[New feature] col2im/ image patch

[roya0045]

Greetings,

I am writing this to officially document my proposition to add something that would separate a dataset into subparts like is it possible to do in Tensorflow and chainer.

The chainer version is the following: https://docs.chainer.org/en/stable/reference/generated/chainer.functions.im2col.html

With Tensorflow it's not that straightforward it can be done with proper reshaping:
https://www.tensorflow.org/versions/master/api_docs/python/tf/extract_image_patches

Here is the CNTK equivalent that I made, though it is different compared to the original chainer function as I made it more viable for broadcasting. I was wondering if it could be possible to integrate this officially as an utility function for all languages.

    def cnwindow(image,window,pad=0,stride=1):
        if pad>0:
            image=C.pad(image, pattern=[(0,0),(0,0),(pad,pad),(pad,pad)],  constant_value=0)
        image_shape=image.shape
        output_shape=(*image_shape[:-2],*window,((image_shape[-2]-window[-2])+1)//stride,((image_shape[-1]-window[-1])+1)//stride)
        holder=None
        for R in range(window[0]):
            j_lim = R + output_shape[-2]
            for H in range(window[1]):
                temp_data=C.slice(image,[-2,-1], [R*stride,H*stride], [j_lim*stride,(H +  output_shape[-1])*stride])
                if holder is None:
                    holder=temp_data
                else:
                    holder=C.splice(holder,temp_data,axis=1)
        output=(C.reshape(C.transpose(C.reshape(holder, shape=output_shape),(0,4,5,3,1,2)), (output_shape[0],*output_shape[5:3:-1],1,*output_shape[3:0:-1])))
        print(output.shape,"CNWINDOWSHAPE")#(0,5,4,3,2,1)
        return(output)

[ebarsoumMS]

Can you submit is PR?

[roya0045]

Sure, I'll just correct the name of the variables in the code, in what folder do you want me to put the code? (unless you are referring to @spandantiwari )