How to use learn2learn task set with new data set?

[brando90]

Say I have 3 splits of disjoint classes to create anew meta-learning data set (say in 3 folders with each folder having the class and the image files). I can lead each in a normal pytorch data set. How would i create a new l2l task set loader?

[brando90]

The main idea is in this file: https://github.com/learnables/learn2learn/blob/master/learn2learn/vision/benchmarks/mini_imagenet_benchmark.py

but my understanding is to:

1. get a normal pytorch data set e.g. https://github.com/learnables/learn2learn/blob/0b9d3a3d540646307ca5debf8ad9c79ffe975e1c/learn2learn/vision/datasets/mini_imagenet.py#L34
2. wrap it in the l2l objects as in https://github.com/learnables/learn2learn/blob/master/learn2learn/vision/benchmarks/mini_imagenet_benchmark.py likely this step make sure the data transforms are set correctly. Also (likely) return the data sets & task transforms here for next step.
3. then return task objects with each data set, each task transform (note the data transforms were already done to the data set objs) and then wrap it in the benchmark taskloader: https://github.com/learnables/learn2learn/blob/master/learn2learn/vision/benchmarks/__init__.py#L127
4. now it should work.

so three files you need:

• data set obj (normal pytorch): https://github.com/learnables/learn2learn/blob/0b9d3a3d540646307ca5debf8ad9c79ffe975e1c/learn2learn/vision/datasets/mini_imagenet.py#L34
• task transforms & (meta) data sets: https://github.com/learnables/learn2learn/blob/master/learn2learn/vision/benchmarks/mini_imagenet_benchmark.py
• wrap in benchmark loader: https://github.com/learnables/learn2learn/blob/master/learn2learn/vision/benchmarks/__init__.py#L127

[seba-1511]

Hello @brando90,

I recommend you read this tutorial, which goes into the details of task creation and task transforms. Also, feel free to post these support questions in our slack (we try to keep the issue tracker for implementation-related issues).

[brando90]

Hello @brando90,

I recommend you read this tutorial, which goes into the details of task creation and task transforms. Also, feel free to post these support questions in our slack (we try to keep the issue tracker for implementation-related issues).

sounds good! If I may respectfully suggest we open a discussion forum for this repo, the issue with l2l slack is that the responses die after a while :(

Thanks for your patience seba!