darknet-pytorch

Implementation of DarkNet19, DarkNet53, CSPDarkNet53 with PyTorch

Table of Contents

• Project Description
• Installation
• Usage
• Contributing
• License

Project Description

1. DarkNet19 - used as a feature extractor in YOLO900.
2. DarkNet53 - used as a feature extractor in YOLOv3.
3. CSPDarkNet53 - Implementation of Cross Stage Partial Networks in DarkNet53.
4. DarkNet53-Elastic - Implementation of ELASTIC with DarkNet53.
5. CSPDarkNet53-Elastic - Implementation of CSP and ELASTIC in DarkNet53.

Architecture of DarkNet19 and DarkNet53:

Installation

 git clone https://github.com/yakhyo/darknet-pytorch.git
 cd darknet-pytorch
 python main.py path/to/imagenet --batch-size 512 --workers 8

Continue from checkpoint:

python main.py path/to/imagenet --batch-size 512 --workers 8 --resume darknet53.pth.tar

Results:

	This Repo.			Official
Model	Acc@1	Acc@5	Params	Acc@1	Acc@5
DarkNet19	70.5	89.7	21M	74.1	91.8
DarkNet53	75.6	92.5	41M	77.2	93.8
CSP-DarkNet53	74.3	92.2	19M	77.2	93.6
DarkNet53-Elastic	70.8	90.2	24M	...	...
CSPDarkNet53-Elastic	...	...	...	76.1	93.3

Weights of DarkNet53 (105th epoch), DarkNet19 (50th epoch), CSPDarkNet53 (80th epoch) and DarkNet53 ELASTIC ( 57th epoch) are available on here.

Trained on ImageNet

• GPU: Tesla V100
• Input size: 3x224x224

Dataset structure:

├── IMAGENET 
    ├── train
         ├── [class_id1]/xxx.{jpg,png,jpeg}
         ├── [class_id2]/xxy.{jpg,png,jpeg}
         ├── [class_id3]/xxz.{jpg,png,jpeg}
          ....
    ├── val
         ├── [class_id1]/xxx1.{jpg,png,jpeg}
         ├── [class_id2]/xxy2.{jpg,png,jpeg}
         ├── [class_id3]/xxz3.{jpg,png,jpeg}

Usage

from model import darknet19, darknet53, darknet53e, cspdarknet53

# darknet19
model = darknet19(num_classes=1000, init_weight=True)

# darknet53
model = darknet53(num_classes=1000, init_weight=True)

# darknet53 elastic
model = darknet53e(num_classes=1000, init_weight=True)

# cspdarknet53
model = cspdarknet53(num_classes=1000, init_weight=True)

Contributing

If you find any issues within this code, feel free to create PR or issue.

License

The project is licensed under the MIT license.