Create a text classifier for the IMDB large movie dataset

[jrxk]

This PR fixes #293.

Description of changes

Added a text classifier for the IMDB large movie dataset based on Texar TF and BERT. The model expects CSV file inputs with columns (content, label, id), which can be generated from a Forte pipeline.

Test Conducted

Added an example and trained locally & got ~84% accuracy.

[hunterhector]

When we said we use Texar to implement the classifier, it means that we won't use tensorflow or pytorch directly. Texar is the layer that hides the actual implementation. And then you can easily switch between pytorch and tensorflow.

Please implement this using Texar-Pytorch instead.

[jrxk]

When we said we use Texar to implement the classifier, it means that we won't use tensorflow or pytorch directly. Texar is the layer that hides the actual implementation. And then you can easily switch between pytorch and tensorflow.

Please implement this using Texar-Pytorch instead.

Sorry, I'm a bit confused. Why do we have to use PyTorch? Is it because we are using PyTorch for the rest of Forte?

I was referring to Texar TF's BertClassfier for this. I will modify the Texar-PyTorch version of the BertClassifier instead. Does that sound good?

[hunterhector]

When we said we use Texar to implement the classifier, it means that we won't use tensorflow or pytorch directly. Texar is the layer that hides the actual implementation. And then you can easily switch between pytorch and tensorflow.
Please implement this using Texar-Pytorch instead.

Sorry, I'm a bit confused. Why do we have to use PyTorch? Is it because we are using PyTorch for the rest of Forte?

I was referring to Texar TF's BertClassfier for this. I will modify the Texar-PyTorch version of the BertClassifier instead. Does that sound good?

1. We would like to use more general interfaces. In this implementation, there are still some specific tensorflow calls.
2. And most of the other implementation examples are texar-pytorch, so let's stick with that.

[codecov]

Codecov Report

Merging #294 (8758d09) into master (66462e7) will decrease coverage by 0.45%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master     #294      +/-   ##
==========================================
- Coverage   79.81%   79.36%   -0.46%     
==========================================
  Files         154      150       -4     
  Lines        9919     9722     -197     
==========================================
- Hits         7917     7716     -201     
- Misses       2002     2006       +4     

Impacted Files	Coverage Δ
...te/processors/ir/elastic_search_index_processor.py	59.09% <0.00%> (-34.02%)	⬇️
forte/processors/base/index_processor.py	50.00% <0.00%> (-13.64%)	⬇️
forte/processors/base/query_processor.py	89.47% <0.00%> (-1.84%)	⬇️
forte/data/base_pack.py	80.10% <0.00%> (-1.58%)	⬇️
forte/processors/writers.py	55.00% <0.00%> (-1.53%)	⬇️
forte/processors/ir/elastic_search_processor.py	43.24% <0.00%> (-1.21%)	⬇️
forte/data/data_pack.py	81.81% <0.00%> (-0.25%)	⬇️
forte/processors/base/__init__.py	100.00% <0.00%> (ø)
...orte/processors/data_augment/data_selector_test.py
forte/processors/data_augment/selectors.py
... and 4 more

---

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 66462e7...3e0a1e8. Read the comment docs.

[jrxk]

Reimplement with texar-pytorch.

[hunterhector]

Could you open this as another PR and make sure this file is not included here directly?

examples/text_classification/data/IMDB_raw/train_id_list.txt

Files in GitHub will be there permanently once committed. So this file would increase the size of the project permanently. Can you simply provide a download link?

When you open another PR, make sure this file is not in your commit history of the branch. This means that you should create a fresh branch from master, add your changes to that branch, and do not include train_id_list.

Note that git rm on the current branch will not reduce the size of this PR.