Resnet classifier sample returns incorrect results on MPS

[malfet]

🐛 Describe the bug

Consider the following example:

from torchvision.io import read_image
from torchvision.models import resnet50, ResNet50_Weights
img = read_image("dog2.jpg").to("mps")
weights = ResNet50_Weights.DEFAULT
model = resnet50(weights=weights).to("mps")
model.eval()
preprocess = weights.transforms()
batch = preprocess(img).unsqueeze(0)
prediction = model(batch).squeeze(0).softmax(0)
class_id = prediction.argmax().item()
score = prediction[class_id].item()

When run on MPS following image is classified as necklace:

Versions

Nightly/1.13.0

cc @ezyang @gchanan @zou3519 @kulinseth @albanD @DenisVieriu97 @razarmehr @abhudev

[malfet]

Upon digging further, it looks like the problem is in torch.nn.functional.interpolate, i.e.:

import torch
import torchvision
import matplotlib.pyplot as plt
img=torchvision.io.read_image("assets/dog2.jpg").to("mps")
y=torch.nn.functional.interpolate(img.unsqueeze(0).to(torch.float32), size=(232, 232), mode='bilinear').squeeze(0).to(torch.uint8)
plt.imshow(y.cpu().permute(1, 2, 0))
plt.show()

Renders the following:

[malfet]

Hmm, looks like the problem is simply with copying contiguous tensor to MPS device:

>>> img=torchvision.io.read_image("assets/dog2.jpg")
>>> img.is_contiguous(), img.shape, img.stride()
(False, torch.Size([3, 500, 500]), (1, 1500, 3))

[malfet]

OK, so the problem is very easy to simulate as follows:

% python -c "import torch; x=torch.arange(27).reshape(3, 3, 3).permute(2, 0, 1); print(x, '\n', x.to('mps'))"
tensor([[[ 0,  3,  6],
         [ 9, 12, 15],
         [18, 21, 24]],

        [[ 1,  4,  7],
         [10, 13, 16],
         [19, 22, 25]],

        [[ 2,  5,  8],
         [11, 14, 17],
         [20, 23, 26]]]) 
 tensor([[[ 0,  9, 18],
         [ 1, 10, 19],
         [ 2, 11, 20]],

        [[ 3, 12, 21],
         [ 4, 13, 22],
         [ 5, 14, 23]],

        [[ 6, 15, 24],
         [ 7, 16, 25],
         [ 8, 17, 26]]], device='mps:0')