Vision Face Matcher Model

This repository provides two equivalent scripts to convert the VGG-Face model into a truncated version saved in the ONNX format. This conversion is useful for optimizing face representation tasks (verification). You can choose to run the script either as a Python file for local execution or as a Jupyter Notebook for interactive use.

Table of Contents

• Vision Face Matcher Model
  • Table of Contents
  • Installation
  • Usage
  • Model Architecture
  • Files
  • References
  • License

Installation

To run the scripts, you need to have the following installed:

• Python (3.6 or later)
• TensorFlow (version 2.x preferred)
• tf2onnx (for converting the model to ONNX format)

You can install the required Python packages with the following command:

pip install -U tensorflow tf2onnx

Usage

1. Download the VGG-Face Weights:

   Download the pre-trained VGG-Face model weights using the command:

   wget https://github.com/serengil/deepface_models/releases/download/v1.0/vgg_face_weights.h5

2. Run the Script:

   There are two options to run the script, providing flexibility based on your environment:

   • Python Script: Execute the model_generator_vgg_face_to_onnx_truncated.py file in your local Python environment:

     python model_generator_vgg_face_to_onnx_truncated.py

   • Jupyter Notebook: Open and run the Model_Generator_VGG_Face_to_ONNX_TRUNCATED.ipynb file using Jupyter Notebook for an interactive experience.

3. Convert Model to ONNX:

   To convert the model to the ONNX format, use the following command (assumed to be handled within the scripts):

   python -m tf2onnx.convert --saved-model saved_vgg_face_model_truncated --output saved_vgg_face_model_truncated.onnx --opset 13

4. Saving the Model:

   After conversion, you will have the generated saved_vgg_face_model_truncated.onnx file.

Model Architecture

The VGG-Face model architecture is defined as a Sequential model in TensorFlow. The model is originally designed to classify 2622 identities. The conversion process truncates the model to output embeddings instead of classifications.

• Convolutional Layers: The model includes several convolutional layers with ReLU activations.
• Pooling Layers: Max-pooling layers reduce spatial dimensions.
• Dense Layers: Fully connected layers are followed by dropout for regularization.

The truncated model outputs embeddings from the second-to-last layer, bypassing the final softmax layer used for classification.

Files

• Model_Generator_VGG_Face_to_ONNX_TRUNCATED.ipynb: Jupyter Notebook for interactive execution.
• model_generator_vgg_face_to_onnx_truncated.py: Python script for local execution.

References

• VGG-Face: GitHub Repository
• TensorFlow: Official Site
• tf2onnx: GitHub Repository

License

This project is open-source and available under the MIT License.

For more information, refer to the original implementation by 2060.io and their open-source tools for building verifiable credential-based services.