1. 环境搭建
1.1. 安装环境依赖¶
📝SigmaStar DLA SDK基于AVX2指令集编写，请使用支援AVX2的Intel®处理器运行。
推荐配置
| 🖥️💻 | ⚙️ |
|---|---|
| CPU | Intel® CoreTM i7 or Higher |
| RAM | 8G or Higher |
最低配置
| 🖥️💻 | ⚙️ |
|---|---|
| CPU | Intel® CoreTM i5 |
| RAM | 6G |
系统依赖¶
| Package | Installation command |
|---|---|
| i386 | dpkg --add-architecture i386 |
| build-essential | sudo apt install build-essential |
| cmake | sudo apt install cmake |
| libc6-dev | sudo apt install libc6-dev libc6-dev:i386 libc6-dev-i386 |
| libbz2-dev | sudo apt install libbz2-dev libbz2-dev:i386 |
| libncurses5-dev | sudo apt install libncurses5-dev libncurses5-dev:i386 |
| libglib2.0-dev | sudo apt install libglib2.0-dev libglib2.0-dev:i386 |
| libsm6 | sudo apt install libsm6 libsm6:i386 |
| libxrender1 | sudo apt install libxrender1 libxrender1:i386 |
| libxext6 | sudo apt install libxext6 libxext6:i386 |
| libgdbm-dev | sudo apt install libgdbm-dev libgdbm-dev:i386 |
| liblzma-dev | sudo apt install liblzma-dev liblzma-dev:i386 |
| libsqlite3-dev | sudo apt install libsqlite3-dev libsqlite3-dev:i386 |
| libssl-dev | sudo apt install libssl-dev libssl-dev:i386 |
| libreadline6-dev | sudo apt install libreadline6-dev libreadline6-dev:i386 |
| libffi-dev | sudo apt install libffi-dev libffi-dev:i386 |
| zlib1g-dev | sudo apt install zlib1g-dev zlib1g-dev:i386 |
| libncursesw5-dev | sudo apt install libncursesw5-dev libncursesw5-dev:i386 |
| libsqlite3-dev | sudo apt install libsqlite3-dev libsqlite3-dev:i386 |
| libgdbm-dev | sudo apt install libgdbm-dev libgdbm-dev:i386 |
| libbz2-dev | sudo apt install libbz2-dev libbz2-dev:i386 |
| checkinstall | sudo apt install checkinstall |
| openssl | sudo apt install openssl |
Python 依赖¶
使用SigmaStar DLA SDK需要安装以下依赖库：
| Software | Installation Command | Tested Version |
|---|---|---|
| Python | 3.7 | |
| enum34 | pip install enum34==1.1.10 | ==1.1.10 |
| numpy | pip install numpy==1.16.6 | ==1.16.6 |
| protobuf | pip install protobuf | >=3.8.0 |
| six | pip install six | >=1.12.0 |
| OpenCV-python | pip install opencv-python | ==4.2.0.34 |
| TensorFlow | pip install tensorflow | ==1.14.0 |
| Cython | pip install cython | >=0.29.13 |
| pycocotools | pip install pycocotools | >=2.0.0 |
| matplotlib | pip install matplotlib | >=3.0.3 |
| SciPy | pip install scipy | >=1.3.1 |
| Pillow | pip install pillow | ==6.1.0 |
| joblib | pip install joblib | ==0.13.2 |
| onnx-simplifier | pip install onnx-simplifier | ==0.2.10 |
| sympy | pip install sympy | ==1.6.1 |
| packaging | pip install packaging | ==20.4 |
| onnx | pip install onnx | ==1.8.1 |
| onnxruntime | pip install onnxruntime | ==1.7.0 |
| onnxoptimizer | pip install onnxoptimizer==0.2.4 | ==0.2.4 |
| python3-tk | sudo apt install python3-tk | |
| libc6 | sudo apt install libc6-dev-i386 | |
| libstdc++6 | sudo apt install libstdc++6 | |
| python-qt4 | sudo apt install python-qt4 | |
