Outputs for classification model are logits or y_pred?

[pusapatiakhilraju]

I am a little confused about the the outputs of a forward pass. In Section-4 (Neural Network Classification) it was mentioned that for a classification, the forward pass will output logits. these logits are then passed through activation functions to make them into y_pred.

However in the computer vision section, the code is as follows:

# training Loop

for epoch in tqdm(range(epochs)):

    print(f"Epoch: {epoch}\n-------")
    train_loss = 0
    # Add a loop to loop through training batches
    
    for batch, (X, y) in enumerate(train_dataloader):
        model_0.train() 
        # 1. Forward pass
        y_pred = model_0(X)
        
        # 2. Calculate loss (per batch)
        loss = loss_fn(y_pred, y)
        train_loss += loss # accumulatively add up the loss per epoch 

Can someone explain why the outputs of this forward pass y_pred and not logits?

See the full code here: https://www.learnpytorch.io/03_pytorch_computer_vision/#33-creating-a-training-loop-and-training-a-model-on-batches-of-data

[pusapatiakhilraju]

Can someone please help me further understand this?

[mrdbourke]

Hi @pusapatiakhilraju,

This looks like it may be a naming issue.

E.g. y_pred in the code you're sharing (see here in the book: https://www.learnpytorch.io/03_pytorch_computer_vision/#33-creating-a-training-loop-and-training-a-model-on-batches-of-data) could also be called y_logits.

Because the loss_fn = nn.CrossEntropyLoss(), this loss function can take raw logits (the raw output of the model, which is called y_pred in this case) directly.

In the code above, if you named y_pred as y_logits, you would get the same results.

[pusapatiakhilraju]

That makes complete sense now! legend! thank you.

[mrdbourke]

This also may help with confusion around torch.softmax() - #314