LstmCellWithProjection can't be wrapped with PytorchSeq2seqWrapper

[kdexd]

Steps to Reproduce

>>> import torch
>>> from allennlp.modules.lstm_cell_with_projection import LstmCellWithProjection
>>> from allennlp.modules.seq2seq_encoders import PytorchSeq2SeqWrapper
>>>
>>> input = torch.Tensor([ 
...      [2, 4, 3, 1],
...      [2, 3, 1, 0]
... ]).long()  # shape: (batch_size=2, max_seq_len=4)
>>> input_mask = (input != 0)  # assume padding_index = 0
>>> 
>>> embedder = torch.nn.Embedding(5, 10)
>>> lstm_cell = LstmCellWithProjection(input_size=10, hidden_size=4, cell_size=4)
>>> wrapper = PytorchSeq2SeqWrapper(lstm_cell)
>>> 
>>> embedded = embed(input)
>>> wrapper(embedded, input_mask)

This is the error I get:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/kd/miniconda3/envs/allennlp/lib/python3.6/site-packages/torch/nn/modules/module.py", line 489, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/kd/miniconda3/envs/allennlp/lib/python3.6/site-packages/allennlp/modules/seq2seq_encoders/pytorch_seq2seq_wrapper.py", line 83, in forward
    self.sort_and_run_forward(self._module, inputs, mask, hidden_state)
  File "/home/kd/miniconda3/envs/allennlp/lib/python3.6/site-packages/allennlp/modules/encoder_base.py", line 116, in sort_and_run_forward
    module_output, final_states = module(packed_sequence_input, initial_states)
  File "/home/kd/miniconda3/envs/allennlp/lib/python3.6/site-packages/torch/nn/modules/module.py", line 489, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/kd/miniconda3/envs/allennlp/lib/python3.6/site-packages/allennlp/modules/lstm_cell_with_projection.py", line 122, in forward
    batch_size = inputs.size()[0]
AttributeError: 'PackedSequence' object has no attribute 'size'

Expected behavior

The inline comments in LstmCellWithProjection.__init__() state that it should be fit to be wrapped by a PytorchSeq2SeqWrapper. I am not sure which one of the two options is feasible/sensible — removing those inline comments, or introducing support as described.

System (please complete the following information):

• OS: Linux (17.10)
• Python version: 3.6.7
• AllenNLP version: 0.8.1
• PyTorch version: 1.0.0

[nelson-liu]

I think introducing support (by enabling forward to take PackedSequence input) would be the right fix here.

[kdexd]

I'll be happy to take this up after 23 January if it still remains open. If someone is willing to take it up before that, feel free to.