Update torch indexing and use Tensor.cpu() (#91)

* Update dan.py

* Update rnn.py

* Update elmo.py

* Update rnn.py

* Update dan.py

* Update elmo.py

Author: GeneralPoxter

qanta/guesser/dan.py

@@ -466,7 +466,7 @@ def run_epoch(self, iterator: Iterator):
 
             out = self.model(input_dict, lengths_dict, qanta_ids)
             _, preds = torch.max(out, 1)
-            accuracy = torch.mean(torch.eq(preds, page).float()).data[0]
+            accuracy = torch.mean(torch.eq(preds, page).float()).cpu().data
             batch_loss = self.criterion(out, page)
             if is_train:
                 batch_loss.backward()
@@ -476,7 +476,7 @@ def run_epoch(self, iterator: Iterator):
                 self.optimizer.step()
 
             batch_accuracies.append(accuracy)
-            batch_losses.append(batch_loss.data[0])
+            batch_losses.append(batch_loss.cpu().data)
 
         epoch_end = time.time()
 

qanta/guesser/elmo.py

@@ -193,14 +193,14 @@ def run_epoch(self, batches, train=True):
                 self.model.zero_grad()
             out = self.model(x_batch.cuda(), length_batch.cuda())
             _, preds = torch.max(out, 1)
-            accuracy = torch.mean(torch.eq(preds, y_batch).float()).data[0]
+            accuracy = torch.mean(torch.eq(preds, y_batch).float()).cpu().data
             batch_loss = self.criterion(out, y_batch)
             if train:
                 batch_loss.backward()
                 torch.nn.utils.clip_grad_norm(self.model.parameters(), 0.25)
                 self.optimizer.step()
             batch_accuracies.append(accuracy)
-            batch_losses.append(batch_loss.data[0])
+            batch_losses.append(batch_loss.cpu().data)
         epoch_end = time.time()
 
         return np.mean(batch_accuracies), np.mean(batch_losses), epoch_end - epoch_start

qanta/guesser/rnn.py

@@ -346,7 +346,7 @@ def run_epoch(self, iterator: Iterator):
 
             out, hidden = self.model(text, lengths, hidden_init, qanta_ids)
             _, preds = torch.max(out, 1)
-            accuracy = torch.mean(torch.eq(preds, page).float()).data[0]
+            accuracy = torch.mean(torch.eq(preds, page).float()).cpu().data
             batch_loss = self.criterion(out, page)
             if is_train:
                 batch_loss.backward()
@@ -356,7 +356,7 @@ def run_epoch(self, iterator: Iterator):
                 self.optimizer.step()
 
             batch_accuracies.append(accuracy)
-            batch_losses.append(batch_loss.data[0])
+            batch_losses.append(batch_loss.cpu().data)
 
         epoch_end = time.time()