optimizer.zero_grad()
ending_sim = net(data)
loss = ce_loss(ending_sim, data['labels'])
loss.backward()
optimizer.step()
running_loss_train += loss.item()
if i % PRINT_EVERY == 0:
print(f'Epoch: {epoch+1}, Step: {i}/{n_iteration},\
Runningloss: {running_loss_train/PRINT_EVERY}')
running_loss_train = 0.0
with torch.no_grad():
for i, val_data in enumerate(val_dataloader):
val_ending_sim = net(val_data)
_, predicted = torch.max(val_ending_sim, 1)
metric_acc.update_batch(predicted, val_data['labels'])
val_accuracy = metric_acc.get_metrics_summary()
metric_acc.reset()
if val_accuracy > val_accuracy_prev:
torch.save(net.state_dict(), 'checkpoints/sentiment_finetuned.pth')
print('checkpoint saved')
val_accuracy_prev = val_accuracy
print(
f'============Epoch: {epoch+1}, ValAccuracy: {val_accuracy}================='
)
print('end')
for epoch in range(NUM_EPOCHS):
running_loss_train = 0.0
running_loss_val = 0.0
for i, train_batch in enumerate(train_dataloader):
optimizer.zero_grad()
logits = net(train_batch)
train_loss = ce_loss(logits, train_batch['labels'])
# output, train_loss = utils.run_step(train_batch, net, tokenizer, ce_loss, device)
train_loss.backward()
optimizer.step()
running_loss_train += train_loss.item()
_, predicted = torch.max(logits, 1)
metric_acc.update_batch(predicted, train_batch['labels'])
if i % PRINT_EVERY == 0:
train_accuracy = metric_acc.get_metrics_summary()
metric_acc.reset()
print(
f'Epoch: {epoch+1}, Step: {i}/{n_iteration}, Accuracy: {train_accuracy}, \
Runningloss: {running_loss_train/PRINT_EVERY}')
running_loss_train = 0
with torch.no_grad():
for i, val_batch in enumerate(val_dataloader):
logits = net(val_batch)
val_loss = ce_loss(logits, val_batch['labels'])
# logits, val_loss = utils.run_step(val_batch, net, tokenizer, ce_loss, device)
print('running on cpu')
stories_test = utils.read_data('data/nlp2_test.csv')
embed_file_test = open("data/dictionary_commonsense_test.pickle", 'rb')
embedding_test = pickle.load(embed_file_test)
embed_file_test.close()
test_dataloader = DataLoader(CombinedData(stories_test, embedding_test,
device),
batch_size=BATCH_SIZE,
shuffle=False)
metric_acc = Accuracy()
with torch.no_grad():
metric_acc.reset()
model = CombinedNet(device, pretrained=(False, False, False))
model.load_state_dict(torch.load('checkpoints/combined_model.pth'))
model.to(device)
for i, test_batch in enumerate(test_dataloader):
logits = model(test_batch)
_, predicted = torch.max(logits, 1)
metric_acc.update_batch(predicted, test_batch['labels'])
test_accuracy = metric_acc.get_metrics_summary()
metric_acc.reset()
print(f'======== TestAccuracy: {test_accuracy} ======')