librazafu / pytorch-darknet53

PyTorch implementation of Darknet53

Darknet53

This is implementation of Darknet53 network discussed in [1] used for feature extractor of YOLOv3.

This new network is more efficient than ResNet-101 or ResNet-152.

Here are some ImageNet results:

• Framework: Darknet [2]
• GPU: Titan X
• Input Shape(CWH): 3 x 256 x 256

Darknet-53 is better than ResNet-101 and 1.5× faster.

Darknet-53 has similar performance to ResNet-152 and is 2× faster [1] .

But when I trained and tested this model with 224x224 input image, I could not get the good results like the above table.

I got 75.xx% accracy on Imagenet validation dataset.

Pretrained model_weight_download

Network Structure

Training

• Download the ImageNet dataset and move validation images to labeled subfolders
  • To do this, you can use the following script
• imagenet data is processed as described here

python train.py --data [imagenet-folder with train and val folders] --gpu 0 -b 64

Benchmark

• Framework: PyTorch
• GPU: GTX 1080 Ti 11GB
• CPU: i7 6550 3.4 GHZ
• RAM: 16 GB
• Batch Size: 1
• Input Shape(CWH): 3 x 224 x 224

On GPU

resnet101 : 0.034906 sec
resnet152 : 0.055852 sec
densenet121 : 0.041888 sec
darknet53 : 0.017952 sec

On CPU

resnet101 : 0.675194 sec
resnet152 : 0.949459 sec
densenet121 : 0.649266 sec
darknet53 : 0.405916 sec

Reference

[1] YOLOv3: An Incremental Improvement

[2] darknet framework

[3] ImageNet training in PyTorch