Tensor Compiler

FRONTEND

A frontend for a tensor compiler that parses neural network models in ONNX format and converts them into an internal graph representation. The graph can be visualized using Graphviz.

Frontend features

• Reads ONNX models (.onnx) using Protocol Buffers;
• Builds a computational graph with tensors and operations;
• Supports operations: Add, Mul, Conv, Relu, MatMul, Gemm;
• Visualizes the graph as a DOT file and renders it to PNG (requires Graphviz);
• Logging system with configurable log file (can be found in ./logs);

Requirements

• C++17
• CMake 3.10+
• Protobuf compiler (protoc) and library (libprotobuf-dev)
• Graphviz
• Python 3 with onnx and numpy (optional, for generating test models)

Installation

git clone https://github.com/a-palonskaa/TensorCompiler
cd TensorCompiler

Build

cmake -S . -B build -DCMAKE-COMPILE-OPTIONS=ON
cmake --build build --parallel

Run

./build/run [name].onnx [optional -d]

-d is a flag for debug vesion with debug text dump Name from ./models. Image of graph and a dot file are located in ./images.

Testing

cd build
ctest --output-on-failure

Notes

To get absl:

git clone https://github.com/abseil/abseil-cpp.git

Authors:

• Palonskaya Alina, B01-405 DREC
• Makarskaya Alexandra, B01-401 DREC