Yolov3 on TensorRT7.0 and Tensorflow2.0

This repository contains Yolov3 inference on tensorrt7.0 and tensorflow 2.0. Model is based on darknet, adapt it to keras and tensorrt platform.

Test environments

ubuntu 16.04
Tensorflow 1.4,1.5 and 2.0
TensorRT 7.0
Nvidia-driver 410.78
cuda 10.0
cudnn 7.6.5
python 3.5.2

optional:

keras2onnx 1.6
onnx 1.6

Models

• Keras model
  Keras model is borrowed from keras-yolo3, which contains a detailed description about how to generate .h5 model.
• TensorRT engine
  TensorRT engine are generated based on Keras model. In this case, Keras model is converted to onnx format and then used to generate TensorRT engine.

Keras model can be converted to onnx as:

python3 model_data/convert_model.py --model_path=your_dir/model.h5 --output_path=output_dir --type=onnx

then specify model path in yolo_tensorrt.py or yolo_test.py, TensorRT engine will be generated after running yolo_test.py.

For int8 engine, calibrate images should be prepared and specify images path in yolo_tensorrt.py

• Download engine
  Or Download engine directly(waiting for upload).

Run test

python3 yolo_test.py --model_path=model_data/yolo_int8.engine --live --platform=tensorrt

more detail refer to :

python3 yolo_test.py --help

Evaluate result

int8 calibration images are 1000 pics selected in val2014

As shown above, fp32 model has completely same result as raw model in twice tests. In my previous tensorrt 6.0.1 practice, the fp32 model has little less mAP than raw model, but the model is converted through uff.