Scripts to run logistic regression and sentence classification examples

[kedarbellare]

Summary

• This PR adds scripts to run the logistic regression and sentence classification examples. The sentence classification script in particular downloads the data set required, the GloVe word vectors and runs the sentence classification using CNN 1D model.
• The PR also makes some minor changes to the README and the example files. It also adds some profiles to project.clj to be able to run the examples.
• Finally it modifies topological sort to use loop-recur instead of recursion so that compute works with deep DAGs.

Testing

• lein test
• Ran ./scripts/logistic-regression-example.sh without any issues
• Ran ./scripts/sentence-classification-example.sh which downloads the data and the words vectors (which can take up to 6G of disk space). While running the cnn 1D model with 10k examples, I get a test accuracy of 85% which is quite different from what's reported in http://aria42.com/blog/2017/11/Flare-Clojure-Neural-Net (not sure what accounts for the difference). Also ran bilstm model but didn't get improved results.

[aria42]

Hey @kedarbellare, thanks for the PR. I've uploaded the train data and test data used for the blog post and here's the CLI output for train/test on a jar built from master head (3e5baa17153b), I can recreate the numbers same as the blog post (note the reported numbers are averages) after a few iterations. I think there are a few differences here:

• You seem to be using the 4B word glove vectors instead of 6B
• I'm white-space tokenizing instead of the PTB style of tokenization your scripts seem to do.
• My randomly sampled train/test examples are different

I'm happy to take this PR with a few changes

• Can we make the script repro my results if we ignore variance from the train/test sample?
• I don't think we need filter-glove
• What's the specific issue with topographic sort? can we add a test that would fail before the change to the existing tests?

[kedarbellare]

hey @aria42, i could reproduce the results using the train/test files you provided and the glove 6B 300d vectors (https://gist.github.com/kedarbellare/764b16e813d5e643a1b14c52f53d08ee). i can modify the tokenization scheme and remove filter-glove. however, there seems to be some discrepancy in the original data and your train/test files. the acl data seems to have much longer texts compared to your data. i also tried:

$ grep combination data/sentiment-train10k.txt | grep fiennes
1 interacting eyeball-to-eyeball and toe-to-toe , hopkins and norton are a winning combination -- but fiennes steals 'red dragon' right from under their noses .
$ grep -R -i combination data/aclImdb/train/pos/ | grep -i fiennes
$

i'm not sure what i'm doing wrong :/

regarding topographic sort, the main issue is that a StackOverflowError is thrown for deep DAGs (e.g. LSTMs with sentences containing many tokens) due to the recursive call. let me add a test that captures this.

[aria42]

Ok, not sure why the sentences are shorter in my sample; it's possible I intentionally selected below a length for some reason. It only matters for reproducing the numbers in the blog post, so I'm fine with accepting this as is. Thanks!