Fix bugs with Mean, Accuracy and BinaryAccuracy metrics.

[hertschuh]

• reduce_to_samplewise_values would not reduce sample_weights correctly because the number of dimensions of values was checked.
• reduce_to_samplewise_values needs to explicitely broadcast sample_weights. Before, it was implicitly broadcast in the multiplication with values. However, the explicit broadcast is needed for the computation of num_samples for the averaging to be correct. This causes a bug when sample_weights is of rank 2 or more and a broadcast happens when doing the multiplication. This logic existed in tf_keras: https://github.com/keras-team/tf-keras/blob/master/tf_keras/metrics/base_metric.py#L508
• Accuracy and BinaryAccuracy were doing a mean reduction too early, before multiplying by sample_weights. This matters when the rank of sample_weights is the same as y_true and y_pred.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 73.92%. Comparing base (4551644) to head (604094c).
Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #19847      +/-   ##
==========================================
- Coverage   78.85%   73.92%   -4.93%     
==========================================
  Files         498      498              
  Lines       45853    45855       +2     
  Branches     8449     8450       +1     
==========================================
- Hits        36156    33899    -2257     
- Misses       7995    10312    +2317     
+ Partials     1702     1644      -58     

Flag	Coverage Δ
keras	73.81% <100.00%> (-4.90%)	⬇️
keras-jax	?
keras-numpy	56.63% <100.00%> (-0.01%)	⬇️
keras-tensorflow	63.68% <100.00%> (-0.01%)	⬇️
keras-torch	62.36% <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.

[fchollet]

Thanks for the PR!

[fchollet]

accuracy and binary_accuracy should reduce on the last dimension at this stage (unless the input has rank 1, in which case the last dimension is the batch dimension). All metrics functions do this.

So if your inputs have rank 2, then the function will return one accuracy scalar score per entry in the batch (e.g. (32, 4) -> (32,))

[hertschuh]

Yes, the reduction is applied in a generic way here (if weights have a smaller rank):
https://github.com/keras-team/keras/blob/master/keras/src/metrics/reduction_metrics.py#L29-L31
and/or here:
https://github.com/keras-team/keras/blob/master/keras/src/metrics/reduction_metrics.py#L39-L41
So it does work as expected for rank 2 and above.

Let me make sure it works as expected for rank 1 ys and add unit tests.

The issue is that we don't prevent weights that are the same rank as y_true and y_pred. What it was doing before was multiplying the average of the equality by the average of the weights. Now it's averaging the equality multiplied by the weights.

The full sequence is now:

1. compute equality (called values)
2. tweak values and sample_weights to get to the same rank, which can involve a mean on the last few dimensions of values
3. multiply values by sample_weights
4. mean reduction of values and sample_weights to rank 1
5. sum values (total) and sample_weights (num_samples)
6. divide total by num_samples.

Before, there was an extra mean between 1. and 2. It did not matter if rank(sample_weights) < rank(values) because step 2. does a mean to lower the rank of values.

Note that it now follows the same pattern as tf_keras where the accuracy function is not the one doing the reduction: https://github.com/keras-team/tf-keras/blob/master/tf_keras/metrics/accuracy_metrics.py#L354-L364
but the reduction is done generically in the Reduce class: https://github.com/keras-team/tf-keras/blob/master/tf_keras/metrics/base_metric.py#L520-L522 and https://github.com/keras-team/tf-keras/blob/master/tf_keras/metrics/base_metric.py#L525

[fchollet]

LGTM, thanks for the explanation!