Chapter 7 Answers

1. What is the difference between ImageNet and Imagenette? When is it better to experiment on one versus the other?

   Imagenet original is a dataset with ~1.3M images and 1000 classes. Imagenette is a much smaller subset of 10 distinct classes

2. What is normalization?

   Normalization is the process of making the data input have 0 mean and std 1

3. Why didn’t we have to care about normalization when using a pretrained model?

   If no Normalization.from_stats() is passed to batch_tfms, FA calculates the stats from a single batch of your data

4. What is progressive resizing?

   Training on images that gradually increase in size as training progresses. For example, first fit_one_cycle on size 128, then fine_tune on 224

5. Implement progressive resizing in your own project. Did it help?

   *** Did not try

6. What is test time augmentation? How do you use it in fastai?

   TTA is when you data augment test images, and average the predictions In FA, do learn.tta(n=4)

7. Is using TTA at inference slower or faster than regular inference? Why?

   Slower, because we need each additional image needs an inference as well

8. What is Mixup? How do you use it in fastai?

   Mixup is when you mixup images by adding weighted versions together. This means the image pixels are added together. Weighted labels are added together as well Implemented in fa2 by adding a callback into the learner You will also want to use the LabelSmoothingCrossEntropy loss func Example: cnn_learner(dls, …, cbs = MixUp(), loss_func=LabelSmoothingCrossEntropy())

9. Why does Mixup prevent the model from being too confident?

   Because it allows the model to optimize predictions to values between 0, and 1. Instead of always forcing to 0/1

10. Why does a training with Mixup for 5 epochs end up worse than a training without Mixup?

    It is harder to see what’s in a MixUp image. Also, now it needs to predict continuous numbers instead of a 1/0

11. What is the idea behind label smoothing?

    If there are mislabels in your dataset, trying to optimize to 1/0 predictions can be harmful. Encourage model to be less confident by having labels between 1/0, and then adding a little “epsilon” error to every other prediction Example for one hot-encoded -> [0, 1, 0, 0] -> [0.1, 0.7, 0.1, 0.1]

12. What problems in your data can label smoothing help with?

    Mislabeled data or harder to label data. I think label smoothing can give a higher fidelity to data if you start adding in prediction + confidence level For example [0, 1, 0, 0] -> [0, 0.9, 0, 0] -> Do you care if all sum to 1? Or just that they are constrainted between 0 and 1 (a range)

13. When using label smoothing with 5 categories, what is the target associated with the index 1?

    1 - epsilon + epsilon/N -> where N = 5

14. What is the first step to take when you want to prototype quick experiments on a new dataset.

    If the dataset is large, find a smaller subset that is representative of the whole. This will let you try experiments much faster. Key is finding smallest dataset that is representative and able to discriminate between techniques For example, Imagenet (1000 classes) -> Imagenette (10 classes)