Evaluation error args = checkpoint['args']

[zazgf]

@brade31919 ,
I debugged the code , and encountered the error in main.py line 199 "args = checkpoint['args']", error notice is KeyError: 'args' I found checkpoint have function value 'arch'.
Can you help me sovle this problem?

[zazgf]

Hi, @brade31919 ,
I think just use arch instead of args, but I'm not sure

args.arch = checkpoint['arch']
print(args)

[zazgf]

@brade31919 ,
Yes, when I edit the code line 199 as above and ran out the desired result.

python main.py --evaluate /home/username/radar_depth/pretrained/resnet18_multistage.pth.tar --data nuscenes --arch resnet18_multistage_uncertainty_fixs --modality rgbd --sparsifier radar --decoder upproj

• set num_workers=0 ,as My computer configuration is relatively low.

Thanks again for @brade31919's help.

[Kirstihly]

With the provided checkpoints, the keys are

dict_keys(['arch', 'model_state_dict', 'optimizer_state_dict'])

But if trained ourselves, the checkpoint will include

dict_keys(['args', 'epoch', 'arch', 'model_state_dict', 'best_result', 'optimizer_state_dict'])

[Abdulaaty]

Have you managed to get similar results to the publication when only evaluating the pretrained models? I got very bad results here (RMSE is in magnitude of 20 )

@brade31919 , Yes, when I edit the code line 199 as above and ran out the desired result.

python main.py --evaluate /home/username/radar_depth/pretrained/resnet18_multistage.pth.tar --data nuscenes --arch resnet18_multistage_uncertainty_fixs --modality rgbd --sparsifier radar --decoder upproj

• set num_workers=0 ,as My computer configuration is relatively low.

Thanks again for @brade31919's help.

[zazgf]

@Abdulaaty Sorry for late reply. This is result what I got.

And I did not commented code else.

[Abdulaaty]

@zazgf Thanks for your reply. I mean did you evaluate the pretrained models on the dataset that they provide in the repo (as in to replicate the results) and got the same results as they state in the paper?

[zazgf]

@zazgf Thanks for your reply. I mean did you evaluate the pretrained models on the dataset that they provide in the repo (as in to replicate the results) and got the same results as they state in the paper?

I did not evaluate the models.