C++ Nerual Network Practice

1. Train the Network with Pytorch

Run the following command. This will train the network with Pytorch and simultaneously convert both the model and the weight into C style format in the header which can be subsequently built in our framework.

python converter.py

Both the network structure and pretrained weights are built into Network.hpp

2. Build & Run the Code

Then you can build the Inference framework from source by running the following commands. (For Windows)

mkdir build
cd build
cmake .. -DCOMPLETE_TEST=ON
cmake --build .
cd ..
build\Debug\Try.exe

3. Supported Operators

• weight converter
• model converter (TorchScript)
• Tensor::permute()
• Tensor::view()
• Tensor::repeat()
• Tensor::copy()
• Linear (no_grad)
• ReLU (no_grad)
• ReLU6 (no_grad)
• Sigmoid (no_grad)
• Tanh (no_grad)
• Conv1d (no_grad)
• Conv2d (no_grad)