Anki
Overview
This chapter contains some of the advanced techniques and tricks that can be used to train SOTA image classification models.
ImageNet: how many images?
1.3M
ImageNet: how many categories?
1000
ImageNet: what size are images?
differing but ~500px across
MNIST: how many images?
50k
MNIST: how many categories?
10
MNIST: what size are images?
28x28
CIFAR10: how many images?
60k
CIFAR_: how many categories? (blank gives the game away!)
10
CIFAR10: what size are images?
32x32
How does Imagenette differ from ImageNet?
Only 10/1000 classes, so about 1% (10k) of the number of images
If experiment iteration is slow, what can you do to the dataset?
Create a simplified version to use while iterating
Training
When using a pre-trained model, what should you transfer as well as the weights?
Normalisation statistics
FastAI: how to transfer normalisation statistics when using a pre-trained model
FastAI automatically adds a
Normalise transform with the correct statisticsWhat is the term for iteratively increasing the size of each datapoint during training?
Progressive resizing
What dataset modification can be used to speed up the training of CV models?
Progressive resizing
Why might progressive resizing not help for transfer learning?
If the pre-training task had a similar image size to the target task, keeping the size the same may be beneficial