* PyTorch 0.4.1 support * Batched Viterbi decoding (thanks JeppeHallgren)
0.5.0
We are now in beta version. Hopefully there won't be anymore breaking changes. Here are the details of what's new:
* Backward compatibility support for `summed` keyword argument in `forward` method is now removed. Keyword argument of `reduce` is preferred to match PyTorch's convention. * Masks are converted to `LongTensor` before summation to avoid overflow. This is a serious bug for a long input sequence (>= 255). * Minor stylistic refactoring.
0.4.1
* Initialize parameters ini `__init__` (fixes 1) * Refactor tests * Deprecate `summed` in favor of `reduce` (fixes 2) * Rename setup.cfg to .flake8
0.4.0
0.3.2
This just introduces the same changes as version 0.4.1, namely fixing issues 1 and 2, and some refactorings. This backporting is done because some still use PyTorch 0.2 so they can't upgrade to 0.4.1 yet.
0.3.0
PyTorch 0.3.0 is out! We upgrade our PyTorch dependency to version 0.3.0, so now we can use its fancy indexing to get transition score instead of manually broadcasting tensors.
Others
* Fix summing `mask` to get length by first converting it to `LongTensor` to avoid overflow. * Specify `dim` explicitly when `squeeze`-ing to prevent squeezing unintended dimensions.