Error on certain XAI metrics for tabular data about quantus HOT 2 CLOSED

Hi @sallana22

As can be found in the logic of the infidelity metric, it "patches the input" meaning that it expects input data. I think the line:

data_applicability = {DataType.IMAGE, DataType.TIMESERIES, DataType.TABULAR}

may have caused the issue, as our implementation of the infidelity metric does not support tabular data. This is removed in the following commit: 0ef57b6.

The second error appeared as we assumed that the input also had a channel, the docstring includes "Checks if value is smaller than input size, assumes batch and channel first dimension."

I have made changes (see PR #326) and will merge to main branch when the tests branch. Just pull from the main and then check that it is working for you.

from quantus.

Hi @annahedstroem,

Thanks for the quick resolution! I will exclude infidelity analysis on the tabular dataset.

For faithfulness correlation, your changes work but hit an error on asserts.py, line 295. Adding an elif on line 295 as below solves the problem:

if len(x.shape) == 2:
        if value >= np.prod(x.shape[1:]):
            raise ValueError(
                f"'{value_name}' must be smaller than input size."
                f" [{value} >= {np.prod(x.shape[1:])}]"
            )
elif value >= np.prod(x.shape[2:]):    
        raise ValueError(
            f"'{value_name}' must be smaller than input size."
            f" [{value} >= {np.prod(x.shape[2:])}]"
)

The metric evaluation, however, is computationally slow on the complete test set. Evaluating on 10% of records (approx. 500) takes about 30 mins on a 2GHz quad-core cpu.

from quantus.