Pack_padded_sequence,flatten_parameters and pad_packed_sequence?

[AutumnRoom]

Are the functions pack_padded_sequence, flatten_parameters, and pad_packed_sequence in PyTorch needed for BiLSTM calculations in Burn?

[laggui]

We do not have any specific RNN utilities at this time like the one you mentioned from pytorch.

If you have inputs of varying sequence lengths that you want to pad, you can use tensor.pad(padding, value) to pad to the max sequence length in the batch. But there is no packing.

Regarding flatten_parameters, it seems to be an optimization that specifically targets cuDNN:

https://github.com/pytorch/pytorch/blob/1eba9b3aa3c43f86f4a2c807ac8e12c4a7767340/torch/nn/modules/rnn.py#L261-L281

We don't depend on cuDNN for better portability, so we don't have this option. Maybe in the future we can have similar careful optimizations, but at this time we don't have this.

Side note: you might be interested in this stacked LSTM implementation.