Skip to content

PyTorch ,ONNX and TensorRT implementation of YOLOv4

License

Notifications You must be signed in to change notification settings

blzheng/pytorch-YOLOv4

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Pytorch-YOLOv4

A minimal PyTorch implementation of YOLOv4.

├── README.md
├── dataset.py            dataset
├── demo.py               demo to run pytorch --> tool/darknet2pytorch
├── demo_darknet2onnx.py  tool to convert into onnx --> tool/darknet2pytorch
├── demo_pytorch2onnx.py  tool to convert into onnx
├── models.py             model for pytorch
├── train.py              train models.py
├── cfg.py                cfg.py for train
├── cfg                   cfg --> darknet2pytorch
├── data            
├── weight                --> darknet2pytorch
├── tool
│   ├── camera.py           a demo camera
│   ├── coco_annotation.py       coco dataset generator
│   ├── config.py
│   ├── darknet2pytorch.py
│   ├── region_loss.py
│   ├── utils.py
│   └── yolo_layer.py

image

0. Weights Download

0.1 darknet

0.2 pytorch

you can use darknet2pytorch to convert it yourself, or download my converted model.

1. Train

use yolov4 to train your own data

  1. Download weight

  2. Transform data

    For coco dataset,you can use tool/coco_annotation.py.

    # train.txt
    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 ...
    ...
    ...
    
  3. Train

    you can set parameters in cfg.py.

     python train.py -g [GPU_ID] -dir [Dataset direction] ...
    

2. Inference

2.1 Performance on MS COCO dataset (using pretrained DarknetWeights from https://github.com/AlexeyAB/darknet)

ONNX and TensorRT models are converted from Pytorch (TianXiaomo): Pytorch->ONNX->TensorRT. See following sections for more details of conversions.

  • val2017 dataset (input size: 416x416)
Model type AP AP50 AP75 APS APM APL
DarkNet (YOLOv4 paper) 0.471 0.710 0.510 0.278 0.525 0.636
Pytorch (TianXiaomo) 0.466 0.704 0.505 0.267 0.524 0.629
TensorRT FP32 + BatchedNMSPlugin 0.472 0.708 0.511 0.273 0.530 0.637
TensorRT FP16 + BatchedNMSPlugin 0.472 0.708 0.511 0.273 0.530 0.636
  • testdev2017 dataset (input size: 416x416)
Model type AP AP50 AP75 APS APM APL
DarkNet (YOLOv4 paper) 0.412 0.628 0.443 0.204 0.444 0.560
Pytorch (TianXiaomo) 0.404 0.615 0.436 0.196 0.438 0.552
TensorRT FP32 + BatchedNMSPlugin 0.412 0.625 0.445 0.200 0.446 0.564
TensorRT FP16 + BatchedNMSPlugin 0.412 0.625 0.445 0.200 0.446 0.563

2.2 Image input size for inference

Image input size is NOT restricted in 320 * 320, 416 * 416, 512 * 512 and 608 * 608. You can adjust your input sizes for a different input ratio, for example: 320 * 608. Larger input size could help detect smaller targets, but may be slower and GPU memory exhausting.

height = 320 + 96 * n, n in {0, 1, 2, 3, ...}
width  = 320 + 96 * m, m in {0, 1, 2, 3, ...}

2.3 Different inference options

  • Load the pretrained darknet model and darknet weights to do the inference (image size is configured in cfg file already)

    python demo.py -cfgfile <cfgFile> -weightfile <weightFile> -imgfile <imgFile>
  • Load pytorch weights (pth file) to do the inference

    python models.py <num_classes> <weightfile> <imgfile> <IN_IMAGE_H> <IN_IMAGE_W> <namefile(optional)>
  • Load converted ONNX file to do inference (See section 3 and 4)

  • Load converted TensorRT engine file to do inference (See section 5)

2.4 Inference output

There are 2 inference outputs.

  • One is locations of bounding boxes, its shape is [batch, num_boxes, 1, 4] which represents x1, y1, x2, y2 of each bounding box.
  • The other one is scores of bounding boxes which is of shape [batch, num_boxes, num_classes] indicating scores of all classes for each bounding box.

Until now, still a small piece of post-processing including NMS is required. We are trying to minimize time and complexity of post-processing.

