ImageNet-21K Pretraining for the Masses

Tal Ridnik; Emanuel Ben-Baruch; Asaf Noy; Lihi Zelnik-Manor

ImageNet-21K Pretraining for the Masses

Tal Ridnik, Emanuel Ben-Baruch, Asaf Noy, Lihi Zelnik-Manor

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

Abstract

ImageNet-1K serves as the primary dataset for pretraining deep learning models for computer vision tasks. ImageNet-21K dataset, which is bigger and more diverse, is used less frequently for pretraining, mainly due to its complexity, low accessibility, and underestimation of its added value.
This paper aims to close this gap, and make high-quality efficient pretraining on ImageNet-21K available for everyone.
Via a dedicated preprocessing stage, utilization of WordNet hierarchical structure, and a novel training scheme called semantic softmax, we show that various models significantly benefit from ImageNet-21K pretraining on numerous datasets and tasks, including small mobile-oriented models.
We also show that we outperform previous ImageNet-21K pretraining schemes for prominent new models like ViT and Mixer.
Our proposed pretraining pipeline is efficient, accessible, and leads to SoTA reproducible results, from a publicly available dataset. The training code and pretrained models are available at: https://github.com/Alibaba-MIIL/ImageNet21K

Original language	English
Title of host publication	Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1)
State	Published - 2021
Externally published	Yes

Access to Document

https://openreview.net/forum?id=Zkj_VcZ6ol

Cite this

@inproceedings{1f6de63a39be4a689603a42747c195fa,

title = "ImageNet-21K Pretraining for the Masses",

abstract = "ImageNet-1K serves as the primary dataset for pretraining deep learning models for computer vision tasks. ImageNet-21K dataset, which is bigger and more diverse, is used less frequently for pretraining, mainly due to its complexity, low accessibility, and underestimation of its added value.This paper aims to close this gap, and make high-quality efficient pretraining on ImageNet-21K available for everyone.Via a dedicated preprocessing stage, utilization of WordNet hierarchical structure, and a novel training scheme called semantic softmax, we show that various models significantly benefit from ImageNet-21K pretraining on numerous datasets and tasks, including small mobile-oriented models. We also show that we outperform previous ImageNet-21K pretraining schemes for prominent new models like ViT and Mixer.Our proposed pretraining pipeline is efficient, accessible, and leads to SoTA reproducible results, from a publicly available dataset. The training code and pretrained models are available at: https://github.com/Alibaba-MIIL/ImageNet21K",

author = "Tal Ridnik and Emanuel Ben-Baruch and Asaf Noy and Lihi Zelnik-Manor",

year = "2021",

language = "الإنجليزيّة",

booktitle = "Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1)",

}

TY - GEN

T1 - ImageNet-21K Pretraining for the Masses

AU - Ridnik, Tal

AU - Ben-Baruch, Emanuel

AU - Noy, Asaf

AU - Zelnik-Manor, Lihi

PY - 2021

Y1 - 2021

N2 - ImageNet-1K serves as the primary dataset for pretraining deep learning models for computer vision tasks. ImageNet-21K dataset, which is bigger and more diverse, is used less frequently for pretraining, mainly due to its complexity, low accessibility, and underestimation of its added value.This paper aims to close this gap, and make high-quality efficient pretraining on ImageNet-21K available for everyone.Via a dedicated preprocessing stage, utilization of WordNet hierarchical structure, and a novel training scheme called semantic softmax, we show that various models significantly benefit from ImageNet-21K pretraining on numerous datasets and tasks, including small mobile-oriented models. We also show that we outperform previous ImageNet-21K pretraining schemes for prominent new models like ViT and Mixer.Our proposed pretraining pipeline is efficient, accessible, and leads to SoTA reproducible results, from a publicly available dataset. The training code and pretrained models are available at: https://github.com/Alibaba-MIIL/ImageNet21K

AB - ImageNet-1K serves as the primary dataset for pretraining deep learning models for computer vision tasks. ImageNet-21K dataset, which is bigger and more diverse, is used less frequently for pretraining, mainly due to its complexity, low accessibility, and underestimation of its added value.This paper aims to close this gap, and make high-quality efficient pretraining on ImageNet-21K available for everyone.Via a dedicated preprocessing stage, utilization of WordNet hierarchical structure, and a novel training scheme called semantic softmax, we show that various models significantly benefit from ImageNet-21K pretraining on numerous datasets and tasks, including small mobile-oriented models. We also show that we outperform previous ImageNet-21K pretraining schemes for prominent new models like ViT and Mixer.Our proposed pretraining pipeline is efficient, accessible, and leads to SoTA reproducible results, from a publicly available dataset. The training code and pretrained models are available at: https://github.com/Alibaba-MIIL/ImageNet21K

M3 - منشور من مؤتمر

BT - Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 1)

ER -

ImageNet-21K Pretraining for the Masses

Abstract

Access to Document

Fingerprint

Cite this