shbaydadaev / yolov5-on-nvidia-jetson

Yolov5 TensorRT Conversion & Deployment on Jetson Nano & TX2 & Xavier [Ultralytics EXPORT]

Notes

This repository is to deploy Yolov5 to NVIDIA Jetson platform and accelerate it through TensorRT technology. Available devices such as Nano (472 GFLOPS), TX2 (1.33 TFLOPS), Xavier NX (21 TOPS) and AGX Xavier (32TOPS).

JetPack

Firstly you should get latest JetPack v4.5.1 from NVIDIA and boot it onto Jetson Nano. You could find JetPack download options here

Conversion Steps

1. Clone this repository and Enter it:

   git clone https://gitee.com/jario-jin/yolov5-on-nvidia-jetson.git
   cd yolov5-on-nvidia-jetson

2. Install requirements for JetPack:

   bash scripts/install_requirements.sh

3. Convert the model:

   bash scripts/convert_models.sh

4. Test detection speed and accuracy:

   wget http://images.cocodataset.org/zips/val2017.zip
   # Download instances_val2017.json
   gdown https://drive.google.com/uc?id=13x1Nc7Jpjz-dDtrAP2A8Jf2LXWscV0q6
   unzip val2017.zip
   # Inference with Python front-end
   python3 inference.py --input "val2017/*.jpg" --gt instances_val2017.json
   # Inference with C++ front-end
   bash scripts/cpp_tensorrt_coco_eval.sh

5. Test with a webcam (id=0):

   python3 inference.py --webcam

Accuracy and speed test results on COCO val

Device	size	FP16/INT8	Front-end	mAPval 0.5:0.95	mAPsmall	Avg.Infer.Time
Nano	640	FP16	Python	33.3	17.3	105.8 ms (+- 13.5 ms)
Nano	640	FP16	C++	33.7	17.6	87.0 ms (+- 4.2 ms)
Nano	320	FP16	Python	28.1	7.7	38.1 ms (+- 3.3 ms)
Nano	320	FP16	C++	28.1	7.3	32.9 ms (+- 4.1 ms)
Xavier NX	640	FP16	Python	33.3	17.3	37.5 ms (+- 4.5 ms)
Xavier NX	640	FP16	C++	33.7	17.6	25.5 ms (+- 2.3 ms)
Xavier NX	320	FP16	Python	28.2	7.8	26.2 ms (+- 1.8 ms)
Xavier NX	320	FP16	C++	28.1	7.1	15.1 ms (+- 1.7 ms)
AGX Xavier	640	FP16	Python	33.3	17.3	30.2 ms (+- 3.3 ms)
AGX Xavier	640	FP16	C++	33.7	17.6	18.5 ms (+- 2.1 ms)
AGX Xavier	320	FP16	Python	28.1	7.8	23.5 ms (+- 2.5 ms)
AGX Xavier	320	FP16	C++	28.1	7.4	10.1 ms (+- 3.2 ms)
AGX Xavier	640	INT8	Python	33.2	17.3	27.7 ms (+- 2.0 ms)
AGX Xavier	640	INT8	C++	14.5	9.5	15.0 ms (+- 2.7 ms)

NOTE

The libmyplugins.so file in build will be needed for inference.

Custom model

For custom model conversion there are some factors to take in consideration.

• Only YoloV5 S (small) version is supported.
• You should use your own checkpoint that only contains network weights (i.e. stripped optimizer, which is last output of YoloV5 pipeline after training finishes)
• Change the CLASS_NUM in yololayer.h - Line 28 to number of classes your model has before building yolo. If you've already built it, you can just run cmake .. && make from build folder after changing the class numbers.
• Change the Input W and Input H according to resolution you've trained the network in yololayer.h on - Lines 29, 30
• Change the CONF and NMS thresholds according to your preferences in yolov5.cpp - Lines 12, 13
• Change the batch size according to your requirements in yolov5.cpp - Line 14.

INT 8 Conversion & Calibration

Exporting YoloV5 network to INT8 is pretty much straightforward & easy.

Before you run last command in Step 5 in conversion steps, you should take in consideration that INT8 needs dataset for calibration. Good news is - we won't need labels, just images. There's no recommended amount of data samples to use for calibration, but as many - as better.

In this case, as long as we're exporting the standard yolov5s, trained on COCO dataset, we'll download val set of images from coco and do calibration on it.

1. Enter the tensorrt_yolov5 folder and run.

wget http://images.cocodataset.org/zips/val2017.zip
unzip val2017.zip
mv val2017 build/coco_calib

2. Change the precision in yolov5.cpp - Line 10

#define USE_FP16 -> #define USE_INT8

3. Run the conversion.

cmake .. && make
sudo ./yolov5 -s yolov5s.wts yolov5s_int8.engine s

It'll do calibration on every image in coco_calib folder, it might take a while. After it's finished, the engine is ready for usage.

Thanks to the following sources/repositories for scripts & helpful suggestions.

• For TensorRT conversion codebase: https://github.com/wang-xinyu/tensorrtx
• For Torch installation procedures: https://qengineering.eu/install-pytorch-on-jetson-nano.html
• For Swap memory management: https://github.com/JetsonHacksNano/installSwapfile