ENH: generic_filter supports axes

[Microdeep-ZL]

Is your feature request related to a problem? Please describe.

most filters accept the parameter axes, I'm wondering if it's possible to make generic_filter support axes as well.

Describe the solution you'd like.

generic_filter(data,getKCC,output=np.float32,footprint=np.ones((3,3,3)),axes=(0,1,2))

Describe alternatives you've considered.

I have a 4D ndarray, which a time series of 3D images. I want to compute Kendall's W (KCC) for every 3x3x3 cluster. Now my workaround is as follows

def getKCC(inputs,length,data):
    inputs=[slice(i,i+length) for i in inputs.astype(np.uint32)]
    indices=np.r_[*inputs]
    data=data.flat[indices].reshape(-1,length)

    # compute KCC
    var=data.mean(axis=0).var()
    return 12*var/(length-1)/(length+1)

# data.shape (40,40,40,3000)
length=data.shape[-1]
indices=np.arange(np.array(data.shape[:3]).prod(),dtype=np.uint32)*length
indices=indices.reshape(data.shape[:3])
generic_filter(indices,getKCC,output=np.float32,footprint=np.ones((3,3,3)),extra_arguments=(length,data))

Additional context (e.g. screenshots, GIFs)

No response

[Microdeep-ZL]

I realized that I've misunderstood how axes works in other filters. Initially I thought axes=(0,1,2) would make the output have 3 dimensions rather than 4. In fact, even though generic_filter supports axes, this function call generic_filter(data,getKCC,output=np.float32,footprint=np.ones((3,3,3)),axes=(0,1,2)) will still result in an array with the same shape as data

Sorry for opening an immature issue. But anyways, I think it would be good to implement axes in generic_filter