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

Related tags

Deep Learning ATS
Overview

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.

You might also like...
Lightweight, Portable, Flexible Distributed/Mobile Deep Learning with Dynamic, Mutation-aware Dataflow Dep Scheduler; for Python, R, Julia, Scala, Go, Javascript and more
Lightweight, Portable, Flexible Distributed/Mobile Deep Learning with Dynamic, Mutation-aware Dataflow Dep Scheduler; for Python, R, Julia, Scala, Go, Javascript and more

Apache MXNet (incubating) for Deep Learning Master Docs License Apache MXNet (incubating) is a deep learning framework designed for both efficiency an

ROCm Software Platform 29 Nov 16, 2022
Source code for paper "Document-Level Relation Extraction with Adaptive Thresholding and Localized Context Pooling", AAAI 2021

ATLOP Code for AAAI 2021 paper Document-Level Relation Extraction with Adaptive Thresholding and Localized Context Pooling. If you make use of this co

Wenxuan Zhou 146 Nov 29, 2022
This is the source code for our ICLR2021 paper: Adaptive Universal Generalized PageRank Graph Neural Network.
This is the source code for our ICLR2021 paper: Adaptive Universal Generalized PageRank Graph Neural Network.

GPRGNN This is the source code for our ICLR2021 paper: Adaptive Universal Generalized PageRank Graph Neural Network. Hidden state feature extraction i

Jianhao 92 Jan 3, 2023
Image data augmentation scheduler for albumentations transforms

albu_scheduler Scheduler for albumentations transforms based on PyTorch schedulers interface Usage TransformMultiStepScheduler import albumentations a

null 19 Aug 4, 2021
This is the code of NeurIPS'21 paper "Towards Enabling Meta-Learning from Target Models".

ST This is the code of NeurIPS 2021 paper "Towards Enabling Meta-Learning from Target Models". If you use any content of this repo for your work, plea

Su Lu 7 Dec 6, 2022
[WWW 2021] Source code for "Graph Contrastive Learning with Adaptive Augmentation"

GCA Source code for Graph Contrastive Learning with Adaptive Augmentation (WWW 2021) For example, to run GCA-Degree under WikiCS, execute: python trai

Big Data and Multi-modal Computing Group, CRIPAC 97 Jan 7, 2023
PyTorch implementation of the supervised learning experiments from the paper Model-Agnostic Meta-Learning (MAML)
PyTorch implementation of the supervised learning experiments from the paper Model-Agnostic Meta-Learning (MAML)

pytorch-maml This is a PyTorch implementation of the supervised learning experiments from the paper Model-Agnostic Meta-Learning (MAML): https://arxiv

Kate Rakelly 516 Jan 5, 2023
Supplementary code for the paper
Supplementary code for the paper "Meta-Solver for Neural Ordinary Differential Equations" https://arxiv.org/abs/2103.08561

Meta-Solver for Neural Ordinary Differential Equations Towards robust neural ODEs using parametrized solvers. Main idea Each Runge-Kutta (RK) solver w

Julia Gusak 25 Aug 12, 2021
Code and data of the ACL 2021 paper: Few-Shot Text Ranking with Meta Adapted Synthetic Weak Supervision

MetaAdaptRank This repository provides the implementation of meta-learning to reweight synthetic weak supervision data described in the paper Few-Shot

THUNLP 5 Jun 16, 2022
Comments
  • Question about the performance of ANIL

    Question about the performance of ANIL

    Hi, I try to extract the ANIL module and ran the experiment(5way-1shot) based on the dataset miniimagenet, but the test set performance is just 44, which not matched with the original paper performance 46.7+-0.4.

    Would you mind sharing some advice ? Thank you.

