EFFICIENT PRETRAINING OF VISION TRANSFORMERS: A LAYER-FREEZING APPROACH WITH LOCAL MASKED IMAGE MODELING

Date

2024-9-03

Author

Topçuoğlu, Utku Mert

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

60
views

49
downloads

Cite This

This thesis explores the acceleration of pre-training Vision Transformers (ViTs) for self-supervised learning by integrating progressive layer freezing with local masked image modeling. The study aims to address the significant computational demands and lengthy training times inherent in training ViTs when employing self-supervised methods like masked image modeling. The core contribution of this research lies in integrating the FreezeOut method into the LocalMIM architecture, enhancing training efficiency by systematically freezing specific layers at strategic points during training. We evaluate whether the FreezeOut method is as effective as proposed in the original paper across different optimizers, acknowledging that learning rate scheduling is optimizer-dependent. Our experimental results demonstrate that the proposed approach can reduce training time by approximately 12.5% with a minimal drop in top-1 accuracy (0.6%). Furthermore, we introduce and validate a novel learning rate scheduling method tailored for ViTs, which achieves an even more negligible accuracy drop of 0.1% with an 83.1% top-1 accuracy. We demonstrate that the number of training epochs and dataset complexity are critical factors for the effectiveness of the FreezeOut method and show that it performs even better with longer training epochs or simpler datasets. Our specially designed learning rate scheduling method showed greater robustness to fewer training epochs and more complex datasets, explaining its superior results in the 100 epoch IN-1K training setup. This research offers a solution for enhancing the efficiency of ViT pre-training, making self-supervised learning more accessible in environments with constrained computational resources. The findings contribute to the broader field of computer vision by highlighting the potential of progressive layer freezing and adaptive learning rate scheduling in optimizing training processes for ViTs. The implementation of our approach is accessible here: https://github.com/utkutpcgl/ViTFreeze.

Subject Keywords

Vision Transformers, Self-Supervised Learning, Local Masked Image Modeling, Progres- sive Layer Freezing, Computational Efficiency, Multi-Scale Reconstruction, Adaptive Learning Rate Scheduling, FreezeOut Learning Rate Scheduling, Training Time Reduction

URI

https://hdl.handle.net/11511/111029

Collections

Graduate School of Informatics, Thesis