Allow tensors in tf.Datasets to have different dimensions.

[hertschuh]

The shape for the tf.TensorSpec for the tf.Dataset is determined by inspecting several batches and keeping dimensions that are common.

Fixes #19124

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 78.37838% with 8 lines in your changes missing coverage. Please review.

Project coverage is 75.73%. Comparing base (c8700f4) to head (228a748).
Report is 632 commits behind head on master.

Files with missing lines	Patch %	Lines
keras/trainers/data_adapters/data_adapter_utils.py	73.91%	3 Missing and 3 partials ⚠️
...s/trainers/data_adapters/generator_data_adapter.py	83.33%	0 Missing and 1 partial ⚠️
...rainers/data_adapters/torch_data_loader_adapter.py	80.00%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #19318      +/-   ##
==========================================
- Coverage   80.14%   75.73%   -4.41%     
==========================================
  Files         341      366      +25     
  Lines       36163    40193    +4030     
  Branches     7116     7814     +698     
==========================================
+ Hits        28982    30442    +1460     
- Misses       5578     8064    +2486     
- Partials     1603     1687      +84     

Flag	Coverage Δ
keras	75.59% <78.37%> (-4.40%)	⬇️
keras-jax	59.95% <78.37%> (-3.11%)	⬇️
keras-numpy	54.54% <78.37%> (-2.54%)	⬇️
keras-tensorflow	61.46% <78.37%> (-3.19%)	⬇️
keras-torch	60.56% <78.37%> (-3.31%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[fchollet]

Thanks for the PR!

[fchollet]

Why not just 2?

[fchollet]

Also, these are batches, not samples

[hertschuh]

I don't know, I just figured the odds of detecting different shapes would be better with more examples (but it will never be 100% correct).

I will rename.

Also, I need to add unit tests.

[fchollet]

The 2 main use cases are for images and sequences. In both cases you have a very high likelihood that 2 consecutive batches will have different image sizes or sequence lengths if they are dynamic. So IMO 2 is good enough and minimizes overhead. Of course, it's not 100% accurate, as you say.

[fchollet]

LGTM, thank you!