Add all the latest models to hubconf #2189
Conversation
|
I think it makes sense to publish those models. @fmassa any ideas? |
hubconf.py
Outdated
| from torchvision.models.detection import fasterrcnn_resnet50_fpn, \ | ||
| maskrcnn_resnet50_fpn, keypointrcnn_resnet50_fpn |
There was a problem hiding this comment.
I think detection models won't work as expected with torchhub.
This is because they require compiled C++ extensions, which do not get compiled in Hub, and would instead try to pick the ops from torchvision.
There was a problem hiding this comment.
Will remove detection models. Segmentation models are okay?
which do not get compiled in Hub, and would instead try to pick the ops from torchvision.
Ah why they worked on my local env.
There was a problem hiding this comment.
because you already have torchvision installed on your system. But if the version of torchvision that you had installed in your system was quite old (say 0.3.0), then the model wouldn't work because the interface for the custom ops has since changed.
New models in torchvison are not in hub yet.