This repository contains the source code for interpreting the intermediate layers of the vision encoder component of the CoCa model. The interpretation leverages the open-source implementation of CoCa by mlfoundations.
| Original CoCa Architecture | Visual latent captioning framework |
|---|---|
 |
 |
-
Objective:
Utilize the model to interpret its internal components through natural language descriptions. This self-interpretation is achieved by passing visual features from every layer through cross-attention within the multimodal text decoder of the same model to generated captions per layer.
The generated captions are then further analysized and categorized into visually detectable features and attributes by the use of a large language model. This will allow to get an insite about learned information within the vision encoder at different layers. -
Components:
- Vision Encoder: Extracts visual features from input images at every layer.
- Multimodal Text Decoder: Verbalizes the extracted visual features at every layer.
- Large Language Model: Categorizes the generated interpretations into visually detectable attributes.
-
Visual Features Extraction of Intermediate Layers:
Use the vision encoder to obtain intermediate visual representations from input data. -
Natural Language Insights:
- Generate human-readable explanations of visual feature representations by passing the extracted features through cross-attention layers in the multimodal text decoder of the same model.
- Categorize the generated interpretation for deeper insite into learned information within the layers of the vision encoder using a large language model.
- Architecture Dependency:
This approach is currently limited to multimodal architectures that include a multimodal text decoder.
To get started with this repository, follow these steps:
-
Clone the Repository:
git clone https://github.com/SogolHaghighat/latent_verbalizer.git
-
Install Dependencies:
Create and activate a python or conda environment with python>=3.10.cd latent_verbalizer pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118 pip install .
-
Prepare the data:
To reproduce the results get access to MSCOCO Karpathy test split (5k samples) and store the webdataset format under data. We stored the data in 5 webdataset shards with 1000 sample per shard. The batch size is limited to the GPU memory for batch processing, and the number of epochs is then calculated accordingly. -
Interpretation:
Refer to latent_verbalizer/demo.ipynb for an interactive example of the framework. For reproducing the results follow these steps:python latent_verbalizer/extract_layer_features_coca.py --dataset "data/{000000..000004}.tar" \ --sample-per-shards 1000 \ --batch-size 500 \ --epochs 10 \ --model "coca_ViT-L-14" \ --pretrained "mscoco_finetuned_laion2B-s13B-b90k" \ --num-layers 24 \ --output "data/interpret"
python latent_verbalizer/generate_layers_captions_coca.py --config latent_verbalizer/interpret.yaml
-
Categorization:
Create and activate a new python or conda environment with python>=3.11. Acquire access to Llama-3.1-70B-Instruct and store the model under models. Prepare the environment as below and run the categorization script:pip install torch --index-url https://download.pytorch.org/whl/cu118 pip install "transformers==4.46.2" pip install "accelerate>=0.26.0" python latent_verbalizer/categorization_using_Llama.py --token <provide your HF access token to llama model>
Contributions are welcome! Please open an issue or submit a pull request for any enhancements or bug fixes.
This project is licensed under the MIT License.