[Bug]: - MuFidelity does not work for Tabular data and time series

[AntoninPoche]

Module

None

Current Behavior

When applying MuFidelity on tabular data or time series, the number of features set to 0 with the mask does not correspond to subset_percent. Indeed, the formula to compute the subset_size depends on grid_size ** 2 which is adapted to images.

Therefore, the number of features set to zero is much higher than expected, all features may be set to zero if grid_size is big. For tabular data or time series, the default value corresponds to the number of features, which is too big and results in fidelity scores of only zeros.

Expected Behavior

Remove the attribute self.subset_size line 91.

Then in lines 201, 208, and 220:
Replace: subset_masks = tf.argsort(subset_masks, axis=-1) > self.subset_size
By: subset_masks = subset_masks > self.subset_percent
!!! See the direction of the > or < depending on the meaning of subset percent.

Version

1.3.1

Environment

No response

Relevant log output

No response

To Reproduce

See Attributions: Time Series and Regression tutorial, in the metric part, for MuFidelity, set subset_percent to 0.5 (perturb half) and grid_size to 48 (treat each time step differently).

It will return only zeros, but the configuration is far from absurd.