mangaocr_v2

Alan Chen
This project processes Manga109 into yolo/coco format, and trains yolov5 with it.

Use Ubuntu 20.04

Install latest Nvidia drivers

Install anaconda

Create conda env for this project

$conda create -n mangaocrV2 python=3.7 (or higher) anaconda
$conda activate mangaocrV2

Download yolov5 latest version

Download the source zip from the latest release and unzip it to ./yolov5 directory ...

Install pytorch with cuda support

$conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch

Install yolov5 dependencies

$conda install --file ./yolov5/requirements.txt
$pip install -r requirements.txt
$pip install wandb

Setup data

Download Manga109 and extract it to Manga109 directory ...
Create directories ./datasets/Manga109
Run manga109_to_yolov5.py
Copy Manga109.yaml to ./yolov5/data

Train yolov5

Dual GPU (DistributedDataParallel)

$python -m torch.distributed.run --nproc_per_node 2 train.py --img 1280 --batch 4 --epochs 300 --data Manga109.yaml --weights yolov5m6.pt --cfg yolov5m6.yaml

Single GPU (DataParallel)

$python train.py --img 1280 --workers 1 --batch 3 --epochs 300 --data Manga109.yaml --weights yolov5l6.pt --cfg yolov5l6.yaml

Test yolov5

$python val.py --weights runs/train/exp18(your path to the model)/weights/best.pt --data data/Manga109.yaml --task test --name manga109_best_1280 --imgsz 1280 --batch-size 2

Known issues:

OMP: Error #15: Initializing libiomp5md.dll, but found libiomp5md.dll already initialized. place this workaround in train.py: import os os.environ['KMP_DUPLICATE_LIB_OK']='True'