ATS

About

Source code of the paper Meta-learning with an Adaptive Task Scheduler.

If you find this repository useful in your research, please cite the following paper:

@inproceedings{yao2021adaptive,
  title={Meta-learning with an Adaptive Task Scheduler},
  author={Yao, Huaxiu and Wang, Yu and Wei, Ying and Zhao, Peilin and Mahdavi, Mehrdad and Lian, Defu and Finn, Chelsea},
  booktitle={Proceedings of the Thirty-fifth Conference on Neural Information Processing Systems},
  year={2021} 
}

Miniimagenet

The processed miniimagenet dataset could be downloaded here. Assume the dataset has been downloaded and unzipped to /data/miniimagenet, which has the following file structure:

-- miniimagenet  // /data/miniimagenet
  -- miniImagenet
    -- train_task_id.pkl
    -- test_task_id.pkl
    -- mini_imagenet_train.pkl
    -- mini_imagenet_test.pkl
    -- mini_imagenet_val.pkl
    -- training_classes_20000_2_new.npz
    -- training_classes_20000_4_new.npz

Then $datadir in the following code sould be set to /data/miniimagenet.

ATS with noise = 0.6

We need to first pretrain the model with no noise. The model has been uploaded to this repo. You can also pretrain the model by yourself. The script for pretraining is as follows:
(1) 1 shot:

python3 main.py --meta_batch_size 2 --datasource miniimagenet --datadir $datadir --num_updates 5 --num_updates_test 10 --update_batch_size 1 --update_batch_size_eval 15 --resume 0  --num_classes 5 --metatrain_iterations 30000 --logdir $logdir --noise 0.0

(2) 5 shot:

python3 main.py --meta_batch_size 2 --datasource miniimagenet --datadir $datadir --num_updates 5 --num_updates_test 10 --update_batch_size 5 --update_batch_size_eval 15 --resume 0  --num_classes 5 --metatrain_iterations 30000 --logdir $logdir --noise 0.0

Then move the model to the current directory:
(1) 1 shot:

mv $logdir/ANIL_pytorch.data_miniimagenetcls_5.mbs_2.ubs_1.metalr0.001.innerlr0.01.hidden32/model20000 ./model20000_1shot

(2) 5 shot:

mv $logdir/ANIL_pytorch.data_miniimagenetcls_5.mbs_2.ubs_5.metalr0.001.innerlr0.01.hidden32/model10000 ./model10000_5shot

Then with this model, we could run the uniform sampling and ATS sampling.
For ATS, the script is:
(1) 1 shot

python3 main.py --meta_batch_size 2 --datasource miniimagenet --datadir $datadir --num_updates 5 --num_updates_test 10 --update_batch_size 1 --update_batch_size_eval 15 --resume 0 --num_classes 5 --metatrain_iterations 30000 --replace 0 --noise 0.6 --logdir $logdir --sampling_method ATS --buffer_size 10  --temperature 0.1 --scheduler_lr 0.001 --warmup 2000 --pretrain_iter 20000

(2) 5 shot

python3 main.py --meta_batch_size 2 --datasource miniimagenet --datadir $datadir --num_updates 5 --num_updates_test 10 --update_batch_size 5 --update_batch_size_eval 15 --resume 0  --num_classes 5 --metatrain_iterations 30000 --replace 0 --noise 0.6 --logdir $logdir --sampling_method ATS --buffer_size 10 --utility_function sample --temperature 0.1 --scheduler_lr 0.001 --warmup 2000 --pretrain_iter 10000

For uniform sampling, we need to use the validation set to finetune the model trained under uniform sampling. The training commands are:
(1) 1 shot

python3 main.py --meta_batch_size 2 --datasource miniimagenet --datadir $datadir --num_updates 5 --num_updates_test 10 --update_batch_size 1 --update_batch_size_eval 15 --resume 0 --num_classes 5 --metatrain_iterations 30000 --logdir $logdir --noise 0.6
mkdir models
mv ANIL_pytorch.data_miniimagenetcls_5.mbs_2.ubs_1.metalr0.001.innerlr0.01.hidden32_noise0.6/model30000 ./models/ANIL_0.4_model_1shot
python3 main.py --meta_batch_size 2 --datasource miniimagenet --datadir $datadir --num_updates 5 --num_updates_test 10 --update_batch_size 1 --update_batch_size_eval 15 --resume 0 --num_classes 5 --metatrain_iterations 30000 --logdir $logdir --noise 0.6 --finetune

(2) 5 shot

python3 main.py --meta_batch_size 2 --datasource miniimagenet --datadir $datadir --num_updates 5 --num_updates_test 10 --update_batch_size 5 --update_batch_size_eval 15 --resume 0  --num_classes 5 --metatrain_iterations 30000 --logdir $logdir --noise 0.6
mkdir models  // if directory "models" does not exist
mv ANIL_pytorch.data_miniimagenetcls_5.mbs_2.ubs_5.metalr0.001.innerlr0.01.hidden32_noise0.6/model30000 ./models/ANIL_0.4_model_5shot
python3 main.py --meta_batch_size 2 --datasource miniimagenet --datadir $datadir --num_updates 5 --num_updates_test 10 --update_batch_size 5 --update_batch_size_eval 15 --resume 0  --num_classes 5 --metatrain_iterations 30000 --logdir $logdir --noise 0.6 --finetune

