Problem with SpaceToDepth layer

ARM-software / ComputeLibrary

The Compute Library is a set of computer vision and machine learning functions optimised for both Arm CPUs and GPUs using SIMD technologies.

MIT License

2.82k stars 774 forks source link

Closed allnes closed 1 year ago

allnes commented 1 year ago

Hi! When I used NESpaceToDepth layer, I got wrong values in random indexes.

That is configuration were with current bug:

arm_compute::DataType::F32 with block_size=2
arm_compute::DataType::F32 with block_size=3 and arm_compute::DataLayout::NCHW
arm_compute::DataType::S8 with block_size=2
arm_compute::DataType::S8 with block_size=3 and arm_compute::DataLayout::NCHW

GGGGxxxxxxxxr commented 1 year ago

Hi @allnes ,

I have not used this operator but I have planned to use it in the future.

Could you please explain the wrong value in more details?

I suppose I would look at the backend implementation of this and try to find out the bug

GGGGxxxxxxxxr commented 1 year ago

I have tried DepthtoSpace Operator for PixelShuffle purpose, I cannot get the correct result either. I have implemented based on NHWC layout.

GGGGxxxxxxxxr commented 1 year ago

I have figured out the issue.

For example, as you configured it in NCHW layout, I would explain in NCHW layout.

If I want to convert Tensor from (2,2,12) to (4,4,3) (blocksize= 2)

The first channel of the output would take data from INPUT Channel 1, 4, 7, 10.

But as you might expect it to take the data from INPUT CHANNEL 1,2,3,4.

allnes commented 1 year ago

thank you for answer!