Jetson/L4T/TRT Customized Example

This page collects information to deploy customized models with TensorRT and some common questions for Jetson.

TensorRT Python

OpenCV with ONNX model

Below is an example to deploy TensorRT from an ONNX model with OpenCV images.

Verified environment:

JetPack4.5.1 + Xavier

import cv2
import time
import numpy as np
import tensorrt as trt
import pycuda.autoinit
import pycuda.driver as cuda

EXPLICIT_BATCH = 1 << (int)(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
TRT_LOGGER = trt.Logger(trt.Logger.INFO)
runtime = trt.Runtime(TRT_LOGGER)

host_inputs  = []
cuda_inputs  = []
host_outputs = []
cuda_outputs = []
bindings = []


def Inference(engine):
    image = cv2.imread("/usr/src/tensorrt/data/resnet50/airliner.ppm")
    image = (2.0 / 255.0) * image.transpose((2, 0, 1)) - 1.0

    np.copyto(host_inputs[0], image.ravel())
    stream = cuda.Stream()
    context = engine.create_execution_context()

    start_time = time.time()
    cuda.memcpy_htod_async(cuda_inputs[0], host_inputs[0], stream)
    context.execute_async(bindings=bindings, stream_handle=stream.handle)
    cuda.memcpy_dtoh_async(host_outputs[0], cuda_outputs[0], stream)
    stream.synchronize()
    print("execute times "+str(time.time()-start_time))

    output = host_outputs[0].reshape(np.concatenate(([1],engine.get_binding_shape(1))))
    print(np.argmax(output))


def PrepareEngine():
    with trt.Builder(TRT_LOGGER) as builder, builder.create_network(EXPLICIT_BATCH) as network, trt.OnnxParser(network, TRT_LOGGER) as parser:
        builder.max_workspace_size = 1 << 30
        with open('/usr/src/tensorrt/data/resnet50/ResNet50.onnx', 'rb') as model:
            if not parser.parse(model.read()):
                print ('ERROR: Failed to parse the ONNX file.')
                for error in range(parser.num_errors):
                    print (parser.get_error(error))
        engine = builder.build_cuda_engine(network)

        # create buffer
        for binding in engine:
            size = trt.volume(engine.get_binding_shape(binding)) * engine.max_batch_size
            host_mem = cuda.pagelocked_empty(shape=[size],dtype=np.float32)
            cuda_mem = cuda.mem_alloc(host_mem.nbytes)

            bindings.append(int(cuda_mem))
            if engine.binding_is_input(binding):
                host_inputs.append(host_mem)
                cuda_inputs.append(cuda_mem)
            else:
                host_outputs.append(host_mem)
                cuda_outputs.append(cuda_mem)

        return engine


if __name__ == "__main__":
    engine = PrepareEngine()
    Inference(engine)

OpenCV with PLAN model

Below is an example to deploy TensorRT from a TensorRT PLAN model with OpenCV images.

Verified environment:

JetPack4.5.1 + Xavier

$ /usr/src/tensorrt/bin/trtexec --onnx=/usr/src/tensorrt/data/resnet50/ResNet50.onnx --saveEngine=trt.plan

import cv2
import time
import numpy as np
import tensorrt as trt
import pycuda.autoinit
import pycuda.driver as cuda

EXPLICIT_BATCH = 1 << (int)(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
TRT_LOGGER = trt.Logger(trt.Logger.INFO)
runtime = trt.Runtime(TRT_LOGGER)

host_inputs  = []
cuda_inputs  = []
host_outputs = []
cuda_outputs = []
bindings = []


def Inference(engine):
    image = cv2.imread("/usr/src/tensorrt/data/resnet50/airliner.ppm")
    image = (2.0 / 255.0) * image.transpose((2, 0, 1)) - 1.0

    np.copyto(host_inputs[0], image.ravel())
    stream = cuda.Stream()
    context = engine.create_execution_context()

    start_time = time.time()
    cuda.memcpy_htod_async(cuda_inputs[0], host_inputs[0], stream)
    context.execute_async(bindings=bindings, stream_handle=stream.handle)
    cuda.memcpy_dtoh_async(host_outputs[0], cuda_outputs[0], stream)
    stream.synchronize()
    print("execute times "+str(time.time()-start_time))

    output = host_outputs[0].reshape(np.concatenate(([1],engine.get_binding_shape(1))))
    print(np.argmax(output))


def PrepareEngine():
    runtime = trt.Runtime(TRT_LOGGER)
    with open('./trt.plan', 'rb') as f:
        buf = f.read()
        engine = runtime.deserialize_cuda_engine(buf)

    # create buffer
    for binding in engine:
        size = trt.volume(engine.get_binding_shape(binding)) * engine.max_batch_size
        host_mem = cuda.pagelocked_empty(shape=[size],dtype=np.float32)
        cuda_mem = cuda.mem_alloc(host_mem.nbytes)

        bindings.append(int(cuda_mem))
        if engine.binding_is_input(binding):
            host_inputs.append(host_mem)
            cuda_inputs.append(cuda_mem)
        else:
            host_outputs.append(host_mem)
            cuda_outputs.append(cuda_mem)

    return engine


if __name__ == "__main__":
    engine = PrepareEngine()
    Inference(engine)

Multi-threading

Below is an example to run TensorRT with threads.

