Skip to content

Latest commit

 

History

History
39 lines (22 loc) · 2.66 KB

README.md

File metadata and controls

39 lines (22 loc) · 2.66 KB
Raw

Cross-modal Few-shot Learning Implementation

Overview

This repository provides a naive implementation of the cross-modal few-shot learning approach presented in this paper https://arxiv.org/pdf/2301.06267.pdf. The primary goal of this paper is to improve few-shot learning. In this implementation, we demonstrated that the technique also enhances the classifier's accuracy in a non few-shot settings while harnessing the significant boost of inference speed compare to methods that don't leverage the power of pre-trained embedding models such as CLIP.

Key Features

Multimodal Foundation: Utilizes models like CLIP that map different modalities to the same representation space. Like the paper, this implementation supports the embedding of three modalities: visual, textual, and audio.

Cross-modal Adaptation: A simple approach that learns from the embedded examples across different modalities using shallow classifiers.

Enhanced Classifier: Our experiments (though many aspects of the design should be made more rigorous and improved) showed superiority in accuracy using an image-text classifier compare to a fresh ResNet50 (image only) and a further-trained top-ranked classifier by https://www.kaggle.com/vlomme (audio only) on a curated subset of Cornell Birdcall Identification dataset.

Future Work

LLM Prompting: We have considered using LLM prompting techniques to generate textual data that aligns with how CLIP are trained for further experiments to see whether this can be an effective model-improving strategy. Segmentation Techniques: Segmentation techniques like SAM have also been thought of to improve the quality and robustness of image data but haven't been implemented yet.

Getting Started

Setting up Dependencies

To set up the dependencies for the project, please follow the instructions below:

  1. Open your command prompt or terminal.

  2. Navigate to the project directory.

  3. Run the command conda env create -f environment.yml. This will create a new conda environment with the name specified in the environment.yml file, and install all the necessary dependencies.

  4. Once the installation is complete, activate the environment by running the command conda activate <environment-name>. Replace <environment-name> with the name of the environment created in step 3.

  5. You're now ready to use the project with all the required dependencies installed.

Note: Make sure you have conda installed on your system before following the above steps.