LLamaRecipes/end-to-end-use-cases/Multi-Modal-RAG

End to End Tutorial on using Llama models for Multi-Modal RAG

Recipe Overview: Multi-Modal RAG using Llama-3.2-11B model:

This is a complete workshop on how to label images using the new Llama 3.2-Vision Models and performing RAG using the image caption capabilities of the model.

• Data Labeling and Preparation: We start by downloading 5000 images of clothing items and labeling them using Llama-3.2-11B-Vision-Instruct model
• Cleaning Labels: With the labels based on the notebook above, we will then clean the dataset and prepare it for RAG
• Building Vector DB and RAG Pipeline: With the final clean dataset, we can use descriptions and 11B model to generate recommendations

Requirements:

Before we start:

1. Please grab your HF CLI Token from here
2. Git clone this dataset inside the Multi-Modal-RAG folder: git clone https://huggingface.co/datasets/Sanyam/MM-Demo (Remember to thank the original author by up voting Kaggle Dataset)
3. Make sure you grab a together.ai token here

Detailed Outline for running:

Order of running files, the notebook establish the method of approaching the problem. Once we establish it, we use the scripts to run the method end to end.

• Notebook 1: Part_1_Data_Preparation.ipynb
• Script: label_script.py
• Notebook 2: Part_2_Cleaning_Data_and_DB.ipynb
• Notebook 3: Part_3_RAG_Setup_and_Validation.ipynb
• Script: final_demo.py

Here's the detailed outline:

Step 1: Data Prep and Synthetic Labeling:

In this step we start with an unlabeled dataset and use the image captioning capability of the model to write a description of the image and categorize it.

Notebook for Step 1 and Script for Step 1

To run the script (remember to set n):

python scripts/label_script.py --hf_token "your_huggingface_token_here" \
    --input_path "../MM-Demo/images_compressed" \
    --output_path "../MM-Demo/output/" \
    --num_gpus N

The dataset consists of 5000 images with some meta-data.

The first half is preparing the dataset for labeling:

• Clean/Remove corrupt images
• Some exploratory analysis to understand existing distribution
• Merging up categories of clothes to reduce complexity
• Balancing dataset by randomly sampling images to have an equal distribution for retrieval

Second Half consists of Labeling the dataset. Llama 3.2, 11B model can only process one image at a time:

• We load a few images and test captioning
• We run this pipeline on random images and iterate on the prompt till we feel the model is giving good outputs
• Finally, we can create a script to label all 5000 images on multi-GPU

After running the script on the entire dataset, we have more data cleaning to perform.

Step 2: Cleaning up Synthetic Labels and preparing the dataset:

Notebook for Step 2

We notice that even after some fun prompt engineering, the model faces some hallucinations-there are some issues with the JSON formatting and we notice that it hallucinates the label categories. Here is how we address this:

• Re-balance the dataset by mapping correct categories. This is useful to make sure we have an equal distribution in our dataset for retrieval
• Fix Descriptions so that we can create a CSV

Now, we are ready to try our vector db pipeline:

Step 3: Notebook 3: MM-RAG using lance-db to validate idea

Notebook for Step 3 and Final Demo Script

With the cleaned descriptions and dataset, we can now store these in a vector-db, here's the steps:

• We create embeddings using the text description of our clothes
• Use 11-B model to describe the uploaded image
• Ask the model to suggest complementary items to the upload
• Try to find similar or complementary images based on the upload

We try the approach with different retrieval methods.

Finally, we can bring this all together in a Gradio App.

For running the script:

python scripts/final_demo.py \
    --images_folder "../MM-Demo/compressed_images" \
    --csv_path "../MM-Demo/final_balanced_sample_dataset.csv" \
    --table_path "~/.lancedb" \
    --api_key "your_together_api_key" \
    --default_model "BAAI/bge-large-en-v1.5" \
    --use_existing_table 

Note: We can further improve the description prompt. You will notice sometimes the description starts with the title of the cloth which causes in retrieval of "similar" clothes instead of "complementary" items

• Upload an image
• 11B model describes the image
• We retrieve complementary clothes to wear based on the description
• You can keep the loop going by chatting with the model

Resources used:

Credit and Thanks to List of models and resources used in the showcase:

Firstly, thanks to the author here for providing this dataset on which we base our exercise here

• Llama-3.2-11B-Vision-Instruct Model
• Lance-db for vector database
• This Kaggle dataset
• HF Dataset Since output of the model can be non-deterministic every time we run, we will use the uploaded dataset to give a universal experience
• Together API for demo