learnables
Checkpoints for all data sets
Hi,
I was wondering if checkpoints could be made available for some of the meta-learning datasets.
I am in particular interested in mini-imagenet checkpoints trained with MAML.
Though, honestly, the more checkpoints with more meta-learning algorithms the better.
Hello @brando90,
This is not our immediate priority as it's unclear which model is worth check-pointing or not. (Different papers use slightly different architecture variations). Hopefully training your own model and saving it (e.g. on Colab) should be as easy as copy-pasting the script and adding the check-pointing code.
I'll leave the issue open in case someone would like to tackle it.
Hello @brando90,
This is not our immediate priority as it's unclear which model is worth check-pointing or not. (Different papers use slightly different architecture variations). Hopefully training your own model and saving it (e.g. on Colab) should be as easy as copy-pasting the script and adding the check-pointing code.
I'll leave the issue open in case someone would like to tackle it.
Hi @seba-1511
I know this is not a priority. But do you have at least the original 5CNN on mini-imagenet that Chelase Finn did but with l2l? I am guessing you must have done some unit test/sanity check to verify your library. So I'm crossing my fingers you used that setting. Did you?
Yes, I carefully tested this implementation. See #280 where I give a configuration which yields decent results.
Closing since we won't provide weights for all possible configs (see get_pretrained_backbones for supervised pretraining ones).