integrated grad fails silently when input dtype is not float

[SeppeDeWinter]

Report

The crested.tl._explainer.integrated_grad function will fail silently (producing 0 as output)
when the data type of the input is not a floating point.

The functions: crested.tl._explainer.saliency_map and crested.tl._explainer.smoothgrad at least give some warning.

import keras
import crested
import numpy as np
from crested.utils._seq_utils import one_hot_encode_sequence
from crested.tl._explainer import integrated_grad, saliency_map, smoothgrad
import random

model_path, output_names_biccn = crested.get_model("DeepBICCN")
model_biccn = keras.models.load_model(model_path)

random.seed(123)
oh = one_hot_encode_sequence("".join(random.choices(["A", "C", "G", "T"], k=model_biccn.input[0].shape[1]))).astype(np.uint8)

integrated_grad(oh, model_biccn, 1).sum()
#0

saliency_map(oh, model_biccn, 1).sum()
#WARNING:tensorflow:The dtype of the watched tensor must be floating (e.g. tf.float32), got tf.uint8 
#AttributeError: 'NoneType' object has no attribute 'numpy'

smoothgrad(oh, model_biccn, 1).sum()
#InvalidArgumentError: cannot compute AddV2 as input #1(zero-based) was expected to be a uint8 tensor but is a float tensor [Op:AddV2] name

Version information

No response

[SeppeDeWinter]

I know that one_hot_encode_sequence by default returns a np.float32 array, but it's not clearly documented why this is needed.

see: #139 for potential fix.

For example, a user (like me :p) might encounter this error by initializing an array like this:

one_hot = np.zeros((1000, 500, 4), dtype = int)