Skip to content

Latest commit

 

History

History
122 lines (84 loc) · 3.8 KB

using-pure-cpp.md

File metadata and controls

122 lines (84 loc) · 3.8 KB
Raw

Using C++-based ML Frameworks in ns-3

TensorFlow C API

libtensorflow Installation

If model/libtensorflow directory exists, targets using TensorFlow C API are automatically enabled.

For x86-64-based operating systems

  1. Download prebuilt library from TensorFlow official website.
  2. Extract tarball under model.
cd YOUR_NS3_DIRECTORY
mkdir contrib/ai/model/libtensorflow
tar -xf PATH_TO_TARBALL -C contrib/ai/model/libtensorflow

For arm64-based macOS

  1. The website does not provide arm64 prebuilt library. You need to build it, or get prebuilt library with brew. You can install libtensorflow with brew:
brew install tensorflow
  1. Then copy the files to ns3-ai.
cd YOUR_NS3_DIRECTORY
cp -r /opt/homebrew/Cellar/libtensorflow/YOUR_VERSION/ contrib/ai/model/libtensorflow

Cmake settings

The following variables are available if libtensorflow is installed correctly:

  • NS3AI_LIBTENSORFLOW_EXAMPLES: Whether libtensorflow is found. Should be ON or OFF.
  • TensorFlow_LIBRARIES: Dynamically-linked libraries.
  • Libtensorflow_INCLUDE_DIR: The include directory.

These variables are helpful for adding libtensorflow-based examples (targets).

Example

  • Cmake target: ns3ai_ltecqi_purecpp

The Python API of TensorFlow provides the full functionality, while the C API is in progress and incomplete. As a result, the LTE-CQI example, which uses LSTM and requires Gradients and Neural Network library, cannot be rewritten into pure C++ version using libtensorflow.

However, a basic example is provided for checking whether libtensorflow is correctly installed. If it is, ./ns3 run ns3ai_ltecqi_purecpp should successfully print TensorFlow's version. The full example using LSTM will be available when TensorFlow offers sufficient C API to developers.

PyTorch C++ API

libtorch Installation

If model/libtorch directory exists, targets using PyTorch C++ API are automatically enabled.

For x86-64-based operating systems

  1. Download prebuilt library from PyTorch official website.
  2. Unzip under model.
cd YOUR_NS3_DIRECTORY
unzip PATH_TO_ZIP -d contrib/ai/model/

For arm64-based macOS

  1. The website does not provide arm64 prebuilt library. You need to build it, or get prebuilt library with brew. You can install libtorch with brew:
brew install pytorch
  1. Then copy the files to ns3-ai.
cd YOUR_NS3_DIRECTORY
cp -r /opt/homebrew/Cellar/pytorch/YOUR_VERSION/ contrib/ai/model/libtorch

Cmake settings

The following variables are available if libtorch is installed correctly:

  • NS3AI_LIBTORCH_EXAMPLES: Whether libtorch is found. Should be ON or OFF.
  • Torch_LIBRARIES: Dynamically-linked libraries.
  • Libtorch_INCLUDE_DIRS: Include directories.

These variables are helpful for adding libtorch-based examples (targets). Additionally, you may need to link with Python (Python_LIBRARIES) in case symbols like _PyBaseObject_Type are missing.

Example

  • Cmake target: ns3ai_rltcp_purecpp

The message-interface version of the RL-TCP example has been modified to run in pure C++, utilizing PyTorch C++ APIs. All RL functionalities are same with original, except for random seed settings that can lead to different results. Running ns3ai_rltcp_purecpp should by default apply deep Q-learning algorithm (DQN) to choose TCP parameters, with states and actions printing in the console.

pip install -r contrib/ai/examples/rl-tcp/requirements.txt
./ns3 run ns3ai_rltcp_purecpp