🚀 Feature

Interpolation output size is incorrectly computed with floor vs opencv/skimage/scipy are using round:

# pytorch
output_size = floor(input_size * scale)

vs

# opencv
output_size = round(input_size * scale)

Details

Currently, interpolation output size if using float scale factor and recompute_scale_factor=False is computed as

In both cases, this can lead to -1 pixel output size comparing to opencv, scikit-image or scipy. For example, let's rescale an image of size (32, 32) with different scale factors: [0.1234, 0.9999, 1.8]. This gives the following output sizes per library:

- check scale_factor=0.1234
Error: different shapes: opencv (4, 4) vs sk-img (4, 4) vs scipy (4, 4) vs pth (3, 3)

- check scale_factor=0.9999
Error: different shapes: opencv (32, 32) vs sk-img (32, 32) vs scipy (32, 32) vs pth (31, 31)

- check scale_factor=1.8
Error: different shapes: opencv (58, 58) vs sk-img (58, 58) vs scipy (58, 58) vs pth (57, 57)

...

Here is a gist with details: https://gist.github.com/vfdev-5/a5bd5b1477b1c82a87a0f9e25c727664

Pitch

It would be helpful to have compatible output size using pytorch interpolate op.

Alternatives

Use output size argument directly instead of scales.

Additional context

Here is equivalent code for

• opencv:
  • https://github.com/opencv/opencv/blob/68d15fc62edad980f1ffa15ee478438335f39cc3/modules/imgproc/src/resize.cpp#L4049-L4058
  • https://github.com/opencv/opencv/blob/68d15fc62edad980f1ffa15ee478438335f39cc3/modules/core/include/opencv2/core/saturate.hpp#L140
• scikit-image: https://github.com/scikit-image/scikit-image/blob/e128e0743bb2cded104654352a13ed5b4e53efb8/skimage/transform/_warps.py#L286
• scipy: https://github.com/scipy/scipy/blob/44e31b1eb202ef91caad059952ba268d2610a60b/scipy/ndimage/interpolation.py#L768-L769

cc @fmassa