About The Project

This project is done in the context of the course "Machine Learning on Microcontrollers" @ ETH Zurich. It contains a reduced Keras implementation of MTCNN and is trained using the WIDER FACE dataset. It also contains a quantization part where the networks (weights and activations) are quantized from float32 to int8 without a major loss of accuarcy.

Installation

1. Clone the repo
2. Setup an environment according to the environment.yml file
3. Set the source node as a environment variable

   export PYTHONPATH="path to your directory"

Usage

1. Download the data with download_data.py

   python data/download_data.py

2. Generate the training images for the pnet, rnet and onet (with arguments 12, 24 and 48 respectively) with generate_training_data.py

   python data/generate_training_data.py 12

3. Train the models with train_pnet.py, train_rnet.py and train_onet.py

   python models/train_pnet.py

4. Quantize the models with quantize_pnet.py, quantize_rnet.py and quantize_onet.py

   python models/quantize_pnet.py

In models, you will find *.h5, *.tflite and .h files, for the keras model, tensorflow lite model and a hex representation of them. The quantization files also give you the possibility to generate validation data for STM32 X-CUBE-AI expansion (use the keras models *.h5 and the *net_bbx.csv, *net_cat.csv and *net_data.csv files).

To increase the performance, you can also forward sample with data/generate_hard_samples for the rnet and onet with the arguments 24 and 48 respectively.

5. To test the performance, run

python detect_faces "path to your image" 

Acknowledgements

You will recognise some code sections or even entire files from the following repositories:

• MTCNN
• MTCNN-Tensorflow