I am trying to replicate this paper.
Since I am using cross-entropy, I won’t add the sigmoid function when training.
The input size of the lstm is 25088 but I am not sure what hidden_size and lstm_layers to use in order to be able to use cross-entropy.
nn.CrossEntropyLoss expects a model output in the shape [batch_size, nb_classes] so you would have to make sure your model creates such outputs containing logits.
Depending on the model architecture (and if the LSTM is the last layer), the hidden_size could define the number of features and could thus match the number of classes. However, if the LSTM output is passed to another layer (e.g. nn.Linear) then this layer would have to output the class logits.
Thank you, I’ve used another nn.Linear layer and fixed that.