This repository develops yolov3 using hasktorch. It is based on https://github.com/eriklindernoren/PyTorch-YOLOv3.

It supports the format of the original yolo config and weight files. See this link for original yolo.

For now, we use pre-trained weights. We do not support testing and training, yet.

The sources of the config and weight file are following links. These are already included in this repository.

https://raw.githubusercontent.com/eriklindernoren/PyTorch-YOLOv3/master/config/yolov3.cfg
https://pjreddie.com/media/files/yolov3.weights

Getting Started

linux+cabal

git clone git@github.com:junjihashimoto/hasktorch-yolo.git
cd hasktorch-yolo
./download-weights.sh
cabal test all

linux+nix

git clone git@github.com:junjihashimoto/hasktorch-yolo.git
cd hasktorch-yolo
./download-weights.sh
nix-build

Inference

Use the following command for inference. An image with a bounding box is output. The execution example is fig.1.

linux+cabal

cabal run yolov3 -- config/yolov3.cfg weights/yolov3.weights test-data/train.jpg out.png

linux+nix

nix-build
./result-2/bin/yolov3 config/yolov3.cfg weights/yolov3.weights test-data/train.jpg out.png

Test

The command to calcurate mAP is as follows. It supports both CPU and CUDA.

CPU

DEVICE=cpu cabal run yolov3-pipelined-test --enable-profiling -- config/yolov3.cfg weights/yolov3.weights ./coco.data  +RTS -p -hc -N3

DEVICE

DEVICE=cuda:0 cabal run yolov3-pipelined-test --enable-profiling -- config/yolov3.cfg weights/yolov3.weights ./coco.data  +RTS -p -hc -N3

Training

The command to train a model is as follows. It supports both CPU and CUDA.

CPU

DEVICE=cpu cabal run yolov3-training --enable-profiling -- config/yolov3.cfg weights/yolov3.weights ./coco.data  +RTS -p -hc -N3

DEVICE

DEVICE=cuda:0 cabal run yolov3-training --enable-profiling -- config/yolov3.cfg weights/yolov3.weights ./coco.data  +RTS -p -hc -N3