Verified environment:

JetPack4.5.1 + Xavier

$ /usr/src/tensorrt/bin/trtexec --onnx=/usr/src/tensorrt/data/mnist/mnist.onnx --saveEngine=mnist.trt
$ cd /usr/src/tensorrt/data/mnist/
$ sudo pip3 install pillow
$ python3 download_pgms.py
$ wget https://raw.githubusercontent.com/AastaNV/JEP/master/elinux/my_tensorrt_code.py -O my_tensorrt_code.py

import threading
import time
from my_tensorrt_code import TRTInference, trt

exitFlag = 0

class myThread(threading.Thread):
   def __init__(self, func, args):
      threading.Thread.__init__(self)
      self.func = func
      self.args = args
   def run(self):
      print ("Starting " + self.args[0])
      self.func(*self.args)
      print ("Exiting " + self.args[0])

if __name__ == '__main__':
    # Create new threads
    '''
    format thread:
        - func: function names, function that we wished to use
        - arguments: arguments that will be used for the func's arguments
    '''

    trt_engine_path = 'mnist.trt'

    max_batch_size = 1
    trt_inference_wrapper = TRTInference(trt_engine_path,
        trt_engine_datatype=trt.DataType.FLOAT,
        batch_size=max_batch_size)

    # Get TensorRT SSD model output
    input_img_path = '/usr/src/tensorrt/data/mnist/3.pgm'

    thread1 = myThread(trt_inference_wrapper.infer, [input_img_path])

    # Start new Threads
    thread1.start()
    thread1.join()
    trt_inference_wrapper.destory();
    print ("Exiting Main Thread")

Deepstream

YoloV4 Tiny

Verified environment:

JetPack4.5.1 + Xavier

Deepstream can reach 60fps with 4 video stream on Xavier:

$ cd /opt/nvidia/deepstream/deepstream-5.1/sources/objectDetector_Yolo
$ wget https://raw.githubusercontent.com/AastaNV/eLinux_data/main/deepstream/yolov4-tiny/yolov4_tiny.patch
$ git apply yolov4_tiny.patch
$ export CUDA_VER=10.2
$ make -C nvdsinfer_custom_impl_Yolo

$ wget https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov4-tiny.cfg -q --show-progress
$ wget https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.weights -q --show-progress
$ wget https://raw.githubusercontent.com/AastaNV/eLinux_data/main/deepstream/yolov4-tiny/deepstream_app_config_yoloV4_tiny.txt
$ wget https://raw.githubusercontent.com/AastaNV/eLinux_data/main/deepstream/yolov4-tiny/config_infer_primary_yoloV4_tiny.txt

$ deepstream-app -c deepstream_app_config_yoloV4_tiny.txt

Custom Parser for SSD-MobileNet Trained by Jetson-inference

Verified environment:

JetPack4.5.1 + Xavier

$ cd /opt/nvidia/deepstream/deepstream-5.1/sources/objectDetector_SSD/
$ sudo wget https://raw.githubusercontent.com/AastaNV/eLinux_data/main/deepstream/ssd-jetson_inference/ssd-jetson_inference.patch
$ sudo git apply ssd-jetson_inference.patch
$ sudo CUDA_VER=10.2 make -C nvdsinfer_custom_impl_ssd/

Update config_infer_primary_ssd.txt:

Ex.

diff --git a/config_infer_primary_ssd.txt b/config_infer_primary_ssd.txt
index e5bf468..81c52fd 100644
--- a/config_infer_primary_ssd.txt
+++ b/config_infer_primary_ssd.txt
@@ -62,15 +62,13 @@ gpu-id=0
 net-scale-factor=0.0078431372
 offsets=127.5;127.5;127.5
 model-color-format=0
-model-engine-file=sample_ssd_relu6.uff_b1_gpu0_fp32.engine
-labelfile-path=ssd_coco_labels.txt
-uff-file=sample_ssd_relu6.uff
+model-engine-file=ssd-mobilenet.uff_b1_gpu0_fp16.engine
+uff-file=ssd.uff
 infer-dims=3;300;300
 uff-input-order=0
 uff-input-blob-name=Input
-batch-size=1
-## 0=FP32, 1=INT8, 2=FP16 mode
-network-mode=0
+labelfile-path=labels.txt
+network-mode=2
 num-detected-classes=91
 interval=0
 gie-unique-id=1

