one-bit-supervision

The source code for the NeurIPS 2020 paper "One-bit Supervision for Image Classification"

Requirements:

Python 3; Pytorch 1.0.0

Training:

To train the models on CIFAR100, run these commands:
For stage 0:

python main_stage0.py --train-subdir trainset_path  --test-subdir testset_path  --arch 'cifar_shakeshake26' --labeled-batch-size 50  -b 512 --epochs 180  --lr 0.2  --lr-rampdown-epochs 210 --nesterov 'true'  --ema-decay 0.97  --dataset cifar100  --consistency 1000  --consistency-rampup 5  --logit-distance-cost 0.01  

For stage 1:

python main_stage1.py --train-subdir trainset_path  --test-subdir testset_path  --arch 'cifar_shakeshake26' --labeled-batch-size 200  -b 512  --epochs 180  --lr 0.2  --lr-rampdown-epochs 210 --nesterov 'true'  --ema-decay 0.97  --dataset cifar100  --consistency 1000  --consistency-rampup 5  --logit-distance-cost 0.01  

For stage 2:

python main_stage2.py --train-subdir trainset_path  --test-subdir testset_path  --arch 'cifar_shakeshake26' --labeled-batch-size 320  -b 512  --epochs 180  --lr 0.2  --lr-rampdown-epochs 210 --nesterov 'true'  --ema-decay 0.97  --dataset cifar100  --consistency 1000  --consistency-rampup 5  --logit-distance-cost 0.01  

Evaluation:

To evaluate the model on CIFAR100, run:

python main_stage2.py --arch 'cifar_shakeshake26'  --evaluate pretrained_model_path

Datasets:

The three datasets we used, namely, CIFAR100, Mini-Imagenet, and Imagenet, are all publicly available.

Results:

Dataset	Top 1 Accuracy
CIFAR100	0.7376
Mini-Imagenet	0.4554
Imagenet	0.6040