Feedback from Anirban

[rayanirban]

• There could be optional ways of exploring the data folders (eg. using seedir library in Python or something similar).
• The exercise to find the total number of pairs in the data folders and their extension (in case you want to keep it) could be a good learning exercise for the students but I would recommend some starting instructions.
• The assert of hash to check correctness may be removed because of unpredictable output based on the environment.
• In the NucleiDataset class, I would recommend some more explanation of the three different types of transforms that re being used.
• In the show_random_dataset_image function in the solutions.ipynb I would recommend that the solution incorporated an additional default parameter idx=None that might be used to visualize a particular image.
• After checkpoint 1, you might want to give students some options to play around with the GaussianBlur augmentation by allowing them to change the RandomCrop and the GaussianBlur parameters and plot the different images together.
• I recommend doing the assert cuda() at an initial stage of the notebook.