dl Flashcards
dl: Only nvidia GPUs support
Cuda
jupyter: To see which arguments a function requires
put cursor in brackets and press shift tab twice
dl: the standard folder structure for image data is
one folder called train and one called valid
each one has one folder per class, and that folder is full of images of that class.
dl: Learning rate annealing is when
you decrease learning rate as you get to later iterations.
Note: A good type is cosine annealing. Keep learning rate high, then drop fast, then keep learning rate flat
dl: A learning rate schedule is
Using an annealing learning rate that jumps back up to the highest learning rate and starts again a few times to make sure its not in a non optimal valley. Works better than ensembles (where they train 10 models hoping one will work better)
dl: The best way to deal with unbalanced datasets is to
make copies of the rare classes