| torch | sudo install torch==1.8.0+cpu | ==1.8.0 |
| torchvision | sudo install torchvision==0.9.0+cpu | ==0.9.0 |
| wheel | sudo install wheel | |
| scikit-image | sudo install scikit-image | |
| scikit-learn | sudo install scikit-learn | |
| pulp | sudo install pulp |
1.2. 快速上手¶
默认设置
请将SGS_Models和SGS_IPU_SDK放到主目录 ~/ 下，以下命令均基于该目录结构进行。请使用Linux环境运行本工具。
1.2.1 安装环境依赖¶
SigmaStar DLA SDK基于AVX2指令集编写，请使用支援AVX2的Intel®处理器运行。如果使用docker等虚拟机环境，请保证虚拟机内最低分配6G内存。
1.2.2 快速安装环境依赖¶
命令如下：
sudo apt update
sudo apt install python3-tk python-qt4 libc6-dev-i386 libstdc++6
cd ~/SGS_IPU_SDK
pip3 install -r Scripts/calibrator/setup/requirements.txt \
–i https://pypi.tuna.tsinghua.edu.cn/simple
1.2.3 环境设置¶
命令如下：
cd ~/SGS_IPU_SDK
source cfg_env.sh
1.2.4 快速上手说明¶
本手册使用caffe训练的mobilenet_v2作为参考例子。
在SGS_IPU_SDK ⽬录下运⾏以下脚本，输出Library的路径：
cd ~/
mkdir caffe_mobilenet_v2
cd caffe_mobilenet_v2
1.2.4.1 原始模型转化为SigmaStar浮点网络模型¶
python3 ~/SGS_IPU_SDK/Scripts/ConvertTool/ConvertTool.py caffe \
--model_file ~/SGS_Models/caffe/caffe_mobilenet_v2/caffe_mobilenet_v2.prototxt \
--weight_file ~/SGS_Models/caffe/caffe_mobilenet_v2/caffe_mobilenet_v2.caffemodel \
--input_arrays data \
--output_arrays prob \
--output_file ./caffe_mobilenet_v2_float.sim \
--input_config ~/SGS_Models/caffe/caffe_mobilenet_v2/input_config.ini
1.2.4.2 SigmaStar浮点网络模型转化为SigmaStar定点网络模型¶
进入caffe_mobilenet_v2文件夹，运行:
python3 ~/SGS_IPU_SDK/Scripts/calibrator/calibrator.py \
-i ~/SGS_Models/resource/classify/ilsvrc2012_calibration_set32/ \
-m ./caffe_mobilenet_v2_float.sim \
-c Classification \
--input_config ~/SGS_Models/caffe/caffe_mobilenet_v2/input_config.ini \
-n caffe_mobilenet_v2
1.2.4.3 SigmaStar定点网络模型转化为SigmaStar离线网络模型¶
进入caffe_mobilenet_v2 文件夹，运行:
python3 ~/SGS_IPU_SDK/Scripts/calibrator/compiler.py \
-m ./caffe_mobilenet_v2_fixed.sim
1.2.5 模型仿真¶
1.2.5.1 使用simulator对SigmaStar浮点网络模型验证¶
python3 ~/SGS_IPU_SDK/Scripts/calibrator/simulator.py \
-i ~/SGS_Models/resource/classify/ilsvrc2012_val_set100 \
-l ~/SGS_Models/resource/classify/caffe_labels.txt \
-m ./caffe_mobilenet_v2_float.sim \
-c Classification \
-t Float \
-n ~/SGS_Models/caffe/caffe_mobilenet_v2/caffe_mobilenet_v2.py \
--num_process 20
1.2.5.2 使用simulator对SigmaStar定点网络模型验证¶
python3 ~/SGS_IPU_SDK/Scripts/calibrator/simulator.py \
-i ~/SGS_Models/resource/classify/ilsvrc2012_val_set100 \
-l ~/SGS_Models/resource/classify/caffe_labels.txt \
-m ./caffe_mobilenet_v2_fixed.sim \
-c Classification \
-t Fixed \
-n ~/SGS_Models/caffe/caffe_mobilenet_v2/caffe_mobilenet_v2.py \
--num_process 20
1.2.5.3 使用simulator对SigmaStar离线网络模型验证¶
python3 ~/SGS_IPU_SDK/Scripts/calibrator/simulator.py \
-i ~/SGS_Models/resource/classify/ILSVRC2012_test_00000002.bmp \
-m caffe_mobilenet_v2_fixed.sim_sgsimg.img \
-l ~/SGS_Models/resource/classify/labels.txt \
-c Classification \
-t Offline \
-n ~/SGS_Models/caffe/caffe_mobilenet_v2/caffe_mobilenet_v2.py