3. Darknet2ONNX

  • This script is to convert the official pretrained darknet model into ONNX

  • Pytorch version Recommended:

    • Pytorch 1.4.0 for TensorRT 7.0 and higher
    • Pytorch 1.5.0 and 1.6.0 for TensorRT 7.1.2 and higher
  • Install onnxruntime

    pip install onnxruntime
  • Run python script to generate ONNX model and run the demo

    python demo_darknet2onnx.py <cfgFile> <weightFile> <imageFile> <batchSize>

3.1 Dynamic or static batch size

  • Positive batch size will generate ONNX model of static batch size, otherwise, batch size will be dynamic
    • Dynamic batch size will generate only one ONNX model
    • Static batch size will generate 2 ONNX models, one is for running the demo (batch_size=1)

4. Pytorch2ONNX

  • You can convert your trained pytorch model into ONNX using this script

  • Pytorch version Recommended:

    • Pytorch 1.4.0 for TensorRT 7.0 and higher
    • Pytorch 1.5.0 and 1.6.0 for TensorRT 7.1.2 and higher
  • Install onnxruntime

    pip install onnxruntime
  • Run python script to generate ONNX model and run the demo

    python demo_pytorch2onnx.py <weight_file> <image_path> <batch_size> <n_classes> <IN_IMAGE_H> <IN_IMAGE_W>

    For example:

    python demo_pytorch2onnx.py yolov4.pth dog.jpg 8 80 416 416

4.1 Dynamic or static batch size

  • Positive batch size will generate ONNX model of static batch size, otherwise, batch size will be dynamic
    • Dynamic batch size will generate only one ONNX model
    • Static batch size will generate 2 ONNX models, one is for running the demo (batch_size=1)

5. ONNX2TensorRT

  • TensorRT version Recommended: 7.0, 7.1

5.1 Convert from ONNX of static Batch size

  • Run the following command to convert YOLOv4 ONNX model into TensorRT engine

    trtexec --onnx=<onnx_file> --explicitBatch --saveEngine=<tensorRT_engine_file> --workspace=<size_in_megabytes> --fp16
    • Note: If you want to use int8 mode in conversion, extra int8 calibration is needed.

5.2 Convert from ONNX of dynamic Batch size

  • Run the following command to convert YOLOv4 ONNX model into TensorRT engine

    trtexec --onnx=<onnx_file> \
    --minShapes=input:<shape_of_min_batch> --optShapes=input:<shape_of_opt_batch> --maxShapes=input:<shape_of_max_batch> \
    --workspace=<size_in_megabytes> --saveEngine=<engine_file> --fp16
  • For example:

    trtexec --onnx=yolov4_-1_3_320_512_dynamic.onnx \
    --minShapes=input:1x3x320x512 --optShapes=input:4x3x320x512 --maxShapes=input:8x3x320x512 \
    --workspace=2048 --saveEngine=yolov4_-1_3_320_512_dynamic.engine --fp16

5.3 Run the demo

python demo_trt.py <tensorRT_engine_file> <input_image> <input_H> <input_W>
  • This demo here only works when batchSize is dynamic (1 should be within dynamic range) or batchSize=1, but you can update this demo a little for other dynamic or static batch sizes.

  • Note1: input_H and input_W should agree with the input size in the original ONNX file.

  • Note2: extra NMS operations are needed for the tensorRT output. This demo uses python NMS code from tool/utils.py.

6. ONNX2Tensorflow

7. ONNX2TensorRT and DeepStream Inference

  1. Compile the DeepStream Nvinfer Plugin
    cd DeepStream
    make 
  1. Build a TRT Engine.

For single batch,

trtexec --onnx=<onnx_file> --explicitBatch --saveEngine=<tensorRT_engine_file> --workspace=<size_in_megabytes> --fp16

For multi-batch,

trtexec --onnx=<onnx_file> --explicitBatch --shapes=input:Xx3xHxW --optShapes=input:Xx3xHxW --maxShapes=input:Xx3xHxW --minShape=input:1x3xHxW --saveEngine=<tensorRT_engine_file> --fp16

Note :The maxShapes could not be larger than model original shape.

  1. Write the deepstream config file for the TRT Engine.

Reference:

@article{yolov4,
  title={YOLOv4: YOLOv4: Optimal Speed and Accuracy of Object Detection},
  author={Alexey Bochkovskiy, Chien-Yao Wang, Hong-Yuan Mark Liao},
  journal = {arXiv},
  year={2020}
}

About

PyTorch ,ONNX and TensorRT implementation of YOLOv4

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 74.5%
  • C++ 21.4%
  • Shell 2.6%
  • Other 1.5%