Installation

The doc contains two parts mainly: C backend module build and the CVM & MRT python package installation.

Pre-requisites

Ubuntu (CentOS, RedHat)

Dependent compilation tools: g++, make, cmake, python3

MacOS

we have already tested the project setup on MacOS, and the system need some extra dependency library installation besides ubuntu requests.

Using the brew command to install libomp in Mac. And if neccessary, install the brew tools in Mac terminal first, please.

brew install libomp

CXX Build

This section gives a guide of compiling CVM.

Setup Configurations for Compilation

Check file config.cmake exists in your project root, or execute the following command to create:

cp cmake/config.cmake .

The config.cmake make some settings to change the behavior of compilation:

• PROFILE: print the inference time cost of each operators in model.

• DEBUG: add the -g flag in cxx compilation to gdb debug.

• ENABLE_CPU: enable CPU inference

• ENABLE_FORMAL: enable FORMALIZATION inference(slow but code easy read)

• ENABLE_CUDA: enable CUDA device

• ENABLE_OPENCL: enable OPENCL device(not well-supported)

• USE_NANO: option that enable the optimizer for NVIDIA jetson-nano device

CVM Build

All the targets locate at the directory build by default. You can build the project with command:

make all

Notice: execute make -j8 lib for the first time might encounter an error like:

Error copying file (if different) from "/alongdirname/cuda_compile_1_generated_broadcast.cu.o.depend.tmp" to "/alongdirname/cuda_compile_1_generated_broadcast.cu.o.depend".
CMake Error at /alongdirname/cuda_compile_1_generated_broadcast.cu.o.cmake:246 (message):
  Error generating
  /alongdirname/./cuda_compile_1_generated_broadcast.cu.o
build.make:83: recipe for target '/alongdirname/cuda_compile_1_generated_broadcast.cu.o' failed

Executing it again will fix the problem.

However the command above will compile all modules in the project, you can specify build target as below sections for a less complication time.

cvm library

CVM project is pure C&C++ backend AI inference framework. It’s easy for developer to compile the target shared library via below commands:

make lib

The command will generate two dynamic libraries: libcvm.so and libcvm_runtime.so, where libcvm_runtime.so is lighter than libcvm.so. Generally library cvm_runtime contains only the inference relative module whereas the cvm library compiles the extra symbol, graph, compile pass module etc.

unit test

This project have some unit test in subdirectory tests, you could compile the test binary with the command:

make tests

It will generate many test binary in the build/tests/ directory.

doc generator

CVM support sphinx documentation, and before compiling the documentation in directory docs, you need to install some required python packages. We have collected the neccessary packages in the docs/requirements.txt, so you can simply run the command to install the pre-requisites:

pip install -r docs/requirements.txt

Besides, make sure that you have installed the doxygen for C++ API generator. Refer to offical website for more details: Doxygen Install.

And then generate the html format with this command:

make html

All the target html resource files locate at directory docs/html as static assets, you can serve it with command:

python -m http.server --directory docs/html [port]

And then read the doc.

Python Installation

CVM & MRT python packages are respectively stored at directories: python/cvm and python/mrt. and you may setup python for the ability of importing cvm and mrt.

Python Environments

• python version: 3.6 and higher

• python dependencies: numpy

CVM requirements

Use the command to install CVM python requirements:

pip install -r install/requirements.txt

More dependency details refer to install/requirements.txt please.

MRT requirements

The mxnet and gluoncv are required. For installation, please respectively refer to:

• mxnet installation
• gluoncv installation

Notice: Mxnet in official website has different versions that whether enables cuda. Install with pip for the right mxnet release version if you want to use GPU, such as the mxnet_cu101 version etc.

Setup

There are two way of installing python packages: using the python default setup tools or export PYTHONPATH environments if you don’t want to pollute your python package environments.

Target Build

The python makefile target have been set for installing the cvm and mrt python package into sitepackages. Run the command as below:

make python

Notice: the command above will install the cvm and mrt packages, so that it’s neccessary to prepare the two requirements before.

Export Environments

Execute the commands to export PYTHONPATH in terminal or add the commands into your .bashrc if you are frustrated about typing export commands before every time opening terminal:

export PYTHONPATH={the project path}/python:${PYTHONPATH}
export LD_LIBRARY_PATH={the project path}/build:${LD_LIBRARY_PATH}

Install Test

After passing the above procedures, try to import the cvm & mrt packages for test:

python -c 'import cvm'
python -c 'import mrt'

Here is a successfull output:

Register cvm library into python module: cvm.symbol