Dataset metadata (i.e. 'trials_properties.json' and 'dataset_info.json') is loaded by each visual search model separately. It would be wiser to load them beforehand and passing them as parameters.
This would have the benefit of:
- Allowing to run the model(s) on a specified subset of the dataset (through command line arguments)
- Adding the capability to resume the computation of human scanpath prediction on a given subject, instead of starting over