Torch.jit.script for torch.nn.utils.rnn.pack_padded_sequence

1

I have struggled with the problem several days.
Now I implemented torch.nn.utils.rnn.pack_padded_sequence with variable batch size and sequence length. The model is already trained and work fine in python. I want to embed it in my c++ project. I used torch.jit.script but it always inform me that
for operator (Tensor 0, Tensor 1, Tensor 2) -> Tensor:
expected a value of type Tensor for argument ‘2’ but found bool
I changed the batch_first to tensor.ones() and in forward prediction, it informs me that
argument ‘batch_first’ (position 3) must be bool
it is really ridiculous. And most examples online are trace implementation.
Can somebody help me or give me some hints?
Thx a lot!

1 Like
2

Indeed, pack_padded_sequences does not work with TorchScript yet.

3

I’m so sad :frowning: This is Important for me…

4

I’m thinking about opening an issue in Github, just to make sure PT devs have this under their radar. Hopefully the necessary change is easy to implement (although I have no idea on how TorchScript internals work…)

5

I have opened an issue in Github to track this: https://github.com/pytorch/pytorch/issues/21282