Some bugs in generation phase #16
Description
Hi,
Thanks for your excellent work.
I downloaded the data and processed them as in the instructions (few of them cannot be downloaded and I modified the data processing part to skip on these ones, I guess should not be a big issue. ) Then I trained a new model and everything looks good. However, when I run with the generation script, it gives me the following errors:
The command:
CUDA_VISIBLE_DEVICES=1 python XSum-ConvS2S/generate.py ./data-convs2s --path ./checkpoints-convs2s/checkpoint-best.pt --batch-size 1 --beam 10 --replace-unk --source-lang document --target-lang summary > test-output-convs2s-checkpoint-best.pt
Output:
0%| | 0/11334 [00:00<?, ?it/s]/home/rasmlnlp/zhiyu/anaconda3/envs/tf1.12/lib/python3.5/site-packages/torch/autograd/function.py:41: UserWarning: mark_shared_storage is deprecated. Tensors with shared storages are automatically tracked. Note that calls to set_() are not tracked
'mark_shared_storage is deprecated. '
Traceback (most recent call last):
File "XSum-ConvS2S/generate.py", line 161, in
main(args)
File "XSum-ConvS2S/generate.py", line 96, in main
for sample_id, src_tokens, target_tokens, hypos in translations:
File "/home/rasmlnlp/zhiyu/XSum/XSum-ConvS2S/fairseq/sequence_generator.py", line 77, in generate_batched_itr
prefix_tokens=s['target'][:, :prefix_size] if prefix_size > 0 else None,
File "/home/rasmlnlp/zhiyu/XSum/XSum-ConvS2S/fairseq/sequence_generator.py", line 90, in generate
return self._generate(src_tokens, src_lengths, beam_size, maxlen, prefix_tokens)
File "/home/rasmlnlp/zhiyu/XSum/XSum-ConvS2S/fairseq/sequence_generator.py", line 250, in _generate
tokens[:, :step+1], encoder_outs, incremental_states)
File "/home/rasmlnlp/zhiyu/XSum/XSum-ConvS2S/fairseq/sequence_generator.py", line 413, in _decode
decoder_out, attn = model.decoder(tokens, encoder_out, incremental_states[model])
File "/home/rasmlnlp/zhiyu/anaconda3/envs/tf1.12/lib/python3.5/site-packages/torch/nn/modules/module.py", line 477, in call
result = self.forward(*input, **kwargs)
File "/home/rasmlnlp/zhiyu/XSum/XSum-ConvS2S/fairseq/models/fconv.py", line 266, in forward
x, attn_scores = attention(x, target_embedding, (encoder_a, encoder_b))
File "/home/rasmlnlp/zhiyu/anaconda3/envs/tf1.12/lib/python3.5/site-packages/torch/nn/modules/module.py", line 477, in call
result = self.forward(*input, **kwargs)
File "/home/rasmlnlp/zhiyu/XSum/XSum-ConvS2S/fairseq/models/fconv.py", line 160, in forward
x = self.bmm(x, encoder_out[0])
File "/home/rasmlnlp/zhiyu/anaconda3/envs/tf1.12/lib/python3.5/site-packages/torch/nn/modules/module.py", line 477, in call
result = self.forward(*input, **kwargs)
File "/home/rasmlnlp/zhiyu/XSum/XSum-ConvS2S/fairseq/modules/beamable_mm.py", line 34, in forward
input1 = input1[:, 0, :].unfold(0, beam, beam).transpose(2, 1)
RuntimeError: invalid argument 3: out of range at /opt/conda/conda-bld/pytorch_1535490206202/work/aten/src/THC/generic/THCTensor.cpp:444
Is there anything bugged here? Or is there any other way to reproduce the results in the paper? (The training only shows the loss or ppl on the validation set, not ROUGE results.)
Thank you.