🐛 Describe the bug

...
device = "cuda" if torch.cuda.is_available() else "cpu"
raft_model = raft_small(pretrained=True, progress=False).to(device)
raft_model = raft_model.eval()
transform = transforms.ToTensor()
with torch.no_grad():
list_of_flows = raft_model(old_batch.to(device), new_batch.to(device))
...

Versions

Hi there,

I am testing the orchvision.models.optical_flow module raft_small, the code is running ok for image size (480, 752), (800,848)..
However, when I test it on Image size: Height 480 , Width 854. The code throw

ValueError: The feature encoder should downsample H and W by 8

I debug the code on https://github.com/pytorch/vision/blob/d3beb52a00e16c71e821e192bcc592d614a490c0/torchvision/models/optical_flow/raft.py#L494

fmaps = self.feature_encoder(torch.cat([image1, image2], dim=0))
fmap1, fmap2 = torch.chunk(fmaps, chunks=2, dim=0)
if fmap1.shape[-2:] != (h // 8, w // 8):
     raise ValueError("The feature encoder should downsample H and W by 8")

Image size: Height 480 , Width 854
where fmap1.shape[-2:] is torch.Size([60, 107]), h // 8 = 60, but w // 8 = 106 which triggered the ValueError.

I think this issue is related to output dimension of self.feature_encoder. Looking for help, thx~