Pit Trader

Welcome to the Pit.

Make your fortune in the open outcry pits. Learn the hand signals and scalp your way to profit!

Pit Trader

pit-trader-2-hd-24-1280-720-med-compressed.mp4

Project Layout

• /train: Training data webcam capture and labeling using opencv + mediapipe.

• /model: notebook(s) ingest training data for NN, Logistic Regression and SVC classifiers. Current deploy uses Logistic Regression.

• /web: React-based game. Uses onnyx runtime and mediapipe for live inference.

---

Docker Dev Notes

Quickstart Dev

1. Clone this repository
2. docker-compose up
3. Pit Trader game: http://localhost:3000
4. Training and model environments:

#
# Models: get access url for notebooks via jupyter server logs
#

docker logs pit-trader_pytorch-minimal-notebook_1


#
# Train: Python env
#

./docker_run.sh
python train/webcam_trainer.py


#
# Frontend Game Env
#
./docker_run_web.sh
npm test
npm run lint
(etc)

Dependent Images:

pytorch-minimal-notebook:`

• jupyter/minimal-notebook:notebook-6.4.12
• python related packages: conda, pytorch, opencv (cv2), mediapipe

web:

• node:18.12.1

X windows + Camera Setup in Docker

These are taken care of in docker-compose.yml, docker_run.sh, but are noted here:

• To "receive" GUI run: xhost +local:docker on host before entering docker environment.
• docker exec necessary to pass DISPLAY env variable
• Need to bind mount /tmp/.X11-unix

/dev/video0

• user:root needed to access dev/video0 webcam and docker group with privileges
• Ensure docker-compose.yml: webcam mapped at /dev/video0 for training in pytorch-minimal-notebook container.