Abstract

The ubiquity of vision transformers (ViTs) for various edge applications, including personalized learning, has created the demand for on-device fine-tuning. However, training with the limited memory and computation power of edge devices remains a significant challenge. In particular, the memory required for training is much higher than that needed for inference, primarily due to the need to store activations across all layers in order to compute the gradients needed for weight updates. Previous works have explored reducing this memory requirement via frozen-weight training as well storing the activations in a compressed format. However, these methods are deemed inefficient due to their inability to provide training or inference speedup. In this paper, we first investigate the limitations of existing on-device training methods aimed at reducing memory and compute requirements. We then present block selective reprogramming (BSR) in which we fine-tune only a fraction of total blocks of a pre-trained model and selectively drop tokens based on self-attention scores of the frozen layers. To show the efficacy of BSR, we present extensive evaluations on ViT-B and DeiT-S with five different datasets. Compared to the existing alternatives, our approach simultaneously reduces training memory by up to 1.4 x and compute cost by up to 2x while maintaining similar accuracy. We also showcase results for Mixture-of-Expert (MoE) models, demonstrating the effectiveness of our approach in multi-task learning scenarios.

Metadata

publication

arXiv preprint arXiv:2405.10951, 2024

year

2024

publication date

2024

authors

Sreetama Sarkar, Souvik Kundu, Kai Zheng, Peter A Beerel

link

https://openaccess.thecvf.com/content/CVPR2024W/ECV24/html/Sarkar_Block_Selective_Reprogramming_for_On-device_Training_of_Vision_Transformers_CVPRW_2024_paper.html

resource_link

https://openaccess.thecvf.com/content/CVPR2024W/ECV24/papers/Sarkar_Block_Selective_Reprogramming_for_On-device_Training_of_Vision_Transformers_CVPRW_2024_paper.pdf

conference

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

pages

8094-8103