def train_classifier(self, max_epochs, train_dataloader,
validation_dataloader, comparison):
epochs = []
train_losses = []
validation_accs = []
early_stopping = EarlyStopping('{}/{}_classifier.pt'.format(
self.model_name, self.dataset_name))
for epoch in range(max_epochs):
if early_stopping.early_stop:
break
for batch_idx, (data, labels) in enumerate(train_dataloader):
self.Classifier.train()
data = data.to(self.device)
labels = labels.to(self.device)
self.Classifier_optim.zero_grad()
preds = self.Classifier(data)
loss = self.Classifier_criterion(preds, labels)
loss.backward()
self.Classifier_optim.step()
if comparison:
epochs.append(epoch)
train_losses.append(loss.item())
validation_accs.append(
accuracy(self.Classifier, validation_dataloader,
self.device))
val = accuracy(self.Classifier, validation_dataloader, self.device)
print('Supervised Epoch: {} Validation acc: {}'.format(epoch, val))
early_stopping(1 - val, self.Classifier)
if early_stopping.early_stop:
early_stopping.load_checkpoint(self.Classifier)
return epochs, train_losses, validation_accs
def train_classifier(self, max_epochs, train_dataloader,
validation_dataloader):
epochs = []
train_losses = []
validation_accs = []
early_stopping = EarlyStopping('{}/{}_inner.pt'.format(
self.state_path, self.model_name))
for epoch in range(max_epochs):
if early_stopping.early_stop:
break
train_loss = 0
for batch_idx, (data, labels) in enumerate(train_dataloader):
self.SDAEClassifier.train()
data = data.to(self.device)
labels = labels.to(self.device)
self.optimizer.zero_grad()
predictions = self.SDAEClassifier(data)
loss = self.criterion(predictions, labels)
loss.backward()
self.optimizer.step()
train_loss += loss.item()
if validation_dataloader is not None:
acc = accuracy(self.SDAEClassifier, validation_dataloader,
self.device)
validation_accs.append(acc)
early_stopping(1 - acc, self.SDAEClassifier)
epochs.append(epoch)
train_losses.append(train_loss / len(train_dataloader))
early_stopping.load_checkpoint(self.SDAEClassifier)
return epochs, train_losses, validation_accs
def full_train(self, unsupervised_dataset, train_dataset,
validation_dataset):
self.reset_model()
# TODO: don't use arbitrary values for batch size
unsupervised_dataloader = DataLoader(dataset=unsupervised_dataset,
batch_size=180,
shuffle=True)
supervised_dataloader = DataLoader(dataset=train_dataset,
batch_size=20,
shuffle=True)
validation_dataloader = DataLoader(
dataset=validation_dataset,
batch_size=validation_dataset.__len__())
for epoch in range(50):
self.train_one_epoch(epoch, supervised_dataloader,
unsupervised_dataloader)
print('Epoch: {} Validation Acc: {}'.format(
epoch,
trainingutils.accuracy(self.D, validation_dataloader,
self.device)))
def test_model(self, test_dataloader):
return accuracy(self.SDAEClassifier, test_dataloader, self.device)
def full_test(self, test_dataset):
test_dataloader = DataLoader(dataset=test_dataset,
batch_size=test_dataset.__len__())
return trainingutils.accuracy(self.D, test_dataloader, self.device)