Add Keras model test with TF2 model

[lantiga]

TF2 supports freezing models, which can ben ran on TF1.15.

This PR adds a test on a MNIST-like model generated with Keras, exported through TF2 and that is ran on RedisAI.

[lantiga]

@hhsecond you might want to add this use case to your export library

[codecov]

Codecov Report

Merging #315 into master will not change coverage.
The diff coverage is n/a.

@@          Coverage Diff           @@
##           master    #315   +/-   ##
======================================
  Coverage    57.1%   57.1%           
======================================
  Files          26      26           
  Lines        5133    5133           
======================================
  Hits         2931    2931           
  Misses       2202    2202

---

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 2bd2eed...14247f2. Read the comment docs.

[hhsecond]

Awesome. I'll do that. Assuming they have added it in 2.1?
Also, I believe this doesn't mean we can run all 2.x models in 1.15 because there are 2.x specific ops.

I know this because we have built the same functionality using 1.15 earlier. i.e a SavedModel directory from 2.x can be opened using 1.15 and then convert it to frozen model from there. And I had issues converting few models from 2.x to 1.15

[hhsecond]

Maybe with this new utility, they might convert it to a backward-compatible representation? Very unlikely

[lantiga]

This is slightly different though since it's TF2 doing the freezing. Since the frozen model is only useful for running in TF1, I'm expecting TF2 to fail if there's no corresponding op in GraphDef, so at least it would fail early on the TF2 side.
Can we try challenging the freezing?

[hhsecond]

That's a fair point. Let me find the models I tried with the old technique.