nanxuanzhao/Good_transfer

Pretrained models for "What Makes Instance Discrimination Good for Transfer Learning?".

What Makes Instance Discrimination Good for Transfer Learning?

What Makes Instance Discrimination Good for Transfer Learning?
Nanxuan Zhao* Zhirong Wu* Rynson W.H. Lau Stephen Lin

Pretrained Models

Different data augmentations for learning self-supervised and supervised representations (Table 1).

Pretraining	Pytorch Augmentation	Download
Unsupervised	+ RandomHorizontalFlip(0.5)	model
	+ RandomResizedCrop(224)	model
	+ ColorJitter(0.4, 0.4, 0.4, 0.1)	model
	+ RandomGrayscale(p=0.2)	model
	+ GaussianBlur(0.1, 0.2)	model
supervised	+ RandomHorizontalFlip(0.5)	model
	+ RandomResizedCrop(224)	model
	+ ColorJitter(0.4, 0.4, 0.4, 0.1)	model
	+ RandomGrayscale(p=0.2)	model
	+ GaussianBlur(0.1, 0.2)	model

Transfer performance with pretraining on various datasets (Table 2).

Pretraining	Pretraining Data	Download
Unsupervised	ImageNet	model
	ImageNet-10%	model
	ImageNet-100	model
	Places	model
	CelebA	model
	COCO	model
	Synthia	model
Supervised	ImageNet	model
	ImageNet-10%	model
	ImageNet-100	model
	Places	model
	CelebA	model
	COCO	model
	Synthia	model

Exemplar-based supervised pretraining (Table 3).

Model	Download
Exemplar v1	model
Exemplar v2	model

Citation

If you use this work in your research, please cite:

@inproceedings{ZhaoICLR2021, 
    author = {Nanxuan Zhao and Zhirong Wu and Rynson W.H. Lau and Stephen Lin}, 
    title = {What Makes Instance Discrimination Good for Transfer Learning?}, 
    booktitle = {ICLR}, 
    year = {2021} 
}