Closed LukasMut closed 1 year ago
I had a bit of trouble getting this to work, but I'm not sure how connected it is to this specific commit or not. I cloned the master branch as this had already been merged and tried to load some of the new
dinov2-vit*
models. I think because I had to step away from thingsvision for a bit, coming back gives me the experience of a more inexperienced user and the issues one might face (useful for testing).First of all, noticed DINOV2 wasn't on the ReadMe and wasn't sure what the source was, so copying example on ReadMe was default
torchvision
and had to look inextractors.py
to see it was the hub. Then I realised I didn't know the keyword for that as it's not specified in the ReadMe (and perhaps should be, comment in the example for what to use to reference different model sources).As
torchvision
is used in the ReadMe example, I left that in accidentally, had errors of missing libraries that I had to install manually (keras_applications
andvalidators
) then the error sayingdinov2-vit*
models were not intorchvision
and another one sayinghub
didn't exist as a source. This is confusing in the installation instructions. I had a peek inhelpers.py
and saw the options I could give there and realised it wasssl
and it was a dumb mistake. So, I think we don't need to referencevssl
ortorch.hub
in those instructions and leave the only option to a new user that the keyword isssl
. Come to think of it, I don't think there is ever a case where a model would be possible to select from different sources so I don't know why the user should specify it. It seems only to introduce a chance for things to go wrong. The mapping in a dictionary could be implemented internally within the code.The other models seem to be working, but I am getting errors saying it can't find any of the dino-v2 models in the
SSLExtractor
even as I'm copying the keys fromMODELS
dictionary directly and it seems it should work.I am leaving for the weekend shortly and can't dig deeper, but just wanted to provide this feedback before I left in case it was useful.
@Alxmrphi, I am not sure I follow. Could you be more specific/concise about the issue?
I had a bit of trouble getting this to work, but I'm not sure how connected it is to this specific commit or not. I cloned the master branch as this had already been merged and tried to load some of the new
dinov2-vit*
models. I think because I had to step away from thingsvision for a bit, coming back gives me the experience of a more inexperienced user and the issues one might face (useful for testing).First of all, noticed DINOV2 wasn't on the ReadMe and wasn't sure what the source was, so copying example on ReadMe was default
torchvision
and had to look inextractors.py
to see it was the hub. Then I realised I didn't know the keyword for that as it's not specified in the ReadMe (and perhaps should be, comment in the example for what to use to reference different model sources).As
torchvision
is used in the ReadMe example, I left that in accidentally, had errors of missing libraries that I had to install manually (keras_applications
andvalidators
) then the error sayingdinov2-vit*
models were not intorchvision
and another one sayinghub
didn't exist as a source. This is confusing in the installation instructions. I had a peek inhelpers.py
and saw the options I could give there and realised it wasssl
and it was a dumb mistake. So, I think we don't need to referencevssl
ortorch.hub
in those instructions and leave the only option to a new user that the keyword isssl
. Come to think of it, I don't think there is ever a case where a model would be possible to select from different sources so I don't know why the user should specify it. It seems only to introduce a chance for things to go wrong. The mapping in a dictionary could be implemented internally within the code.The other models seem to be working, but I am getting errors saying it can't find any of the dino-v2 models in the
SSLExtractor
even as I'm copying the keys fromMODELS
dictionary directly and it seems it should work.I am leaving for the weekend shortly and can't dig deeper, but just wanted to provide this feedback before I left in case it was useful.