add the kl divergence loss

[haoyuLucas]

1. Implement the KL divergence loss for UDA(unsupervised data augmentation);
2. Support distribution sharpening and confidence masking.

[haoyuLucas]

The Travis build failed because of an error in "tests/core/attention_mechanism_test.py". Any idea why this would happen?

[gpengzhi]

The Travis build failed because of an error in "tests/core/attention_mechanism_test.py". Any idea why this would happen?

The line that triggers the segmentation fault happens to be torch._C._log_api_usage_once("python.nn_module")

I suspect it is something related to sentencepiece==0.1.94 that prevents torch binaries from normally register its resources. Could you downgrade sentencepiece to 0.1.91? This might resolve such issue.

[codecov[bot]]

Codecov Report

Merging #328 (b5a8edb) into master (f016043) will increase coverage by 0.03%. The diff coverage is 95.45%.

@@            Coverage Diff             @@
##           master     #328      +/-   ##
==========================================
+ Coverage   80.14%   80.17%   +0.03%     
==========================================
  Files         134      135       +1     
  Lines       11195    11217      +22     
==========================================
+ Hits         8972     8993      +21     
- Misses       2223     2224       +1     

Impacted Files	Coverage Δ
texar/torch/losses/info_loss.py	95.45% <95.45%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update f016043...b5a8edb. Read the comment docs.

[haoyuLucas]

Downgrading the sentencepiece solved the problem.

[haoyuLucas]

Shall we merge this PR so that we can use this in Forte for Data Augmentation? Thanks!