Cannot unsqueeze empty tensor

[ksurya]

System (please complete the following information):

• OS: OSX
• Python version: 3.6.5
• AllenNLP version: 0.8.1
• PyTorch version: 0.4.1

Question
How can I have empty fields.ListField?

I am having fields.ListField([ text_field.empty_field() ]) as suggested in #1391

but Torch raises cannot unsqueeze empty tensor on executing instance.as_tensor_dict() or iterator(..)

[brendan-ai2]

Here's a repro:

import torch
from allennlp.data.token_indexers.single_id_token_indexer import SingleIdTokenIndexer
token_indexers = {"tokens": SingleIdTokenIndexer()}
from allennlp.data.tokenizers.word_tokenizer import WordTokenizer
tokenizer = WordTokenizer()
tokens = tokenizer.tokenize("Foo")
from allennlp.data.fields.text_field import TextField
from allennlp.data.fields.list_field import ListField
text_field = TextField(tokens, token_indexers)
list_field = ListField([text_field.empty_field()])
fields = {'list': list_field}
from allennlp.data.instance import Instance
instance = Instance(fields)
instance.as_tensor_dict()

[brendan-ai2]

I think this bug is due to there not be any non-empty lists in your batch. If there were, the aggregated padding would prevent this bug. I have a speculative fix in #2697, but I'm really not sure if it's the right approach.

[DeNeutoy]

Hi @ksurya - could you let us know a bit more about how you are using the ListField(textfield.empty_field())? We've found that in many cases, you might be wanting to model something in a way which might be easier if approached differently, but equally we don't want allennlp to crash in this scenario either. It would be helpful to see some code, (maybe your DatasetReader) and hear a bit about the model you are trying to build. Thanks!

[ksurya]

Hi guys, appreciate your interest on this issue. I will get back with a Jupyter Notebook reproducing this problem and let you know.