Modelling improvements

[yanirs]

• Go through notebooks and apply documented thoughts
• Revisit learning rate finder experiments: fit_one_cycle() or fit()
• Incorporate object detection (e.g., auto-crop and train without too much of the background)
• Incorporate text descriptions (as input/output)
• Revisit and collect notes from more advanced fast.ai notebooks
• Look into https://github.com/libffcv/ffcv
• Look into a smaller model without losing accuracy: https://wandb.ai/fastai/fine_tune_timm/reports/Finding-the-New-ResNet18--VmlldzoyMDI0MjU3