Ensures mask kwarg is used for output mask propagation

[JyotinderSingh]

This fix addresses a bug in the Layer class where an explicitly passed mask keyword argument was ignored during the computation of the output mask for subsequent layers.

The Problem

Currently, when a layer is called with an explicit mask (e.g., layer(inputs, mask=my_mask)), the framework correctly uses this mask for the layer's internal call() computation. However, when preparing to propagate a mask to the next layer, the framework incorrectly ignores this explicit mask. Instead, it only checks for a _keras_mask attribute attached directly to the input tensor.

If the input tensor has no such attribute, the output mask is computed as None, effectively breaking the mask propagation chain.

Solution

The fix modifies how the previous_mask variable is determined within the Layer.__call__ method. The logic has been updated to prioritize the explicitly passed mask keyword argument.

The new order of operations is:

1. Check if a mask was provided in kwargs. If yes, use it as previous_mask.
2. If no explicit mask is found, fall back to checking for a _keras_mask attribute on the first input tensor.

This ensures that the mask actually used by the layer is the same one used to compute the mask for its output, restoring correct and consistent mask propagation behavior.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 82.71%. Comparing base (ff17868) to head (d827a25).

Additional details and impacted files

@@           Coverage Diff           @@
##           master   #21449   +/-   ##
=======================================
  Coverage   82.71%   82.71%           
=======================================
  Files         565      565           
  Lines       55255    55257    +2     
  Branches     8614     8615    +1     
=======================================
+ Hits        45703    45705    +2     
  Misses       7443     7443           
  Partials     2109     2109           

Flag	Coverage Δ
keras	82.52% <100.00%> (+<0.01%)	⬆️
keras-jax	63.34% <100.00%> (+<0.01%)	⬆️
keras-numpy	58.52% <100.00%> (+<0.01%)	⬆️
keras-openvino	33.78% <100.00%> (+<0.01%)	⬆️
keras-tensorflow	63.75% <100.00%> (+<0.01%)	⬆️
keras-torch	63.36% <100.00%> (+<0.01%)	⬆️

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.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[fchollet]

LGTM, thanks for the fix!