play-with-torch

Repository for playing the computer vision apps: People analytics on Raspberry Pi.

Tools

Tested Hardware

• RasberryPi 4 Model B here, RAM: 4 GB and Processor 4-core @ 1.5 GHz
• microSD Card 64 GB
• 5M USB Retractable Clip 120 Degrees WebCam Web Wide-angle Camera Laptop U7 Mini or Raspi Camera

Tested Software

• Ubuntu Desktop 20.10 aarch64 64 bit, install on RasberriPi 4
• PyTorch: torch 1.6.0 aarch64 and torchvision 0.7.0 aarch64
• Python min. ver. 3.6 (3.8 recommended)

Install the prerequisites

• Install packages

$ sudo apt install build-essential make cmake git python3-pip libatlas-base-dev
$ sudo apt install libssl-dev
$ sudo apt install libopenblas-dev libblas-dev m4 python3-yaml
$ sudo apt install libomp-dev

• make swap space to 2048 MB

$ free -h
$ sudo swapoff -a
$ sudo dd if=/dev/zero of=/swapfile bs=1M count=2048
$ sudo mkswap /swapfile
$ sudo swapon /swapfile
$ free -h

• Install torch 1.6.0

$ pip3 install torch-1.6.0a0+b31f58d-cp38-cp38-linux_aarch64.whl

Folder Structure

play-with-torch/
├── config/
│    ├── config.json - holds configuration for training
│    └── parse_config.py - class to handle config file and cli options
│
├── docker/
│   ├── Dockerfile
│   └── requirements.txt
│
├── data/ - default directory for storing input data
│
├── docs/ - for documentation
│   └── play-with-torch.tex
│
├── models/ - models, losses, and metrics
│   ├── model.py
│   ├── metric.py
│   └── loss.py
│
├── samples/
│
├── saved/
│   ├── checkpoints/
│   ├── traced_model/
│   ├── models/ - trained models are saved here
│   └── logs/ - default logdir for tensorboard and logging output
│
├── site
├── templates/ - for serving model on Flask
│   └── index.html
├── tests/
├── utils/ - small utility functions
│   ├── data/
│   └── ...
│
├── inference.py - main script to inference model
├── README.md
├── trace_model.py - main script to convert model
└── train.py - main script to start training  

Usage

Run inference

$ git clone https://github.com/mheriyanto/play-with-torch.git
$ cd play-with-torch/
$ python3 inference.py video --config config/nanodet-m.yml --model saved/models/nanodet_m.ckpt --path video.mp4

Convert model

$ python3 trace_model.py --cfg_path config/nanodet-m.yml --model_path saved/models/nanodet_m.ckpt --input_shape 320,320

Training

$ python3 train.py config/nanodet_custom_xml_dataset.yml

TO DO

• Implement Unit-Test: Test-Driven Development (TDD)

Credit to

• Share PyTorch binaries built for Raspberry Pi

Reference

• NanoDet: Super fast and lightweight anchor-free object detection model. here
• Yunjey Choi - PyTorch Tutorial for Deep Learning Researchers here
• Victor Huang - PyTorch Template Project (here)