TensorFlow: Fixed shape issue where unknown shape will be induced from data.

[zeahmed]

This PR fixes #2458.

The problem addressed in this PR are

1. Input is used to determine the shape. If there is one unknown dimension, it is computed from the input data. If there are several unknown dimension, the value computed from the input is equally distributed among unknown dimension.
2. There is no batch dimension assumption on input now.
3. Output shape are computed as previously done. However, if there are more than one unknown dimensions in the shape, a variable length vector is returned then.
4. Created test to check for different type of shapes can be pass in and out of TF.

[codecov]

Codecov Report

Merging #2475 into master will increase coverage by 0.01%.
The diff coverage is 98.79%.

@@            Coverage Diff             @@
##           master    #2475      +/-   ##
==========================================
+ Coverage   71.22%   71.23%   +0.01%     
==========================================
  Files         788      788              
  Lines      141071   141131      +60     
  Branches    16118    16115       -3     
==========================================
+ Hits       100484   100541      +57     
- Misses      36120    36124       +4     
+ Partials     4467     4466       -1

Flag	Coverage Δ
#Debug	71.23% <98.79%> (+0.01%)	⬆️
#production	67.56% <92.3%> (-0.01%)	⬇️
#test	85.35% <100%> (+0.03%)	⬆️

[eerhardt]

[wschin]

For (1), is it consistent to TF convention? #Resolved

[Ivanidzo4ka]

[zeahmed]

For the case of one unknown dimension, yes. If there are more than one unknown dimension than there is no solution. This is just a way to it other way is to make a reshape transformer in ML.NET which user can use to define their input shape which then be mapped to TF shapes easily.

---

In reply to: 461891682 [](ancestors = 461891682)