ValueError: lengths array has to be sorted in decreasing order about pytorch-seq2seq HOT 6 CLOSED

@kyteague Thanks for the very detailed issue. Currently in requirements.py we refer to a specific commit of torchtext, did you installed that commit or the latest torchtext?

Nice to know that torchtext is available on pypi now, we will keep this issue in mind when we make seq2seq depend on the pypi version.

from pytorch-seq2seq.

@kyteague If you want to work on this, I think the simplest and intuitive fix is this:

1. Fix a version of torchtext in requirements.py, it doesn't matter if the sorting is reversed or not in torchtext as long as our implementation complies with that version;
2. Intercept the pack_padded_sequence error, and in addition to the original error message, add a suggestion to install the right version of torchtext.

from pytorch-seq2seq.

Unfortunately there are some recent things on master in torchtext that I need, so I can't use the version listed in requirements.txt. Also, requirements.txt isn't mean for distribution. That's what setup.py is for.

I think the best solution is to simply detect the default sorting direction used in the torchtext batching code. That way the seq2seq code remains compatible with both versions of torchtext. It's unfortunate that we have to do this as I think their change was unwise.

from pytorch-seq2seq.

Sounds good.

from pytorch-seq2seq.

@kyteague Looks like torchtext has taken the reverse operation out in their latest version, and the tests of seq2seq past with the pipy version of torchtext. Could you confirm that your use case works with the text branch?

from pytorch-seq2seq.

Fixed in #87

from pytorch-seq2seq.