num_samples Error

[wang-zm18]

When I run "test/test_metric.py ", it occurs this error:

ValueError: num_samples should be a positive integer value, but got num_samples=0

Even I specify num_samples by using "--num_samples 5"

[PACerv]

Thank you for letting me know about this!
This error is raised by a pytorch dataloader when the dataset has len(dataset)==0. Could you explain where exactly you encounter this error? It could be that the path_dataset is not set properly. If so, I might add an error message to catch this case.

[wang-zm18]

Thanks for your reply, the code I use is python test/test_metrics.py --path_results ./results --path_classifier pretrained_models/humanact12/checkpoint_5000.pth.tar --variable_length_testing --metrics. And the path_dataset is set "./dataset/HumanAct12/HumanAct12". the volume of my dataset is below:

And the "./dataset/HumanAct12/HumanAct12" contains all .npy files of HumanAct12.

[PACerv]

I've noticed that line 805 should be commented out.

Regarding the ValueError, could you specify which line raises this error? I am not sure why the dataset would be empty since your path_dataset seems correct to me.

A maybe unrelated issue is that you specify --path_classifier pretrained_models/humanact12/checkpoint_5000.pth.tar. I guess that these are the pretrained weights of the motion generation model from ACTOR?
--path_classifier should actually point to a folder containing the weights of classifiers used for evaluation and not the motion generation models. The motion generation models should be found in --path_results. The structure of a results folder should be as follows:

The name of the file containing the model weights should be specified under --epoch and is by default latest.pt. (In hindsight I realize this might be confusing).
The code does not support loading the pretrained ACTOR models though, but it should be possible to adapt the code for it.