$ deepstream-app -c deepstream_app_config_ssd.txt

VPI

VPI with Jetson-utils

Below is an example to use VPI with jetson-utils

Verified environment:

JetPack4.6 + XavierNX

import numpy as np
import jetson.utils
import vpi


display = jetson.utils.glDisplay()

camera = jetson.utils.gstCamera(1920, 1280, '0')
camera.Open()

while display.IsOpen():
    frame, width, height = camera.CaptureRGBA(zeroCopy=1)
    input = vpi.asimage(np.uint8(jetson.utils.cudaToNumpy(frame)))
    with vpi.Backend.CUDA:
        output = input.convert(vpi.Format.U8) 
        output = output.box_filter(11, border=vpi.Border.ZERO).convert(vpi.Format.RGB8)
        vpi.clear_cache()

    display.RenderOnce(jetson.utils.cudaFromNumpy(output.cpu()), width, height)
    display.SetTitle("{:s} | {:d}x{:d} | {:.1f} FPS".format("Camera Viewer", width, height, display.GetFPS()))

camera.Close()

VPI with Deepstream

Please find the following link for the example:

https://forums.developer.nvidia.com/t/deepstream-sdk-vpi-on-jetson-tx2/166834/20

VPI with Argus Camera

Please find the following link for the example:

https://forums.developer.nvidia.com/t/how-do-i-get-image-from-cudabayerdemosaic-and-connect-to-vpi/213529/18

Installation Steps

Darknet with cuDNN-8 Support

Below are the steps to build darknet with cuDNN-8 support.

Verified environment:

JetPack4.5.1 + Xavier

1. Get source

$ git clone https://github.com/pjreddie/darknet.git
$ cd darknet/
$ wget https://raw.githubusercontent.com/AastaNV/JEP/master/script/topics/0001-fix-for-cudnn_v8-limited-memory-to-default-darknet-s.patch
$ wget https://raw.githubusercontent.com/AastaNV/JEP/master/elinux/opencv-darknet.patch -O opencv-darknet.patch
$ git am 0001-fix-for-cudnn_v8-limited-memory-to-default-darknet-s.patch
$ git am opencv-darknet.patch

2. Update Makefile based on your device

GPU=1
CUDNN=1
OPENCV=1

Xavier & XavierNX:

ARCH= -gencode arch=compute_72,code=sm_72 \
      -gencode arch=compute_72,code=[sm_72,compute_72]

TX2:

ARCH= -gencode arch=compute_62,code=sm_62 \
      -gencode arch=compute_62,code=[sm_62,compute_62]

Nano:

ARCH= -gencode arch=compute_53,code=sm_53 \
      -gencode arch=compute_53,code=[sm_53,compute_53]

3. Build and Test

$ make -j8
$ wget https://pjreddie.com/media/files/yolov3-tiny.weights
$ ./darknet detector demo cfg/coco.data cfg/yolov3-tiny.cfg yolov3-tiny.weights [video]

TensorRT Python Bindings

Below are the steps to build TensorRT Python 3.9 bindings.

Verified environment:

JetPack4.6 + Xavier

1. Building python3.9

$ sudo apt install zlib1g-dev libncurses5-dev libgdbm-dev libnss3-dev libssl-dev libreadline-dev libffi-dev libsqlite3-dev libbz2-dev
$ wget https://www.python.org/ftp/python/3.9.1/Python-3.9.1.tar.xz
$ tar xvf Python-3.9.1.tar.xz Python-3.9.1/

$ mkdir build-python-3.9.1
$ cd build-python-3.9.1/
$ ../Python-3.9.1/configure --enable-optimizations
$ make -j $(nproc)
$ sudo -H make altinstall
$ cd ../

2. Build cmake 3.13.5

$ sudo apt-get install -y protobuf-compiler libprotobuf-dev openssl libssl-dev libcurl4-openssl-dev
$ wget https://github.com/Kitware/CMake/releases/download/v3.13.5/cmake-3.13.5.tar.gz
$ tar xvf cmake-3.13.5.tar.gz
$ rm cmake-3.13.5.tar.gz

$ cd cmake-3.13.5/
$ ./bootstrap --system-curl
$ make -j$(nproc)

$ echo 'export PATH='${PWD}'/bin/:$PATH' >> ~/.bashrc
$ source ~/.bashrc
$ cd ../

