search

lmzintgraf / cavia

Code for "Fast Context Adaptation via Meta-Learning"
MIT License
139 stars 40 forks source link

split miniimagenet to train val and test #1

Closed vainaixr closed 5 years ago

vainaixr commented 5 years ago

hello, i ran the code for classification on google colab, downloading files on google drive.

the images folder containing 60,000 images crashes as google drive cannot handle so many files within one folder.

is a modification in the code, that is split into train, test and val folders, that contain images possible, this is the link for these splits https://meta-transfer-learning.yaoyao-liu.com/download/

thanks

ok I was able to run it

lmzintgraf commented 5 years ago

Glad to hear you got it to work! Sorry I didn't have time to look into this yet. Did you use the split or did it work without? I like the suggestion of using the split. If you managed to integrate this it would be great if you can open a pull request :)

vainaixr commented 5 years ago

I modified the classification code , I ran it with adabound optimizer, cosine annealing scheduler, and added dropout, should I create a pull request, I have modified certain things, I forked and uploaded it there.

I split dataset into train, val and test, and placed them in miniimagenet/{train}{val}{test} folders.

I added datetime, to save the results in new folder.

also, how to make cavia transductive, or proto cavia, just like we have transductive maml, and proto maml.

lmzintgraf commented 5 years ago

Cool! Did you run the experiment with those changes? I'm very curious what will happen and if you can get an improvement.

A pull request with just the train/va/test set split would be great.

I haven't looked into something like transductive / proto cavia. For Proto-MAML, are you referring to this paper https://arxiv.org/pdf/1903.03096.pdf? I think something similar can be done with CAVIA as well. I'm not familiary with transductive MAML, could you point me to a paper? (Also happy to continue this via email - luisa.zintgraf@cs.ox.ac.uk)