• Make a new figure for the StarGAN (to be exactly like what we're doing)
• Have a pre-trained stable network ready (not sure whether to make it explicit in the notebook or just give it after a checkpoint)
• Have a "data visualization" at the end of the notebook, that shows that for this task it is possible to figure out what the classes are by just looking at enough images, sorting them in interesting ways, etc.

• In cell 2 (create a balanced sampler): I think that you have the solution copied in the exercise (or do you intend for there to be nothing to change for this cell. The line "Below define a sampler" madde me think that there was something I had to do.
• Could be useful to include an illustration of the VGG network (https://paperswithcode.com/method/vgg), they've seen it before in florians talk
• in cell 12 (task 1.2): you define a variable as "ground_truths" then call "ground_truth" on the last line, producing an error. Also "prediction" and "predictions"
• also in cell 12: the solution is still present in exercises

ran out of GPU memory, will try to finish the notebook later tonight

• Standardize whether to use “Task” or “Exercise”
• Remove training of the VGG classifier. Instead load the model and do evaluation (and use the extra space to give more details). Alternatively, give them a link to a script to train their own model if they choose to.
• For Exercise 1.4, no good reason to have two methods here, just do one
• Make questions tasks: have them discuss in Element/Matrix
• Exercise 2: typo in attributions
• Goes from 2.1 to 2.3
• Exercise 2.1: remove hints! (too obvious) “read the docs”, hint to number of lines (2)
• Part 3: add more links to everything in the explanation, e.g. the lecture slides (CycleGAN, counterfactual explanation)
• remove exercise 3.1: It is more about data wrangling than about actual training. We could make this available to them.
• Exercise 3.2: Options good! Add the numbers from the top to the choices in the code
• Checkpoint 3: Add learning objectives
• Memory errors is a problem. Perhaps we should just have a smaller dataset, or smaller models.
• Remove exercise 4.2
• Move the TODOs inside the blue boxes.
• Part 4 is all called 5 for some reason

• Combines with failure modes by sharing the dataset

Assumes that the VGG that we need has been trained in the Failure modes exercise.