Trying to backward through the graph a second time (or directly access saved tensors after they have already been freed). Saved intermediate values of the graph are freed when you call .backward() or autograd.grad(). Specify retain_graph=True if you need

autograd
1

I’m trying to optimize two models, a generator and a discriminator both of which use the embeddings generated by the generator model.

This is my code:

generator_source_output = generator_task_model(source_data,
                        token_type_ids = None, 
                        attention_mask = source_masks,
                        labels = source_labels)
            
            source_CE_loss = generator_source_output[0]
            source_last_hidden_state = generator_source_output[2][-1][:, 0, :]
            source_contrastive_loss = compute_contrastive_loss(temperature, 
                                                        source_last_hidden_state.cpu().detach().numpy(), source_labels)
            task_loss = (lambd * source_contrastive_loss) + (1-lambd)*(source_CE_loss)
            
            
             #Invoking the generator (+ task) model on target data
            generator_target_output = generator_task_model(target_data,
                        token_type_ids = None, 
                        attention_mask = target_masks)
            
            target_last_hidden_state = generator_target_output[1][-1][:, 0, :]
            
            # Training the language discriminator 
            d_english = discriminator_model(source_last_hidden_state)
            d_target = discriminator_model(target_last_hidden_state)
            d_english_loss = criterion(d_english.to(device), torch.ones(d_english.size()).to(device))
            d_target_loss = criterion(d_target.to(device), torch.zeros(d_target.size()).to(device))
            discriminator_loss = d_english_loss + d_target_loss
            discriminator_optimizer.zero_grad()
            discriminator_loss.backward()
            discriminator_optimizer.step()

            total_d_loss = total_d_loss + discriminator_loss.item()
            
            # Training the generator
            g_english = discriminator_model(source_last_hidden_state)
            g_target = discriminator_model(target_last_hidden_state)
            g_english_loss = criterion(g_english.to(device), torch.zeros(g_english.size()).to(device))
            g_target_loss = criterion(g_target.to(device), torch.ones(g_target.size()).to(device))
            generator_loss = g_english_loss + g_target_loss + task_loss
            generator_optimizer.zero_grad()
            generator_loss.backward()
            generator_optimizer.step()

I’m trying to optimize two models, a generator and a discriminator both of which use the embeddings generated by the generator model.

This is my code:

generator_source_output = generator_task_model(source_data,
                        token_type_ids = None, 
                        attention_mask = source_masks,
                        labels = source_labels)
            
            source_CE_loss = generator_source_output[0]
            source_last_hidden_state = generator_source_output[2][-1][:, 0, :]
            source_contrastive_loss = compute_contrastive_loss(temperature, 
                                                        source_last_hidden_state.cpu().detach().numpy(), source_labels)
            task_loss = (lambd * source_contrastive_loss) + (1-lambd)*(source_CE_loss)
            
            
             #Invoking the generator (+ task) model on target data
            generator_target_output = generator_task_model(target_data,
                        token_type_ids = None, 
                        attention_mask = target_masks)
            
            target_last_hidden_state = generator_target_output[1][-1][:, 0, :]
            
            # Training the language discriminator 
            d_english = discriminator_model(source_last_hidden_state)
            d_target = discriminator_model(target_last_hidden_state)
            d_english_loss = criterion(d_english.to(device), torch.ones(d_english.size()).to(device))
            d_target_loss = criterion(d_target.to(device), torch.zeros(d_target.size()).to(device))
            discriminator_loss = d_english_loss + d_target_loss
            discriminator_optimizer.zero_grad()
            discriminator_loss.backward()
            discriminator_optimizer.step()

            total_d_loss = total_d_loss + discriminator_loss.item()
            
            # Training the generator
            g_english = discriminator_model(source_last_hidden_state)
            g_target = discriminator_model(target_last_hidden_state)
            g_english_loss = criterion(g_english.to(device), torch.zeros(g_english.size()).to(device))
            g_target_loss = criterion(g_target.to(device), torch.ones(g_target.size()).to(device))
            generator_loss = g_english_loss + g_target_loss + task_loss
            generator_optimizer.zero_grad()
            generator_loss.backward()
            generator_optimizer.step()

On running this, I get the error: Trying to backward through the graph a second time (or directly access saved tensors after they have already been freed). Saved intermediate values of the graph are freed when you call .backward() or autograd.grad(). Specify retain_graph=True if you need to backward through the graph a second time or if you need to access saved tensors after calling backward. on this line generator_loss.backward().

I tried setting discriminator_loss.backward(retain_graph=True) but it takes too long to train and I read that it not the most appropriate solution. So, can someone please help me get rid of this error?