Coreference resolution fails on short texts

[darabos]

Coreference resolution seems to work fine for longer texts. But with short texts I get various errors. (I'm on Debian, Python 3.7.0, PyTorch 1.0.1, AllenNLP 0.8.4.)

from allennlp.predictors import predictor
coref = predictor.Predictor.from_path('https://s3-us-west-2.amazonaws.com/allennlp/models/coref-model-2018.02.05.tar.gz')
coref.predict('I am Joe.')

...
~/anaconda3/lib/python3.7/site-packages/allennlp/modules/seq2vec_encoders/cnn_encoder.py in forward(self, tokens, mask)
    103             convolution_layer = getattr(self, 'conv_layer_{}'.format(i))
    104             filter_outputs.append(
--> 105                     self._activation(convolution_layer(tokens)).max(dim=2)[0]
    106             )
    107 

~/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py in __call__(self, *input, **kwargs)
    487             result = self._slow_forward(*input, **kwargs)
    488         else:
--> 489             result = self.forward(*input, **kwargs)
    490         for hook in self._forward_hooks.values():
    491             hook_result = hook(self, input, result)

~/anaconda3/lib/python3.7/site-packages/torch/nn/modules/conv.py in forward(self, input)
    185     def forward(self, input):
    186         return F.conv1d(input, self.weight, self.bias, self.stride,
--> 187                         self.padding, self.dilation, self.groups)
    188 
    189 

RuntimeError: $ Torch: invalid memory size -- maybe an overflow? at /opt/conda/conda-bld/pytorch_1550784378911/work/aten/src/TH/THGeneral.cpp:188

Same error with I am Bob.. But no error with I am David..

There is a slightly different error too, which also affects short texts:

coref.predict('I have a head.')

...
~/anaconda3/lib/python3.7/site-packages/allennlp/modules/time_distributed.py in forward(self, pass_through, *inputs, **kwargs)
     49             reshaped_kwargs[key] = value
     50 
---> 51         reshaped_outputs = self._module(*reshaped_inputs, **reshaped_kwargs)
     52 
     53         if some_input is None:

~/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py in __call__(self, *input, **kwargs)
    487             result = self._slow_forward(*input, **kwargs)
    488         else:
--> 489             result = self.forward(*input, **kwargs)
    490         for hook in self._forward_hooks.values():
    491             hook_result = hook(self, input, result)

~/anaconda3/lib/python3.7/site-packages/allennlp/modules/seq2vec_encoders/cnn_encoder.py in forward(self, tokens, mask)
    103             convolution_layer = getattr(self, 'conv_layer_{}'.format(i))
    104             filter_outputs.append(
--> 105                     self._activation(convolution_layer(tokens)).max(dim=2)[0]
    106             )
    107 

RuntimeError: cannot perform reduction function max on tensor with no elements because the operation does not have an identity

No error for I have hands. though.

I've done a little debugging, and in the second exception the problem seems to be that convolution_layer has a 5-element window. So if tokens is smaller than that, the output is empty.

[kl2806]

@DeNeutoy, can you look at this?

[kl2806]

This is the same issue as a previous one https://github.com/allenai/allennlp/pull/2541, you can make a similar fix for the coref model. PR welcome!