
f"The valid_dict must have the key named 'loss'. "
'Please check the returned keys as defined in MyTrainer._valid_step().'
)
losses = valid_dict.get('losses')
metrics = valid_dict.get('metrics')
outputs = valid_dict.get('outputs')
if (i + 1) == len(dataloader) and not dataloader.drop_last:
batch_size = len(
dataloader.dataset) % dataloader.batch_size
else:
batch_size = dataloader.batch_size
epoch_log.update(batch_size, loss, losses, metrics)
pbar.set_postfix(**epoch_log.on_step_end_log)
return epoch_log.on_epoch_end_log, batch, outputs
def _run_epoch(self, mode):
"""Run an epoch for training.
Args:
mode (str): The mode of running an epoch ('train' or 'valid').
Returns:
log (dict): The log information.
batch (dict or sequence): The last batch of the data.
outputs (torch.Tensor or sequence of torch.Tensor): The corresponding model outputs.
"""
for fn in self.metric_fns:
fn.reset_count()
if mode == 'train':
self.net.train()
dataloader = self.train_dataloader
dataloader_iterator = iter(dataloader)
outter_pbar = tqdm(total=len(dataloader), desc=mode, ascii=True)
epoch_log = EpochLog()
for i in range(len(dataloader)):
batch = next(dataloader_iterator)
train_dict = self._train_step(batch)
loss = train_dict['loss']
losses = train_dict.get('losses')
metrics = train_dict.get('metrics')
outputs = train_dict.get('outputs')
lr = self.optimizer.param_groups[0]['lr']
loss.backward()
if self.grad_norm:
torch.nn.utils.clip_grad_norm_(self.net.parameters(), 1.0)
self.optimizer.step()
self.optimizer.zero_grad()
if (i + 1) == len(dataloader) and not dataloader.drop_last:
batch_size = len(
dataloader.dataset) % dataloader.batch_size
else:
batch_size = dataloader.batch_size
epoch_log.update(batch_size, loss, losses, metrics, lr)
outter_pbar.update()
outter_pbar.set_postfix(**epoch_log.on_step_end_log)
outter_pbar.close()
else:
self.net.eval()
dataloader = self.valid_dataloader
pbar = tqdm(dataloader, desc=mode, ascii=True)
epoch_log = EpochLog()
for i, batch in enumerate(pbar):
with torch.no_grad():
valid_dict = self._valid_step(batch)
loss = valid_dict['loss']
losses = valid_dict.get('losses')
metrics = valid_dict.get('metrics')
outputs = valid_dict.get('outputs')
if (i + 1) == len(dataloader) and not dataloader.drop_last:
batch_size = len(
dataloader.dataset) % dataloader.batch_size
else:
batch_size = dataloader.batch_size
epoch_log.update(batch_size, loss, losses, metrics)
pbar.set_postfix(**epoch_log.on_step_end_log)
return epoch_log.on_epoch_end_log, batch, outputs
def _run_epoch(self, mode):
"""Run an epoch for training.
Args:
mode (str): The mode of running an epoch ('train' or 'valid').
Returns:
log (dict): The log information.
batch (dict or sequence): The last batch of the data.
outputs (torch.Tensor or sequence of torch.Tensor): The corresponding model outputs.
"""
if mode == 'train':
self.net.train()
dataloader = self.train_dataloader
dataloader_iterator = iter(dataloader)
outter_pbar = tqdm(total=len(dataloader), desc=mode, ascii=True)
epoch_log = EpochLog()
for i in range(len(dataloader)):
batch = next(dataloader_iterator)
train_dict = self._train_step(batch)
loss = train_dict['loss']
losses = train_dict.get('losses')
metrics = train_dict.get('metrics')
outputs = train_dict.get('outputs')
lr = self.optimizer.param_groups[0]['lr']
loss.backward()
self.optimizer.step()
self.optimizer.zero_grad()
# if isinstance(self.lr_scheduler, (CyclicLR, OneCycleLR)):
# self.lr_scheduler.step()
# elif isinstance(self.lr_scheduler, CosineAnnealingWarmRestarts):
# self.lr_scheduler.step((self.epoch - 1) + i / len(dataloader))
if (i + 1) == len(dataloader) and not dataloader.drop_last:
batch_size = len(
dataloader.dataset) % dataloader.batch_size
else:
batch_size = dataloader.batch_size
epoch_log.update(batch_size, loss, losses, metrics, lr)
outter_pbar.update()
outter_pbar.set_postfix(**epoch_log.on_step_end_log)
outter_pbar.close()
else:
self.net.eval()
dataloader = self.valid_dataloader
pbar = tqdm(dataloader, desc=mode, ascii=True)
epoch_log = EpochLog()
for i, batch in enumerate(pbar):
with torch.no_grad():
valid_dict = self._valid_step(batch)
loss = valid_dict['loss']
losses = valid_dict.get('losses')
metrics = valid_dict.get('metrics')
outputs = valid_dict.get('outputs')
if (i + 1) == len(dataloader) and not dataloader.drop_last:
batch_size = len(
dataloader.dataset) % dataloader.batch_size
else:
batch_size = dataloader.batch_size
epoch_log.update(batch_size, loss, losses, metrics)
pbar.set_postfix(**epoch_log.on_step_end_log)
return epoch_log.on_epoch_end_log, batch, outputs