LSTM input dims error

I have the below classifier model and i want to pass in my input which has 39 features to it as
Shape: torch .rand(1, 1,39) and it works but according to the documentation it should be (1, 39, hidden size).
Why is this so in this case?
How can i choose the hidden size of the LSTM based on?

class SequenceModel(nn.Module):
    def __init__(self, n_features, n_classes=5, n_hidden=256, n_layers=3):

        self.lstm = torch.nn.LSTM(input_size=n_features,
        self.classifier = torch.nn.Linear(n_hidden, n_classes)

    def forward(self, x):
      _, (hidden, _) = self.lstm(x)
      out = hidden[-1] # get the last state of the last layer
      return self.classifier(out)

#Print smaple output
c = torch.nn.CrossEntropyLoss()
sample = next(iter(train_loader))
out = SequenceModel(39, 5)(sample["sequence"])
# print("Model output:",  out.cpu().detach().numpy()) # Change batch size
print("loss:", c(out, sample["label"]))

According to the documentation:

Since you passed batch_first=True, it expects an input of shape (N, L, H_in) where:

  • N is the batch size
  • L is the sequence length
  • H_in is the number of input features

So when you pass torch.rand(1, 1, 39) you are passing it a batch size of 1, a sequence length of 1, and 39 input features, which matches how you set up your LSTM, therefore it works.

I think you could pass an input of size torch.rand(1, X, 39) where X is arbitrary, and it still should work. In other words the sequence length can (and typically is, if you’re using a LSTM) >> 1.

Hidden size should never matter for your input. I assume what you said was a typo and you meant input size. Regardless, the behavior seems consistent with the docs.

Hope this helps!

1 Like