Unable to reproduce the results on DementiaBank Pitt

[AlexTo]

I've been trying to replicate your results without success so I'd like to reach out to see if there is anything that I'm missing.

Here are the steps I've done:

Get DementiaBank Pitt and put it in data folder

Install OpenSmile 3.0.1 (version 2.3.0 is no longer available) and update the script "generate_compare_feats.sh" accordingly so that "compare" features are generated as csv file in "data/compare" folder.
Update "config.py" to point to the correct "data_dir". I've attached this file in this email.
I also noticed that, at line 10 in "dataset.py", you have
      p = np.random.permutation(552) # n_samples = 108

     I've updated it from 108 to 552 as there are 552 samples
Run "main.py" without changing any code except listed above.

I noticed that, fold splitting is done by samples (config.split_reference == 'samples')

Splitting by subjects is not implemented. It should cause data leak as recordings from the same subject may appear in both train/val split.

Is there any pre-processing steps or configuration that I missed to reproduce the results? Any help would be very much appreciated.

Thank you