search

tristandeleu / pytorch-meta

A collection of extensions and data-loaders for few-shot learning & meta-learning in PyTorch
https://tristandeleu.github.io/pytorch-meta/
MIT License
1.98k stars 256 forks source link

Clarification of some meta-learning terms (for documentation improvement) #75

Open renesax14 opened 4 years ago

renesax14 commented 4 years ago

This might help the documentation later so I want to ask (before I attempt to do any PRs).

The meta-set is usually a collection of data-sets (as outlined in the paper). With this in mind is:

  1. is a task equivalent to one data-set sampled from a meta-set? (to me 1 task is the same as 1 data set)
  2. How many tasks do we sample during 1 episode during meta-training? (for me during one episode we sample meta_batch number of tasks
  3. Is 1 episode the same as 1 batch_idx e.g
    with tqdm(dataloader, total=args.num_batches) as pbar:
        for batch_idx, batch in enumerate(pbar):

same as

    with tqdm(dataloader, total=args.num_batches) as pbar:
        for episode, batch in enumerate(pbar):

(to me episode === batch_idx)

  1. I think I saw that it was used the term somewhere that episode is the same as 1 data set but I always thought 1 episode was sample a set of tasks (data sets)

hope referencing the numbers makes it more straight forward to address my questions.