pytorch-YOLOv4
A minimal PyTorch implementation of YOLOv4.
This Rep forked from Tianxiaomo/pytorch-YOLOv4. See Original_README.
Hardware
- Ubuntu 18.04 LTS
- Intel(R) Core(TM) i7-6900K CPU @ 3.20GHz
- 31 RAM
- NVIDIA RTX 1080 8G * 4
Reproducing Submission
To reproduct my submission without retrainig, do the following steps:
Pytorch Weights Download
- google (provided by Tianxiaomo/pytorch-YOLOv4)
- yolov4.pth(https://drive.google.com/open?id=1wv_LiFeCRYwtpkqREPeI13-gPELBDwuJ)
- yolov4.conv.137.pth(https://drive.google.com/open?id=1fcbR0bWzYfIEdLJPzOsn4R5mlvR6IQyA)
Dataset Preparation
All required files except images are already in data directory. If you generate CSV files (duplicate image list, split, leak.. ), original files are overwritten. The contents will be changed, but It's not a problem.
Prepare Images
After downloading images, the data directory is structured as:
train.txt
+- data/
| +- train/
| +- test/
| +- training_labels.csv
| +- val.txt
Download Classes Image
Smaill SVHN Dataset: https://drive.google.com/drive/u/1/folders/1Ob5oT9Lcmz7g5mVOcYH3QugA7tV3WsSl
Download and extract tain.tar.gz and test.tar.gz to data directory.
Transform data
Use construct_datasets.py to make train.txt .
# train.txt and val.txt
# left(x1) top(y1) right(x2) bottom(y2) label
image_path1 x1,y1,x2,y2,id x1,y1,x2,y2,id x1,y1,x2,y2,id ...
image_path2 x1,y1,x2,y2,id x1,y1,x2,y2,id x1,y1,x2,y2,id ...
...
Names file example is in data/SVHN.names
# names file
Label1
Label2
Label3
...
Training
Setting
You can setting bach size and epoch in cfg.py
Train models
To train models, run following commands.
$ python3 train.py -d data/ -classes 10 -g 0 -pretrained ./weight/yolov4.conv.137.pth
The expected training times are:
| Model | GPUs | Image size | Training Epochs | Training Time | Bach Size |
|---|---|---|---|---|---|
| YOLOv4 | 1x NVIDIA T4 | 608x608 | 1 | 2.5 hours | 4 |
| YOLOv4 | 4x NVIDIA GTX 1080 | 608x608 | 1 | 0.6 hour | 32 |
Muti-GPU Training
$ python3 train.py -d data/ -classes 10 -g 0,1,2,3 -pretrained ./weight/yolov4.conv.137.pth
Inference
Inference single images
$ python3 models.py 10 Yolov4_epoch10.pth data/test/1.png 608 608 data/SVHN.names
Inference images in folder
$ python3 models_mut.py 10 Yolov4_epoch22_pre.pth data/test/ 608 608 data/SVHN.names
Result
mAP: 0.51742 90ms per image
Reference:
- Tianxiaomo/pytorch-YOLOv4
- https://github.com/eriklindernoren/PyTorch-YOLOv3
- https://github.com/marvis/pytorch-caffe-darknet-convert
- https://github.com/marvis/pytorch-yolo3
- Paper Yolo v4: https://arxiv.org/abs/2004.10934
- Source code:https://github.com/AlexeyAB/darknet
- More details: http://pjreddie.com/darknet/yolo/