Set evaluation epoch iterator object to None for each model.fit call

[SuryanarayanaY]

When using model.fit() with validation_data, during first call to fit() it will create an EpochIterator object for evaluation and cache it. This cache will be deleted after completion of all epochs.

However if we try to change the validation_data shapes for any reason and called model.fit() again it will raise an exception like Graph execution error which is intended. After this if we call model.fit() with correct validation_data that of first training call it won't work.

The reason being is after first call is success evaluation EpochIterator will be deleted at the end and during second training call a new EpochIterator will be generated again with changed shape.But during evaluate() call it will raise an exception terminating the process without deleting evaluation EpochIterator object and it remains in cache. During third call though with correct validation_data it will not create new EpochIterator object as one already exists in cache which makes it fail again.

Hence proposing a fix for this with this commit.

Fixes #18653

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 0% with 2 lines in your changes missing coverage. Please review.

Project coverage is 57.73%. Comparing base (6e54f7f) to head (8a468cf).
Report is 1069 commits behind head on master.

Files with missing lines	Patch %	Lines
keras/backend/jax/trainer.py	0.00%	1 Missing ⚠️
keras/backend/tensorflow/trainer.py	0.00%	1 Missing ⚠️

❗ There is a different number of reports uploaded between BASE (6e54f7f) and HEAD (8a468cf). Click for more details.

HEAD has 6 uploads less than BASE

Flag	BASE (6e54f7f)	HEAD (8a468cf)
keras	4	1
keras-jax	1	0
keras-torch	1	0
keras-tensorflow	1	0

Additional details and impacted files

@@             Coverage Diff             @@
##           master   #18659       +/-   ##
===========================================
- Coverage   78.50%   57.73%   -20.77%     
===========================================
  Files         334      335        +1     
  Lines       32596    32946      +350     
  Branches     6376     6450       +74     
===========================================
- Hits        25589    19023     -6566     
- Misses       5452    12497     +7045     
+ Partials     1555     1426      -129     

Flag	Coverage Δ
keras	57.73% <0.00%> (-20.66%)	⬇️
keras-jax	?
keras-numpy	57.73% <0.00%> (-0.22%)	⬇️
keras-tensorflow	?
keras-torch	?

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 fix -- it looks good to me. Please add a simple unit test for the failure case.

[fchollet]

LGTM, thank you