When do curricula work?

This repository provides the sample codes for the paper When do curricula work which was accepted at ICLR 2021 for Oral Presentation. It contains the code for experiments on CIFAR10, CIFAR100 and CIFAR100 with noisy labels.

Citation

If you find this code useful in your research then please cite:

@inproceedings{
wu2021when,
title={When Do Curricula Work?},
author={Xiaoxia Wu and Ethan Dyer and Behnam Neyshabur},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=tW4QEInpni}
}

Baseline

python main_w_test.py --ordering standard This command is our standard setup (ResNet50 architecture with SGD Momentum and cosine scheduler) on CIFAR10 dataset with @ 100 epoch.

Curricula includes curriculum, anti-curriculum and random-curriculum learning

curriculum: linear pacing function with paremeter a=0.2 and b=0.8

python main_w_test.py --pacing-a 0.2 --pacing-b 0.8 --pacing-f linear --half

anti-curriculum: linear pacing function with paremeter a=0.2 and b=0.8

python main_w_test.py --pacing-a 0.2 --pacing-b 0.8 --pacing-f linear --ordering anti-curr --half

random-curriculum: linear pacing function with paremeter a=0.2 and b=0.8

python main_w_test.py --pacing-a 0.2 --pacing-b 0.8 --pacing-f linear --ordering random --half

Note "main_wo_test.py" contains no test set but validation set only, while "main_w_test.py" contains validation and test dataset. Our paper uses the best validation accuracy to pick the corresponding test performance (i.e., main_w_test.py).

options:

argument	options
--dataset	cifar10(default) cifar100 cifar100N
--rand-fraction	0(default) 0.2 0.4 0.6 0.8
--ordering	curr(default) anti-curr random standard
--pacing-f	linear(default) quad root step exp log
--pacing-a	[0,10] 1(default)
--pacing-b	[0,1] 1(default)
--order-dir	cifar10-cscores-orig-order.npz (default)
	cifar100-cscores-orig-order.npz
	cifar10.order.pth cifar100.order.pth
	cifar100N02.order.pth cifar100N04.order.pth
	cifar100N06.order.pth cifar100N08.order.pth
--optimizer	sgd(default) nesterov_sgd rmsprop adam
--scheduler	cosine(default) step2 exponential

Sample codes for Section 2 and Section 3

We provide the sample code for reimplementing the experiments for Figure 2 in Section 2:IMPLICIT CURRICULA

mkdir record
python learnT_order.py --logdir record

We provide the sample code for reimplementing the experiments for Figure 3 in Section 3: loss-based c-score.

python loss-based-cscore.py --logdir orders

For more information on this code, please see the '.orders/README.md'

Disclaimer

This is not an official Google product.