Added TensorFlow module + BERT model

[tovbinm]

Related issues
TransmogrifAI supports neither training nor inferencing deep learning models #248

Describe the proposed solution

• Add additional TensorFlow module

• Add ability to create, load and run TensorFlow models

• Add BERT model

• Compare with JohnSnow performance - https://github.com/JohnSnowLabs/spark-nlp/blob/master/src/main/scala/com/johnsnowlabs/nlp/embeddings/BertEmbeddings.scala. Update: JohnSnow dev team investigation is here - Performance of BertEmbeddings model JohnSnowLabs/spark-nlp#570

• TBD BERT model inference is very slow. Additional investigation is required. And since JohnSnow library works better, perhaps we should wrap around it?

Describe alternatives you've considered
NA

[codecov]

Codecov Report

Merging #355 into master will decrease coverage by 1.52%.
The diff coverage is 29.59%.

@@            Coverage Diff             @@
##           master     #355      +/-   ##
==========================================
- Coverage   80.29%   78.77%   -1.53%     
==========================================
  Files         339      342       +3     
  Lines       11240    11537     +297     
  Branches      374      367       -7     
==========================================
+ Hits         9025     9088      +63     
- Misses       2215     2449     +234

Impacted Files	Coverage Δ
...ain/scala/com/salesforce/op/aggregators/Maps.scala	96.55% <ø> (ø)	⬆️
...la/com/salesforce/op/test/TestFeatureBuilder.scala	100% <ø> (ø)	⬆️
...e/op/stages/impl/feature/SmartTextVectorizer.scala	95.61% <ø> (+0.87%)	⬆️
...alesforce/op/aggregators/TimeBasedAggregator.scala	0% <ø> (ø)	⬆️
...la/com/salesforce/op/features/FeatureBuilder.scala	28.27% <0%> (-0.4%)	⬇️
...m/salesforce/op/utils/stages/NameDetectUtils.scala	0% <0%> (ø)
...scala/com/salesforce/op/features/types/Lists.scala	58.33% <0%> (-19.45%)	⬇️
.../scala/com/salesforce/op/dsl/RichTextFeature.scala	79.16% <0%> (-1.12%)	⬇️
...rce/op/stages/impl/feature/HumanNameDetector.scala	0% <0%> (ø)
...s/impl/feature/OPCollectionHashingVectorizer.scala	96.55% <100%> (+0.07%)	⬆️
... and 41 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 bcd5070...c4bc447. Read the comment docs.

[tovbinm]

I think it makes sense to replace BERT with ALBERT - https://ai.googleblog.com/2019/12/albert-lite-bert-for-self-supervised.html

@kinfaikan ?

[tovbinm]

This isn't going anywhere. Closing for now.