Open Radar

 
Area:
Metal

Summary:
MPSCNNConvolutionTranspose encodeCommand outputs a image with the same size of the input image.

MPSCNNConvolutionTranpose encode function has following signature:
- (MPSImage *)encodeToCommandBuffer:(id )commandBuffer 
                        sourceImage:(MPSImage *)sourceImage 
                   convolutionState:(MPSCNNConvolutionState *)convolutionState;

I assume that we feed source image and this function will return a result image after running ConvolutionTranspose?

Steps to Reproduce:
API seems to be not tested? Or is there any misunderstanding of the API?

Expected Results:
For an input image of dimension (N, C, H, W) and a kernel size(3, 3), we should have output image of size (N, C, (H - 1) + 3, (W - 1) + 3)

Actual Results:
I got output image of size (N, C, H, W) (same as input image)

Version/Build:
11.0

Configuration: