This repository is a curated collection of research papers on Parameter-Efficient Transfer Learning (PETL) for computer vision and multimodal domains. It serves as a valuable resource for researchers and practitioners looking to adapt large pre-trained models efficiently, addressing challenges like overfitting and high computational costs associated with full fine-tuning.

How It Works

The collection categorizes PETL methods into "Prompt Learning" and "Adapter" techniques, with an "Others" section for related approaches. Prompt learning involves adding task-specific prompts to the input or model architecture, while adapters insert small, trainable modules into the pre-trained model. This approach allows for significant adaptation with minimal parameter updates, preserving the knowledge of the large pre-trained model.

Quick Start & Requirements

This is a curated list of papers, not a software library. No installation or execution is required. Links to papers and code are provided for each entry.

Highlighted Details

• Extensive coverage of prompt learning methods, including visual prompting, unified prompt tuning, and multimodal prompting.
• Detailed listing of adapter-based techniques, such as VL-Adapter, AdaptFormer, and various convolutional adapters.
• Includes papers on automated PETL configuration and unified perspectives on tuning methods.
• Covers applications across image, video, and multimodal tasks, including retrieval and segmentation.

Maintenance & Community

The repository structure and contribution guidelines are inspired by thunlp/DeltaPapers. Contributions are welcomed following a specified format for new paper entries.

Licensing & Compatibility

The repository itself does not have a specific license mentioned, but it links to research papers, each with its own licensing and usage terms. Compatibility for commercial or closed-source use depends on the licenses of the individual papers and their associated code.

Limitations & Caveats

This is a reference list and does not provide runnable code or benchmarks. Users must consult the individual papers for implementation details, performance claims, and specific requirements.