def validate(self, epoch=-1):
test_loss = 0
correct = 0
total = 0
progress_bar = ext.ProgressBar(len(self.val_loader))
self.model.eval()
with torch.no_grad():
for inputs, targets in self.val_loader:
inputs, targets = inputs.to(self.device), targets.to(
self.device)
outputs = self.model(inputs)
test_loss += self.criterion(outputs,
targets).item() * targets.size(0)
prediction = outputs.max(1, keepdim=True)[1]
correct += prediction.eq(
targets.view_as(prediction)).sum().item()
total += targets.size(0)
progress_bar.step('Loss: {:.5g} | Accuracy: {:.2f}%'.format(
test_loss / total, 100. * correct / total))
test_loss /= total
accuracy = correct * 100. / total
self.vis.add_value('test loss', test_loss)
self.vis.add_value('test accuracy', accuracy)
self.logger(
'Test on epoch {}: average loss={:.5g}, accuracy={:.2f}% ({}/{}), time: {}'
.format(epoch, test_loss, accuracy, correct, total,
progress_bar.time_used()))
if not self.cfg.test and accuracy > self.best_acc:
self.best_acc = accuracy
self.saver.save_model('best.pth')
self.logger('==> best accuracy: {:.2f}%'.format(self.best_acc))
return accuracy, test_loss
def train_epoch(self, epoch):
self.logger(
'\nEpoch: {}, lr: {:.2g}, weight decay: {:.2g} on model {}'.format(
epoch, self.optimizer.param_groups[0]['lr'],
self.optimizer.param_groups[0]['weight_decay'],
self.model_name))
self.model.train()
train_loss = 0
correct = 0
total = 0
progress_bar = ext.ProgressBar(len(self.train_loader))
for i, (inputs, targets) in enumerate(self.train_loader, 1):
inputs = inputs.to(self.device)
targets = inputs if self.cfg.arch == 'AE' else targets.to(
self.device)
#print(inputs.size())
# compute output
outputs = self.model(inputs)
losses = self.criterion(outputs, targets)
# compute gradient and do SGD step
self.optimizer.zero_grad()
losses.backward()
self.optimizer.step()
# measure accuracy and record loss
train_loss += losses.item() * targets.size(0)
if self.cfg.arch == 'AE':
correct = -train_loss
else:
pred = outputs.max(1, keepdim=True)[1]
correct += pred.eq(targets.view_as(pred)).sum().item()
total += targets.size(0)
if i % 10 == 0 or i == len(self.train_loader):
progress_bar.step(
'Loss: {:.5g} | Accuracy: {:.2f}%'.format(
train_loss / total, 100. * correct / total), 10)
train_loss /= total
accuracy = 100. * correct / total
self.vis.add_value('train loss', train_loss)
self.vis.add_value('train accuracy', accuracy)
self.logger(
'Train on epoch {}: average loss={:.5g}, accuracy={:.2f}% ({}/{}), time: {}'
.format(epoch, train_loss, accuracy, correct, total,
progress_bar.time_used()))
return