PyTorch's torch.nn.functional.interpolate behaving strangely

I'm having issues with PyTorch's tensor-resizing options. In the following code, x is a dataset of 888 64x64 RGB images of pokemon. xs is to be a dictionary of the same dataset at different resolutions.

def load_data():
    pokemon = []
    for png in os.listdir("pokemon"):
        pokemon.append(imageio.imread("pokemon/" + png))
    pokemon = np.array(pokemon)
    s = pokemon.shape
    pokemon = pokemon.reshape((s[0], s[-1], s[1], s[2]))
    pokemon = (pokemon.astype(np.float32) - 127.5)/127.5
    return(pokemon)
x = load_data()

ss = [2, 1, 1/2, 1/4, 1/8, 1/16]
xs = {s : nn.functional.interpolate(torch.from_numpy(x), scale_factor = s, mode='nearest') for s in ss}

As expected, the output shapes include all 888 RGB images with new heights and widths.

print("Data shape:")
for x in xs.values():
    print(" ", x.shape)

> Data shape:
>  torch.Size([888, 3, 128, 128])
>  torch.Size([888, 3, 64, 64])
>  torch.Size([888, 3, 32, 32])
>  torch.Size([888, 3, 16, 16])
>  torch.Size([888, 3, 8, 8])
>  torch.Size([888, 3, 4, 4])

However, the images weren't sensibly resized.

• When s = 1, the pokemon are in their ordinary 64x64 format, as expected:

  

• When s = 2, the pokemon become Pollocks, overlapping in different colors:

  

• When s = 1/2, the pokemon are cut into pieces and rearranged:

  

Is there an alternative PyTorch tool for resizing? I encounter roughly the same problem using AvgPool2d and Upsample instead. I will be resizing non-image tensors, too, so PIL isn't always an option. In any case, using the GPU instead of CPU would be helpful.