ATS with limited budgets

In this setting, pretraining is not needed. You can directly run the following code:
uniform sampling, 1 shot

python3 main.py --meta_batch_size 3 --datasource miniimagenet --datadir ./miniimagenet/ --num_updates 5 --num_updates_test 10 --update_batch_size 1 --update_batch_size_eval 15 --resume 0  --num_classes 5 --metatrain_iterations 30000 --limit_data 1 --logdir ../train_logs --limit_classes 16

uniform sampling, 5 shot

python3 main.py --meta_batch_size 3 --datasource miniimagenet --datadir ./miniimagenet/ --num_updates 5 --num_updates_test 10 --update_batch_size 5 --update_batch_size_eval 15 --resume 0  --num_classes 5 --metatrain_iterations 30000 --limit_data 1 --logdir ../train_logs --limit_classes 16

ATS 1 shot

python3 main.py --meta_batch_size 3 --datasource miniimagenet --datadir ./miniimagenet/ --num_updates 5 --num_updates_test 10 --update_batch_size 1 --update_batch_size_eval 15 --resume 0  --num_classes 5 --metatrain_iterations 30000 --replace 0 --limit_data 1 --logdir ../train_logs --sampling_method ATS --buffer_size 6 --utility_function sample --temperature 1 --warmup 0 --limit_classes 16

ATS 5 shot

python3 main.py --meta_batch_size 3 --datasource miniimagenet --datadir ./miniimagenet/ --num_updates 5 --num_updates_test 10 --update_batch_size 5 --update_batch_size_eval 15 --resume 0  --num_classes 5 --metatrain_iterations 30000 --replace 0 --limit_data 1 --logdir ../train_logs --sampling_method ATS --buffer_size 6 --utility_function sample --temperature 0.1 --warmup 0 --limit_classes 16

Drug

The processed dataset could be downloaded here.
Assume the dataset has been downloaded and unzipped to /data/drug which has the following structure:

-- drug  // /data/drug
  -- ci9b00375_si_001.txt  
  -- compound_fp.npy               
  -- drug_split_id_group2.pickle  
  -- drug_split_id_group6.pickle
  -- ci9b00375_si_002.txt  
  -- drug_split_id_group17.pickle  
  -- drug_split_id_group3.pickle  
  -- drug_split_id_group9.pickle
  -- ci9b00375_si_003.txt  
  -- drug_split_id_group1.pickle   
  -- drug_split_id_group4.pickle  
  -- important_readme.md

Then $datadir in the following script should be set as /data/.

ATS with noise=4.

Uniform Sampling:

python3 main.py --datasource=drug --metatrain_iterations=20 --update_lr=0.005 --meta_lr=0.001 --num_updates=5 --test_num_updates=5 --trial=1 --drug_group=17 --noise 4 --data_dir $datadir
python3 main.py --datasource=drug --metatrain_iterations=20 --update_lr=0.005 --meta_lr=0.001 --num_updates=5 --test_num_updates=5 --trial=1 --drug_group=17 --noise 4 --data_dir $datadir --train 0

ATS:

python3 main.py --datasource=drug --metatrain_iterations=20 --update_lr=0.005 --meta_lr=0.001 --num_updates=5 --test_num_updates=5 --trial=1 --drug_group=17 --sampling_method ATS --noise 4 --data_dir $datadir
python3 main.py --datasource=drug --metatrain_iterations=20 --update_lr=0.005 --meta_lr=0.001 --num_updates=5 --test_num_updates=5 --trial=1 --drug_group=17 --sampling_method ATS --noise 4 --data_dir $datadir --train 0

ATS with full budgets

Uniform Sampling:

python3 main.py --datasource=drug --metatrain_iterations=20 --update_lr=0.005 --meta_lr=0.001 --num_updates=5 --test_num_updates=5 --trial=1 --drug_group=17 --data_dir $datadir
python3 main.py --datasource=drug --metatrain_iterations=20 --update_lr=0.005 --meta_lr=0.001 --num_updates=5 --test_num_updates=5 --trial=1 --drug_group=17 --data_dir $datadir --train 0

ATS:

python3 main.py --datasource=drug --metatrain_iterations=20 --update_lr=0.005 --meta_lr=0.001 --num_updates=5 --test_num_updates=5 --trial=1 --drug_group=17 --sampling_method ATS --data_dir $datadir
python3 main.py --datasource=drug --metatrain_iterations=20 --update_lr=0.005 --meta_lr=0.001 --num_updates=5 --test_num_updates=5 --trial=1 --drug_group=17 --sampling_method ATS --data_dir $datadir --train 0

For ATS, if you need to use 1 for calculating the loss as the input of the scheduler instead of 1, you can add --simple_loss after the script above.

GitHub

View Github