Adapters-for-All: Nyun Zero enables building high-performant AI models through cost-effective fine-tuning
Background
With the growing size of AI models, full fine-tuning of pre-trained models has become increasingly expensive and infeasible. Multiple efficient fine-tuning methods have been proposed in the literature like LoRA which substantially bring down the cost of fine-tuning large models. Recent research has shown that not just large models but even smaller AI models like Resnets, Vision transformers, Yolos have better performance when fine-tuned using efficient fine-tuning methods. With this motivation, Nyun Zero has a brand new plugin that can help users save massive costs in fine-tuning any AI model while surpassing full-finetuning performance - Nyun Adapt!
Efficiency-for-All
Nyun Adapt leverages (Q)LoRA, (Q)SSF and the newly introduced DoRA as the base fine-tuning algorithm over almost any target task including Image Classification, Object Detection, LLMs, Pose Estimation, among others. Adapt supports state-of-the-art vision models like RTMO (Pose Estimation), RT-Detr (Object Detection) and LLMs like Mistral, LLaMA, QWEN, etc. Users can adapt any pre-trained model on their target data with minimal cost and time while attaining superior performance!
Exampler: Pose Estimation with DoRA
To show the superiority of Adapt fine-tuned models we pick a relatively common computer vision task of object detection. We use the current SoTA, a pre-trained RTMDet model and use DoRA to fine-tune it on the Face Detection Dataset. The exact hyperparameters are presented in the table below along with a fair comparison with traditional full fine-tuning.
| Method | Trainable Parameters (%) | Rank | mAP |
|---|---|---|---|
| Traditional | 100.00 | - | 45.5 |
| Adapt - LoRA | 33.08 | 128 | 43.6 |
| Adapt - SSF | 4.68 | - | 45.6 |
Adapt provides direct access to these fine-tuned model weights without any API or middleware enabling users to deploy these high-performing models as they see fit!
Conclusion
With the cost-effective high-performing Adapt plugin on Nyun Zero, users can ship high-performing models with a fractional cost, resource and time. We believe that Nyun Adapt can help revolutionize the traditional deep-learning fine-tuning paradigm and the reduced cost can help onboard the next generation of users onto the deep learning wagon!