Mismatching between the checkpoints and the config file
Closed this issue · 11 comments
Hi, thanks for the good work.
I download the checkpoints from the google drive link, and use the config file you mentioned. However, when I use the image_demo.py for testing, there are some mismatching problems while loading the weights. Do you happen to know the reasons and the ways to resolve it? Thanks!
Can you confirm that the output is fine when you run the demo ? I mean does your demo image get some detections after running the script ?
It seems the backbone in the checkpoint is form HRNet because I saw many keywords from HRNet (I guess maybe you upload the wrong checkpoints?) . Yes, the detections will be messed up.
Yes, the detections will be messed up. Will be or is it messed up ?
Paste the whole error log along with the command you use to run the demo.
Thanks. The detection results are messed up because the weights are not loaded. The error logs are two folds:
(1) Unexpected keys like backbone.bn1 (it seems from HRNet backbone) as in the first screen shot, and
(2) Missing keys are the keys in the Swin-transformer like patch_embed.
I run the code by using:
The checkpoints.pth is renamed from the downloaded one in Google drive.
Thanks!
Sorry for bothering you. Could you share the new weights with me? Thanks so much!
The issue is, I currently do not have access (I am away from the server) to the model. As I am away from office and cannot access it remotely. There is definitely a mess up between the original model and the uploaded one. It will take some time.
Thanks for the explanation. Looking forward to the correct model weights!
Hopefully will do it soon.
@yzhouas Apologies for late response, it took so long because of the access issues I had to face. I have updated the links and also provided the config file, it should work now, incase you require.
Thank you @hasanirtiza . The new model weight and config file work now.
Great, I will close the issue then.