Skip to content

Latest commit

 

History

History
63 lines (40 loc) · 2.34 KB

README.md

File metadata and controls

63 lines (40 loc) · 2.34 KB

Neural Network Method for Presence Monitoring in Thermal Images

This directory contains source code enabling the reproduction of a neural network model for presence monitoring tasks in low-resolution thermal images. It is required to download Thermo Presence data (in h5 format) directly to data folder before starting the training process. In order to do so, run get_data.sh script from the console.

bash get_data.sh

Model Training

To train the density estimation model first check the parameters configuration in config.yaml file. Then call train.py script as shown below:

python train.py -p configs/config.yaml

The conducted experiments were tracked using Neptune platform. If one has an account and wants to use this experiment tracker should provide the below data within the configs/credentials.yaml file and then run an experiment with the log_neptune flag.

neptune:
  project: <Project Name>
  api_token: <Project API Token>
$ python train.py --help

Usage: train.py [OPTIONS]

Options:
  -p, --config_path TEXT  Path to file with config
  --log_neptune           Use Neptune experiment tracker
  --help                  Show this message and exit.

Neural Network Architecture

The neural network is based on U-Net architecture with shallow structure, single-channel input, and output. The proposed model has only 46 577 parameters.

Optimization and quantization

Neural network optimization, quantization, and conversion processes are described in the convert_optimize_quantize.ipynb Jupyter Notebook.

Metrics

Metric Name MAE MSE Counting MAE Counting MSE Counting MRAPE [%] Accuracy F1 Score Model size [kB]
TF FP32 0.1057 0.0323 0.0226 0.0234 0.81 0.9778 0.9782 667