Label Layout in Multi-Class Classification

1

I am a bit new to ML and PyTorch and I’ve been trying to get to grips with this all. I’m trying to train an LSTM that gives me probabilites for my multi-class data. For example, I have a sentence and each sentence can be given a label from 0 to 4 (and only one label, so it’s not multi-label classification).

Right now, I was working from an example (using my own data), but it seems that was for binary classification. When passing y to my model during training, it appears that it is only a tensor of labels corresponding to each sentence like [1,4,2,0,2,2,3,1] (so that would be a batch of 8 sentences, each sentence given one label). If I want to do multi-class, must I pass y so that it looks like [0,0,1,0,0], which correspomds to label 2? I’m just not sure at all what way my labels must be passed for this.

Thanks in advance.

2

The first approach of passing the class indices to the criterion, which should be nn.CrossEntropyLoss or nn.NLLLoss, is correct. Both loss functions expect a target containing the class indices and are not expecting a one-hot encoded target (which would be the second approach).