The default version of Jepack 4.5.1 after brushing is cuda 10.2, but for project requirements, cuda 10.2 should be used 0. Online materials say that this version can't install the lower version of cuda10 0. I tried and didn't install it. I couldn't install it. The official website also said no. The leader said don't brush cuda10 when brushing the machine 2 and cudnn, opencv and python. Try again. If you can't do it again, you really can't install it. I found a lot of information, tried various settings, experienced several hours, and finally succeeded.
cuda10.0 and cudnn7 network disk download address:
Link: https://pan.baidu.com/s/1jBLv4M38fiKX9VMZJtPAew
Extraction code: BwP5
The file is 5.1G in total, including jepack 4 2 SDK, don't worry about it. You can download the files you need.
Now start TX2 I jepack4 5.1 install cuda10 0
Prompt: before installing, determine whether to install with root or user. You can choose the default. Take me for example. When I want to delete useless space as much as possible after installation, if I install it as root, I can delete everything under home. If I install it as user, I can't delete it. I didn't know until later.

sudo dpkg -i cuda-repo-l4t-10-0-local-10.0.166_1.0-1_arm64.deb   #Install cuda10 0
sudo dpkg -i graphsurgeon-tf_5.0.6-1+cuda10.0_arm64.deb

sudo dpkg -i libcudnn7_7.3.1.28-1+cuda10.0_arm64.deb
sudo dpkg -i libcudnn7-dev_7.3.1.28-1+cuda10.0_arm64.deb

Install some more dependencies. The current status is as follows. Other dependencies need to be reinstalled.
Go to * / var/cuda-repo-10-0-local-10.0.166*

sudo dpkg -i cuda-license-10-0_10.0.166-1_arm64.deb
sudo dpkg -i cuda-cublas-10-0_10.0.166-1_arm64.deb
sudo dpkg -i cuda-cublas-dev-10-0_10.0.166-1_arm64.deb

sudo apt-get install cuda-toolkit-8-0
I implemented all of the following, but later I wondered if it was OK not to install python2. So you can try not to install it
cd to your download installation package path:

sudo dpkg -i libnvinfer5_5.0.6-1+cuda10.0_arm64.deb
sudo dpkg -i libnvinfer-dev_5.0.6-1+cuda10.0_arm64.deb
sudo dpkg -i libnvinfer-samples_5.0.6-1+cuda10.0_all.deb # Personally, I think it's OK not to install it, but I haven't tried it. If I have the opportunity to install it again, will I try it? Will there be any mistakes if I don't install it
sudo dpkg -i tensorrt_5.0.6.3-1+cuda10.0_arm64.deb
sudo dpkg -i uff-converter-tf_5.0.6-1+cuda10.0_arm64.deb
sudo dpkg -i python-libnvinfer_5.0.6-1+cuda10.0_arm64.deb #This will also try not to install,
sudo dpkg -i python-libnvinfer-dev_5.0.6-1+cuda10.0_arm64.deb #This one will also try not to install, but this one should be synchronized with the above one. If you want to i install it, don't install it.
sudo dpkg -i python3-libnvinfer_5.0.6-1+cuda10.0_arm64.deb
sudo dpkg -i python3-libnvinfer-dev_5.0.6-1+cuda10.0_arm64.deb

Fix dependencies again:

sudo apt --fix-broken install
sudo apt-get install libtbb2

The following three instructions are installed in opencv, but not in the installation package path of python 3. I want to save space, so I don't want to install this version. I want to change to other controllable opencv versions, but python can call this 3.3. I don't know if there will be an error if you don't install it, but you can try it and install it again if you report an error.
libopencv_3.3.1-2-g31ccdfe11_arm64.deb
libopencv-dev_3.3.1-2-g31ccdfe11_arm64.deb
libopencv-python_3.3.1-2-g31ccdfe11_arm64.deb

After installing the above process, it is found that cuda is wrong and there are no corresponding lib and include files:
Execute the following instructions

sudo apt-get install cuda-toolkit-10-0