freekang / noisy-maml Goto Github PK

To reproduce the results, we offer the code for "MAML is a noisy contrastive learner" submitted for NeurIPS 2021.

To avoid conflict with your current setup, please create and activate a virtual environment. The author implements the code on Python 3.7 platform. Please install the required packages with pip install -r requirements.txt.

For experiments on miniImagenet dataset, please manually download the miniImagenet dataset here to ./data_miniImagenet folder and unzip it. (ref1 and ref2)

cd ./data_miniImagenet
gdown https://drive.google.com/u/0/uc?id=1HkgrkAwukzEZA0TpO7010PkAOREb2Nukt
unzip mini-imagenet.zip

For experiments on Omniglot dataset, the dataset will be download automatically.

The four folders below provide the code to reproduce the results in Figure.3 ~ Figure.6.

./omniglot_main
./omniglot_memorization
./miniimagenet_main
./miniimagenet_memorization

To run the code, one can run experiment_command.txt inside each folders to get the results. To faithfully reproduce the results, it is worth noted that we use random seed of 222-225.

cd ./miniimagenet
. experiment_command.txt

To visualize the contrastiveness of the MAML algorithm, please go to ./contrastiveness_visualization and run ./contrastivemess_visualization.py to train models and calculate the cosine similarities. One can also refer to the ipython notebook to directly visualize the results.