Coremltools: Invalid dst shape

[G-mel]

In my model, I'm attempting to mimick a tf.image.resize_nearest_neighbor operation in TensorFlow (which reshapes a layer from (512, 6, 8) to (512, 5, 8), where shape is (C,H,W)) with the builder.add_reshape(..., target_shape=(1, 512, 5, 8)) function in CoreMLTools.

I get the error RuntimeError: Error compiling model: "compiler error: Invalid dst shape8 x 6 x 512 x 1->1->8 x 5 x 512 x 0 x 1 x ". when the line model = MLModel(builder.spec) is executing.

Any idea what this error message means?

[G-mel]

Also, using builder.add_upsample with mode='NN' can't accept float scaling factors for me to reshape from (512, 6, 8) to (512, 5, 8). I need to reshape to (512, 5, 8) in order to concat with another layer with shape (512, 5, 8)

[slin07]

(1) Reshaping. The reshape layers in CoreML requires the total number of elements does not change. This means np.prod(np.array(input_shape)) == np.prod(np.array(output_shape)). In your case, height is reduced from 6 to 5, so reshaping layer cannot handle this case.
(2) CoreML's upsampling layer and downsampling layer currently restricts the scaling factors in be integers.
(3) As a memory-unfriendly workaround, you might try upsampling your blob by 5 then downsampling it by 6.

[G-mel]

@slin07 The upsampling and downsampling work-around did it for me! Does it matter if my downsampling is of type layer_type = 'AVERAGE' or layer_type = 'MAX'?

[slin07]

@G-mel I'm not so sure which is better in this case. Neither option can give you the same output that truly follows nearest neighbor behavior. For example, resizing a length 3 array [2,1,3] to length 4 by nearest neighbor rule is [2,1,1,3], however,

• using the workaround with average pooling gives you [2,1.33,1.67, 3]
• using the workaround with max pooling gives you [2,2,3,3]
  Therefore I would prefer average over max.

[G-mel]

@slin07 Thanks for the great answer, I'll go with Average for now and test out the performance of both.