    2331638536266_ pic_hd

    opened by peterdarkdarkgogo 2
  • Questions about scheduler update

    Questions about scheduler update

    Hiļ¼Œthanks for releasing this great work. I'm a little confused about equation 8 $$\phi^{(k+1)} \leftarrow \phi^{(k)}-\gamma \nabla_{\phi^{(k)}} \log P\left(\mathbf{W}^{(k)} ; \phi^{(k)}\right)\left(\frac{1}{N_{v}} \sum_{i=1}^{N_{v}} R_{i}^{(k)}-b\right)$$ The reward $R$ is calculated from the updated temporal model $\tilde{\theta}_{0}^{(k+1)}$ on $N_v$ validation tasks, I don't quite understand why this can be used to update the scheduler as part of the loss, it seems that the scheduler is not involved in the calculation of the reward $R$?

    opened by toooooodo 0
Owner
Huaxiu Yao
Postdoctoral Scholar at CS@Stanford
Huaxiu Yao
GitHub
Code for the ICML 2021 paper "Bridging Multi-Task Learning and Meta-Learning: Towards Efficient Training and Effective Adaptation", Haoxiang Wang, Han Zhao, Bo Li.

Bridging Multi-Task Learning and Meta-Learning Code for the ICML 2021 paper "Bridging Multi-Task Learning and Meta-Learning: Towards Efficient Trainin

AI Secure 57 Dec 15, 2022
DexterRedTool - Dexter's Red Team Tool that creates cronjob/task scheduler to consistently creates users

DexterRedTool Author: Dexter Delandro CSEC 473 - Spring 2022 This tool persisten

null 2 Feb 16, 2022
PyTorch implementation of the paper: "Preference-Adaptive Meta-Learning for Cold-Start Recommendation", IJCAI, 2021.

PAML PyTorch implementation of the paper: "Preference-Adaptive Meta-Learning for Cold-Start Recommendation", IJCAI, 2021. (Continuously updating ) Int

null 15 Nov 18, 2022
Source code for CVPR 2020 paper "Learning to Forget for Meta-Learning"

L2F - Learning to Forget for Meta-Learning Sungyong Baik, Seokil Hong, Kyoung Mu Lee Source code for CVPR 2020 paper "Learning to Forget for Meta-Lear

Sungyong Baik 29 May 22, 2022
Implementation of "Meta-rPPG: Remote Heart Rate Estimation Using a Transductive Meta-Learner"

Meta-rPPG: Remote Heart Rate Estimation Using a Transductive Meta-Learner This repository is the official implementation of Meta-rPPG: Remote Heart Ra

Eugene Lee 137 Dec 13, 2022
DeepMind Alchemy task environment: a meta-reinforcement learning benchmark

The DeepMind Alchemy environment is a meta-reinforcement learning benchmark that presents tasks sampled from a task distribution with deep underlying structure.

DeepMind 188 Dec 25, 2022
PyTorch Implementation of Meta-StyleSpeech : Multi-Speaker Adaptive Text-to-Speech Generation

StyleSpeech - PyTorch Implementation PyTorch Implementation of Meta-StyleSpeech : Multi-Speaker Adaptive Text-to-Speech Generation. Status (2021.06.13

Keon Lee 140 Dec 21, 2022
Lightweight, Portable, Flexible Distributed/Mobile Deep Learning with Dynamic, Mutation-aware Dataflow Dep Scheduler; for Python, R, Julia, Scala, Go, Javascript and more

Apache MXNet (incubating) for Deep Learning Apache MXNet is a deep learning framework designed for both efficiency and flexibility. It allows you to m

The Apache Software Foundation 20.2k Jan 8, 2023
Lightweight, Portable, Flexible Distributed/Mobile Deep Learning with Dynamic, Mutation-aware Dataflow Dep Scheduler; for Python, R, Julia, Scala, Go, Javascript and more

Apache MXNet (incubating) for Deep Learning Apache MXNet is a deep learning framework designed for both efficiency and flexibility. It allows you to m

The Apache Software Foundation 20.2k Jan 5, 2023
Lightweight, Portable, Flexible Distributed/Mobile Deep Learning with Dynamic, Mutation-aware Dataflow Dep Scheduler; for Python, R, Julia, Scala, Go, Javascript and more

Apache MXNet (incubating) for Deep Learning Apache MXNet is a deep learning framework designed for both efficiency and flexibility. It allows you to m

The Apache Software Foundation 19.3k Feb 12, 2021