enable explicit dtype cast in alibi detect server

[RafalSkolasinski]

This issue is on behaviour currently present in Alibi Detect Server.

Alibi Detect library expects float input to be float32 type in most cases when numerical data is expected.
Currently in alibi-detect-server, e.g. here

seldon-core/components/alibi-detect-server/adserver/ad_model.py

Line 63 in ec06181

X = np.array(inputs)

we cast input to np.array without specifying the dtype. This is bad as the default will be float64

>>> np.array([1.1, 2.2, 3.3]).dtype
dtype('float64')

Some detectors will ignore it but some that relies strongly on Tensorflow will unfortunately fail.

For SC 1.13 that uses AD 0.8.1 this can be worked-around by adding a dummy preprocessing in detector artifact:

preprocess_fn = lambda x: x.astype(np.float32)
dd = DetectorClassHere(x_ref=..., preprocess_fn = preprocess_fn, ...)

but this solution is not ideal.

One idea that comes to my mind would be on the server side to do following casting:

• if dtype is float64 -> cast to float32
• if dtype is int64 -> cast to int32
• else: do not do anything

Environment

SC up to 1.13.0