improvements to dataloader.py

[wkvong]

some ways to improve the dataloader code:

• in the __init__ function, I would add code to create a CSV file that contains the relevant properties that you need by reading through the image directories for properties such as (image_filename, shape_category, texture_category, shape_instance, texture_instance) and save that to file. Then, you can just read from this CSV directly in the initialisation step, and then the __getitem__ function can directly index values for each item. Also, one thing I would do here is filter out any trials where the shape and category labels are the same before saving the CSV file since they aren't cue conflict images
• also, rather than using os.listdir and filtering out the .DS_Store files, I often prefer to use the functionality from the glob library, e.g. glob.glob('stimuli-shape/style-transfer/*/*.png') which allows me to filter by filetype more easily (and also returns the full directory + filename)

[wkvong]

one more thing, I'd also return the image first (rather than last) from the output of __getitem__ since that is the part that gets passed into the model and its the thing you actually want the dataloader for

[alexatartaglini]

I incorporated all of these suggestions! I'm currently saving the relevant properties for the images in a JSON file (which is on the repo now)