Correcting incorrect types

[zhiqwang]

Due to torchscript support torch.device now, I updated the Device to torch.device

[zhiqwang]

Hi, It's unexpected that the modification of type Device to torch.device would affect a lot in the unittest, so I closed it );

[zhiqwang]

Learned from pytorch/vision#3059, I used string for class name annotation to avoid type annotation error, but there is also a segmentation fault ;(

[fmassa]

Hi,

Thanks for the PR!

I checked the CI logs from DETR, and in here pip is picking the nightly from Nov 29, while pytorch/vision#3059 is taking (via conda) the nightly from Nov 30, which might explain the segfault.

This means that maybe if we try re-running the tests in a couple of days the segfault should disappear.