PyTorch 0.4.1 support

There isn't much new stuff in this release:

• PyTorch 0.4.1 support
• Batched Viterbi decoding (thanks @JeppeHallgren)

We are now in beta!

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.

Fixing issues and refactorings

• Initialize parameters ini __init__ (fixes #1)
• Refactor tests
• Deprecate summed in favor of reduce (fixes #2)
• Rename setup.cfg to .flake8

Backport from v0.4.1

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.

Upgrade to PyTorch 0.3.0 and minor changes

Upgrade to PyTorch 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.