Abstract
Parameter efficient finetuning (PEFT) methods are widely used in LLMs and generative models in computer vision. Especially one can use multiple of these during inference to change the behavior of the base model. In this paper we investigated whether multiple LoRA adapters trained on computer vision tasks can be merged together and used during inference without loss in performance. By achieving this, multitask models can be created just by merging different LoRAs. Merging these will reduce inference time and it will not require any additional retraining. We have trained adapters on six different tasks and evaluated their performance when they are merged together. For comparison we used a model with a frozen backbone and finetuned its head. Our results show that even with simple merging techniques creating a multitask model by merging adapters is achievable by slightly loosing performance in some cases. In our experiments we merged up to three adapters together. Depending on the task and the similarity of the data adapters were trained on, merges can outperform head finetuning. We have observed that LoRAs trained with dissimilar datasets tend to perform better compared to model trained on similar datasets.
Abstract (translated)
参数高效微调（PEFT）方法在大型语言模型和计算机视觉生成模型中广泛使用。特别是，人们可以在推理过程中使用多种此类技术来改变基础模型的行为。本文研究了是否可以将多个针对计算机视觉任务训练的LoRA适配器合并起来，并在推理时使用而不影响性能。通过实现这一点，可以通过合并不同的LoRAs创建多任务模型。合并这些适配器将减少推理时间，并且不需要任何额外的再训练。我们在六个不同的任务上训练了适配器，并评估了当它们被合并在一起时的表现。为了进行比较，我们使用了一个带有冻结主干网络并微调其头部的模型。我们的结果显示，即使采用简单的合并技术，在某些情况下通过合并适配器来创建多任务模型也是可行的，尽管可能会略微损失性能。在我们的实验中，我们将最多三个适配器进行了合并。根据任务和数据集的相似性，这些合并可以优于头部微调的表现。我们观察到，使用不相似的数据集训练的LoRAs通常比用相似数据集训练的模型表现更好。
URL
https://arxiv.org/abs/2411.14064
