harvard-edge/Wake_Vision

The code used to generate the Wake Vision dataset for person detection

Wake Vision

Wake Vision is a Dataset for TinyML person detection. This repository contains the code to generate and filter the dataset from Open Images V7, as well as code to train and evaluate MobileNetV2 models on the dataset. We also provide a suite of benchmarks to evaluate the performance of a person detection model on challenging subsets.

Installation

To install the required packages, run the following command:

pip install -r requirements.txt

Download and build Open Images

Instructions in partial_open_images_v7/README.md

Train a model

To train a MobileNetV2 model using the base config:

python train.py

You can change the config by passing arguments to the train.py script. For example, to change the experiment name and model size, run the following command:

python train.py --experiment_name="name" --model_size=0.5

Alternatively you can change experiment_config.py directly.

Evaluate the model

To run the benchmark suite:

python benchmark_suite.py