Fine-tuning pre-trained foundation models has gained significant popularity in various research fields. Existing methods for fine-tuning can be roughly divided into two categories, namely Parameter-Efficient Fine-Tuning and High-Performance Fine-Tuning. The former aims at improving efficiency, while the latter focuses on enhancing performance. Beyond these methods, we demonstrate that Partial Fine-Tuning can be an innovative and promising direction capable of concurrently enhancing both efficiency and accuracy. We first validate eight manually-defined partial fine-tuning strategies across kinds of datasets and vision transformer architectures, and find that some partial fine-tuning strategies (e.g., ffn only or attention only) can achieve better performance with fewer tuned parameters than full fine-tuning, and selecting appropriate layers is critical to partial fine-tuning. Thus, we propose a novel fine-tuned angle metric to guide the selection of appropriate layers for partial fine-tuning, making it flexible to be adapted to various scenarios for more practicable partial fine-tuning. Additionally, we show that partial fine-tuning can serve as a new dimension for Model Soups, improving both the model performance and generalization with fewer tuned parameters. Comprehensive experiments on a wide range of datasets and models validate the great potential of partial fine-tuning.
This work targets to merge various Vision Transformers (ViTs) trained on different tasks (i.e., datasets with different object categories) or domains (i.e., datasets with the same categories but different environments) into one unified model, yielding still good performance on each task or domain. Previous model merging works focus on either CNNs or NLP models, leaving the ViTs merging research untouched. To fill this gap, we first explore and find that existing model merging methods cannot well handle the merging of the whole ViT models and still have improvement space. To enable the merging of the whole ViT, we propose a simple-but-effective gating network that can both merge all kinds of layers (e.g., Embedding, Norm, Attention, and MLP) and select the suitable classifier. Specifically, the gating network is trained by unlabeled datasets from all the tasks (domains), and predicts the probability of which task (domain) the input belongs to for merging the models during inference. To further boost the performance of the merged model, especially when the difficulty of merging tasks increases, we design a novel metric of model weight similarity, and utilize it to realize controllable and combined weight merging. Comprehensive experiments on kinds of newly established benchmarks, validate the superiority of the proposed ViT merging framework for different tasks and domains. Our method can even merge beyond 10 ViT models from different vision tasks with a negligible effect on the performance of each task.
Structural re-parameterization is a general training scheme for Convolutional Neural Networks (CNNs), which achieves performance improvement without increasing inference cost. As Vision Transformers (ViTs) are gradually surpassing CNNs in various visual tasks, one may question: if a training scheme specifically for ViTs exists that can also achieve performance improvement without increasing inference cost? Recently, Mixture-of-Experts (MoE) has attracted increasing attention, as it can efficiently scale up the capacity of Transformers at a fixed cost through sparsely activated experts. Considering that MoE can also be viewed as a multi-branch structure, can we utilize MoE to implement a ViT training scheme similar to structural re-parameterization? In this paper, we affirmatively answer these questions, with a new general training strategy for ViTs. Specifically, we decouple the training and inference phases of ViTs. During training, we replace some Feed-Forward Networks (FFNs) of the ViT with specially designed, more efficient MoEs that assign tokens to experts by random uniform partition, and perform Experts Weights Averaging (EWA) on these MoEs at the end of each iteration. After training, we convert each MoE into an FFN by averaging the experts, transforming the model back into original ViT for inference. We further provide a theoretical analysis to show why and how it works. Comprehensive experiments across various 2D and 3D visual tasks, ViT architectures, and datasets validate the effectiveness and generalizability of the proposed training scheme. Besides, our training scheme can also be applied to improve performance when fine-tuning ViTs. Lastly, but equally important, the proposed EWA technique can significantly improve the effectiveness of naive MoE in various 2D visual small datasets and 3D visual tasks.