# DDRNet.TensorRT

**Repository Path**: leeyueleeyue/DDRNet.TensorRT

## Basic Information

- **Project Name**: DDRNet.TensorRT
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2025-06-23
- **Last Updated**: 2025-06-23

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# DDRNet

TensorRT implementation of the official [DDRNet](https://github.com/ydhongHIT/DDRNet)

<p align="center">
<img src="./results/result_mainz_000001_009328_leftImg8bit.png">
</p>

[DDRNet-23-slim](https://paperswithcode.com/paper/deep-dual-resolution-networks-for-real-time) outperform other light weight segmentation method,[see](https://paperswithcode.com/sota/real-time-semantic-segmentation-on-cityscapes)
<p align="center">
<img src="./results/Screenshot from 2021-04-21 19-25-48.png">
</p>



## Compile&Run

* 1. get model.wts
 
  Convert pytorch model to wts model using getwts.py, or download the wts [model](url: https://pan.baidu.com/s/1Cm1A2mq6RxCFhUJrOJBSrw  ;passworld: p6hy ) convert from official implementation.

  note  that we do not use extral segmentation head while inference(set augment=False in https://github.com/ydhongHIT/DDRNet/blob/76a875084afdc7dedd20e2c2bdc0a93f8f481e81/segmentation/DDRNet_23_slim.py#L345).

* 2. cmake and make

     config ur cmakelist and
  
  ```
  mkdir build
  cd build
  cmake ..
  make -j8
  ./ddrnet -s             // serialize model to plan file i.e. 'DDRNet.engine'
  ./ddrnet -d  ../images // deserialize plan file and run inference, the images in samples will be processed.
  ```

for INT8 support:

```
#define USE_INT8  // comment out this if want to use INT8
//#define USE_FP16  // comment out this if want to use FP32
```

mkdir "calib" and put around 1k images(cityscape val/test images) into folder "calib".

## FPS

Test on RTX2070

| model          | input           | FPS  |
| -------------- | --------------- | ---- |
| Pytorch-aug    | （3,1024,1024） | 107  |
| Pytorch-no-aug | （3,1024,1024） | 108  |
| TensorRT-FP32  | （3,1024,1024） | 117  |
| TensorRT-FP16  | （3,1024,1024） | 215  |
| TensorRT-INT8  | （3,1024,1024） | 232  |

Pytorch-aug means augment=True.

## Difference with official

we use Upsample with "nearest" other than "bilinear",which may lead to lower accuracy .

Finetune with "nearest" upsample may recover the accuracy.

Here we convert from the official model directly.

## Train 

1. refer to:https://github.com/chenjun2hao/DDRNet.pytorch
2. generate wts model with getwts.py

## Train customer data
https://github.com/midasklr/DDRNet.Pytorch
wirte your own dataset and finetune the model with cityscape.