is that possible to replace your TrilinearInterpolation with torch.nn.functional.grid_sample?

[Dongshengjiang]

is that possible to replace your TrilinearInterpolation with torch.nn.functional.grid_sample? If ok, it would be more convenient.

[HuiZeng]

Hi, this is an interesting idea.
But it seems not that easy since their interpolation details are different.

[tuxa]

Hi! grid_sample actually uses trilinear interpolation if a 5d tensor is passed when mode is set to bilinear.

So, instead of using

_, result = trilinear_(LUT, img)

one can use:

# scale im between -1 and 1 since its used as grid input in grid_sample
img = (img - .5) * 2.

# grid_sample expects NxDxHxWx3 (1x1xHxWx3)
img = img.permute(0, 2, 3, 1)[:, None]

# add batch dim to LUT
LUT = LUT[None]

# grid sample
result = F.grid_sample(LUT, img, mode='bilinear', padding_mode='border', align_corners=True)

# drop added dimensions and permute back
result = result[:, :, 0].permute(0, 2, 3, 1)

[coallar]

changing to:ndarr = result.squeeze().mul_(255).add_(0.5).clamp_(0, 255).permute(1, 2, 0).to('cpu', torch.uint8).numpy() to ndarr = result.squeeze().mul_(255).add_(0.5).clamp_(0, 255).to('cpu', torch.uint8).numpy()
is necessary

[ZedFu]

Has anyone ever used grid_sample instead of Trilinear Interpolation? Why is the result of using grid_sample worse?

[11923303233]

scale im between -1 and 1 since its used as grid input in grid_sample

img = (img - .5) * 2.

grid_sample expects NxDxHxWx3 (1x1xHxWx3)

img = img.permute(0, 2, 3, 1)[:, None]

add batch dim to LUT

LUT = LUT[None]

grid sample

result = F.grid_sample(LUT, img, mode='bilinear', padding_mode='border', align_corners=True)

drop added dimensions and permute back

result = result[:, :, 0].permute(0, 2, 3, 1)

How to set batch_size > 1? LUT = LUT[None] --> LUT = LUT[None].repeat(bs, 1,1,1,1) is ok?

[fedral]

Hi! grid_sample actually uses trilinear interpolation if a 5d tensor is passed when mode is set to bilinear.

So, instead of using

_, result = trilinear_(LUT, img)

one can use:

# scale im between -1 and 1 since its used as grid input in grid_sample
img = (img - .5) * 2.

# grid_sample expects NxDxHxWx3 (1x1xHxWx3)
img = img.permute(0, 2, 3, 1)[:, None]

# add batch dim to LUT
LUT = LUT[None]

# grid sample
result = F.grid_sample(LUT, img, mode='bilinear', padding_mode='border', align_corners=True)

# drop added dimensions and permute back
result = result[:, :, 0].permute(0, 2, 3, 1)

good trick!