【2021年】NVIDIA Jetsonで JetPack 4.5環境で OpenPoseよりも軽量で爆速な trt_pose_handを動かす方法
|
|
|
|
【2021年】NVIDIA Jetsonで JetPack 4.5環境で OpenPoseよりも軽量で爆速な trt_pose_handを動かす方法| Xavier NX | Nano | |
| 性能比 | 6～30倍 | 1 |
| AI性能比 | 21 TOPs | 472 GFLOPs |
| メモリ | 8GB | 4GB |
| M.2 SSD | 対応 | 非対応 |
| 電源 | 付属 | 別売り |
| ファン | 内蔵 | 別売り |
| SDカード | 別売り | 別売り |
| ケース | 別売り | 別売り |
| 消費電力 | 10W / 15W | 5W / 10W |
 NVIDIA Jetson Xavier NX Developer Kit 開発キット クラウドネイティブ AIコンピュータ ディープラーニング ASIN: B088TTD1QT ※ 低価格のハイエンドモデル |
 NVIDIA Jetson Nano Development Kit B01 開発キット AI コンピュータ 人工知能 エッジコンピューティング ASIN: B085NQTKS9 ※ 入門用の低価格 4GB版 |
 GeeekPi Jetson Nanoケース（Jetson Nano B01およびA02バージョンをサポート）NVIDIA Jetson Nano開発者キットに適したファン付きJetson NanoケースSmall AI Powerful Computer ASIN: B085PZ4F3X |
 SanDisk ( サンディスク ) 128GB microSD Extreme PRO microSDXC A2 SDSQXCY-128G-GN6MA ［ 海外パッケージ品 ］ ASIN: B07G3H5RBT ※ 高速、大容量 SDカード |
# ====
# Step 1 - Install Dependencies
# Install PyTorch and Torchvision. To do this on NVIDIA Jetson, we recommend following this guide
# ====
# 1-1
# Pytorch v1.7.0 / torchvision v0.8.1 / Python 3.6
cd
bash ./Jetson_Convenience_Script/PyTorch/inst_PyTorch_v1_7_Python3.sh
# ====
cd
git clone https://github.com/NVIDIA-AI-IOT/trt_pose_hand --depth 1
cd trt_pose_hand
pip3 install traitlets
# ModuleNotFoundError: No module named 'sklearn'
# Use scikit-learn instead
# sudo pip3 install sklearn
# https://pypi.org/project/scikit-learn/
# scikit-learn 0.24.1, Requires: Python >=3.6
sudo pip3 install scikit-learn==0.24.1
# Successfully installed joblib-1.0.0 scikit-learn-0.24.1 threadpoolctl-2.1.0
# sudo pip3 install scipy==1.5.4
# numpy.distutils.system_info.NotFoundError: No lapack/blas resources found.
sudo apt-get install -y libatlas-base-dev gfortran
# ModuleNotFoundError: No module named 'numpy.testing.nosetester' #1
# https://github.com/NVIDIA-AI-IOT/trt_pose_hand/issues/1
# sudo pip uninstall numpy
# sudo pip3 install numpy==1.16.4
# https://pypi.org/project/numpy/
# numpy 1.20.1, Requires: Python >=3.7
# sudo pip3 install numpy==1.18
# https://pypi.org/project/scipy/
# scipy 1.6.0, Requires: Python >=3.7
sudo pip3 install scipy==1.5.4
# sudo pip3 install scikit-learn==0.21.3
# pandas 1.1.5, Requires: Python >=3.6.1
# pip install pandas==1.1.5
# from numpy.testing import rundocs
# Model
# hand_pose_resnet18_att_244_244.pth
# https://drive.google.com/open?id=1NCVo0FiooWccDzY7hCc5MAKaoUpts3mo
FILE_ID=1NCVo0FiooWccDzY7hCc5MAKaoUpts3mo
FILE_NAME=hand_pose_resnet18_att_244_244.pth
wget "https://drive.google.com/uc?export=download&id=${FILE_ID}" -O ${FILE_NAME}
mv ${FILE_NAME} ./model/
# ====
# Jupyter Notebook
# https://jupyter.org/
sudo apt-get install -y libbz2-dev libsqlite3-dev libffi-dev
# ====
# notebook 6.2.0 Released: Jan 14, 2021
# https://pypi.org/project/notebook/
# Jupyter Notebook
sudo -H pip3 install notebook
# ====
# Enable Jupyter Widgets UI Library ipywidgets
# ipywidgets: Interactive HTML Widgets
# https://github.com/jupyter-widgets/ipywidgets
# ImportError: IProgress not found. Please update jupyter and ipywidgets.
# See https://ipywidgets.readthedocs.io/en/stable/user_install.html
sudo pip3 install ipywidgets
# Successfully installed ipywidgets-7.6.3 jupyterlab-widgets-1.0.0 widgetsnbextension-3.5.1
# can be skipped for notebook version 5.3 and above
# sudo jupyter nbextension enable --py --sys-prefix widgetsnbextension
# jupyter nbextension enable --py widgetsnbextension
jupyter nbextension enable --py widgetsnbextension
# Enabling notebook extension jupyter-js-widgets/extension... - Validating: OK
cd
cd trt_pose_hand
ls -l
# live_hand_pose.ipynb
# gesture_classifier.py
# jupyter nbextension enable --py widgetsnbextension
# ====
# USB Camera Logicool HD Webcam C270
# ====
# CSI Camera Raspberry Pi Camera Module V2
# camera = USBCamera(width=WIDTH, height=HEIGHT, capture_fps=30, capture_device=1)
# #camera = CSICamera(width=WIDTH, height=HEIGHT, capture_fps=30)
sed -i 's/camera = USBCamera/# camera = USBCamera/' live_hand_pose.ipynb
sed -i 's/#camera = CSICamera/camera = CSICamera/' live_hand_pose.ipynb
# Notebook live_hand_pose.ipynb is not trusted
jupyter trust live_hand_pose.ipynb
jupyter notebook --ip=* --no-browser
