Clojure example for fixed label-width captcha recognition

[kedarbellare]

Description

Adds an example for training a captcha OCR model and using the learned model for performing inference.

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
• Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
• Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
• Code is well-documented:
• For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable
• To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

• Captcha OCR training and inference with unit tests for training

Comments

I ran this on both osx and ubuntu linux but only with :cpu. However, I encountered memory issues while trying to train on a :gpu:

Exception in thread "main" org.apache.mxnet.MXNetError: [19:18:33] src/storage/./pooled_storage_manager.h:143: cudaMalloc failed: out of memory

Reviewers

@gigasquid

[gigasquid]

Cool! Excited to look at this closer 😸

[Roshrini]

@kedarbellare Thanks for this contribution! On which gpu machine, did you run this example?

@mxnet-label-bot Add [pr-awaiting-review, Clojure]

[kedarbellare]

@Roshrini : I ran this on my personal desktop with Nvidia GTX 1070 with 8GB memory (ubuntu 18.04, cuda-9.0). I tried varying the batch sizes but that didn't help. Let me know if you want more details.

[gigasquid]

@kedarbellare - getting to try this out - how long did your training take locally (how many epochs) ?

• Never mind I just found the info in the README :)

prefix `ocr-`. I was able to achieve an exact match accuracy of ~0.954 and
~0.628 on training and validation data respectively.

[aaronmarkham]

I left some suggested edits.
Also, it would be nice to have the files with code to have comments or some indication of what that file does.

[kedarbellare]

It's pretty slow with CPU training. On my macbook pro, with lein train (default one CPU training) the speedometer shows ~25 samples/sec (end-to-end training took over an hour for 10 epochs). My beefier deskop gets double that with CPU training and GPU training shows over 600 samples/sec.

[kedarbellare]

@aaronmarkham I followed your suggestions for the README. I'll add more comments in the code in a subsequent commit.

[gigasquid]

Great addition! I tested it out and everything works perfectly.
Thank you for this contribution 👍

[gigasquid]

Great way to normalize the pixels 💯