3. Prepare header

$ mkdir python3.9
$ mkdir python3.9/include
$ wget http://ftp.us.debian.org/debian/pool/main/p/python3.9/libpython3.9-dev_3.9.9-2_arm64.deb
$ ar x libpython3.9-dev_3.9.9-2_arm64.deb
$ tar -xvf data.tar.xz
$ cp ./usr/include/aarch64-linux-gnu/python3.9/pyconfig.h python3.9/include/
$ cp -r Python-3.9.1/Include/* python3.9/include/

4. Build TensorRT pybinding

$ git clone https://github.com/pybind/pybind11.git
$ git clone -b release/8.0 https://github.com/NVIDIA/TensorRT.git
$ cd TensorRT
$ git submodule update --init --recursive

$ cd python/
$ TRT_OSSPATH=${PWD}/.. EXT_PATH=${PWD}/../.. TARGET=aarch64 PYTHON_MINOR_VERSION=9 ./build.sh
$ python3.9 -m pip install build/dist/tensorrt-8.0.1.6-cp39-none-linux_aarch64.whl

Caffe

Below are the steps to build the Caffe library.

Verified environment:

JetPack4.6 + Xavier

$ wget https://raw.githubusercontent.com/AastaNV/JEP/master/elinux/install_caffe_jp46.sh -O install_caffe_jp46.sh
$ wget https://raw.githubusercontent.com/AastaNV/JEP/master/elinux/0001-patch-for-jp4.6.patch -O 0001-patch-for-jp4.6.patch
$ ./install_caffe_jp46.sh
$ source ~/.bashrc

MXNet

Below are the steps to build the MXNet 1.8.0 library.

Verified environment:

JetPack4.5.1 + Xavier

$ wget https://raw.githubusercontent.com/AastaNV/JEP/master/elinux/mxnet_v1.8.x.patch -O mxnet_v1.8.x.patch
$ wget https://raw.githubusercontent.com/AastaNV/JEP/master/elinux/autobuild_mxnet.sh -O autobuild_mxnet.sh
$ sudo chmod +x autobuild_mxnet.sh
$ ./autobuild_mxnet.sh Xavier
$ cd mxnet/build/
$ pip3 install mxnet-1.8.0-py3-none-any.whl

PyInstaller with OpenCV

Currently, the OpenCV version between JetPack default and Pyinstaller is not consistent.

To solve this issue, you can either upgrade the python-opencv version or downgrade the PyInstaller version.

Upgrade python-opencv

$ pip3 install opencv-python

Downgrade pyinstaller and pyinstaller-hooks-contrib

$ sudo pip3 install pyinstaller==4.2
$ sudo pip3 install pyinstaller-hooks-contrib==2021.2

$ pyinstaller --onefile --paths="/usr/lib/python3.6/dist-packages/cv2/python-3.6" myfile.py

Common Issues

"Unsupported ONNX data type: UINT8 (2)"

This error is from TensorRT. The root cause is that ONNX expects the input image to be INT8 but TensorRT uses Float32.

To solve this issue, you can modify the input data format of ONNX with our graphsurgeon API.

$ sudo apt-get install python3-pip libprotobuf-dev protobuf-compiler
$ git clone https://github.com/NVIDIA/TensorRT.git
$ cd TensorRT/tools/onnx-graphsurgeon/
$ make install

import onnx_graphsurgeon as gs
import onnx
import numpy as np

graph = gs.import_onnx(onnx.load("model.onnx"))
for inp in graph.inputs:
    inp.dtype = np.float32

onnx.save(gs.export_onnx(graph), "updated_model.onnx")

"Illegal instruction (core dumped)"

This is a known issue in NumPy v1.19.5.

To solve this issue, you can either downgrade your NumPy into 1.19.4 or manually update an environment variable.

Downgrade NumPy

$ sudo apt-get install python3-pip
$ pip3 install Cython
$ pip3 install numpy==1.19.4

Update environment variable

$ export OPENBLAS_CORETYPE=ARMV8

"Long delays when submitting several cudaMemcpy "

Please try to increase computing channel

$ export CUDA_DEVICE_MAX_CONNECTIONS=32

A document can be found here: https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars

Jetson/L4T/TRT Customized Example

Contents

TensorRT Python

OpenCV with ONNX model

OpenCV with PLAN model

Multi-threading

Deepstream

YoloV4 Tiny

Custom Parser for SSD-MobileNet Trained by Jetson-inference

VPI

VPI with Jetson-utils

VPI with Deepstream

VPI with Argus Camera

Installation Steps

Darknet with cuDNN-8 Support

TensorRT Python Bindings

Caffe

MXNet

PyInstaller with OpenCV

Common Issues

"Unsupported ONNX data type: UINT8 (2)"

"Illegal instruction (core dumped)"

"Long delays when submitting several cudaMemcpy "

Navigation menu

Personal tools

Namespaces

Variants

Views

More

Search

Navigation

Where else to find us

Tools