CODEWITHSUNDEEP
python-3.xnumpytensorflow
ellamunde
ellamunde

Reputation: 23

Is there any tensorflow version of numpy.view_as_windows?

I am working with python3 and tensorflow to generate image patches by using numpy view_as_windows but because of numpy can't run on GPU, is there any way to do it with tensorflow?

ex: view_as_windows(array2d, window_shape, stride)

Thanks

Upvotes: 1

Views: 161

Answers (1)

GPhilo
GPhilo

Reputation: 19123

Note: This answer does not answer the OP's exact question, but addresses the actual need of the OP as clarified in the comments (i.e., generate image patches, quickly). I just thought this would fit better here than in a badly-formatted comment.

If all you need to do is generating image patches, Tensorflow (and generally GPU acceleration) is not the right tool for this, because the actual computation is trivial (extract a sub-area of an image) and the bottleneck would be memory transfer between GPU and CPU.

My suggestion is, then, to write CPU-only code that uses view_as_windows and parallelize it via multiprocessing to split the workload on all your CPU cores.

Should you need to feed those patches to a Tensorflow graph afterwards, the way to go would be to first generate the patches on the CPU (with whatever input pipeline you like), batch them and then feed them to the GPU for the graph computation.

Upvotes: 2

Related Questions