ImageNet-21K Pretraining for the Masses
This work makes high-quality pretraining on a large, diverse dataset accessible to a broad audience in computer vision, addressing a practical bottleneck for researchers and practitioners.
The paper tackled the underutilization of the ImageNet-21K dataset for pretraining due to complexity and accessibility issues, showing that their efficient pipeline with semantic softmax significantly improves model performance across various tasks, including achieving state-of-the-art results for models like ViT and Mixer.
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