nnview, neural network viewer

nnview can display weights and graph structure of neural network model.

Status

Very early stage. W.I.P.

• LinearFunction
• ReLU
• Convolution2D
• and more

How it works

• Read graph definition in JSON
• Read weight data
• Plot graph
• Plot weight data

Build

Requirements

• C++14 compiler(imgui-node-editor requires C++14. Other component can be compiled with C++11)
• Cmake
• OpenGL 3.x(imgui-node-editor requires GL 3.x)

cmake options

• NNVIEW_USE_CCACHE On/Off : Compile with ccache
• NNVIEW_USE_NATIVEFILEDIALOG On/Off Use NativeFileDialog. default on for Windows and macOS
• SANITIZE_ADDRESS=On : Enable address sanitizer. Requires clang or recent gcc.

BUild on Linux and macOS

See scripts/bootstrap-linux.sh and scripts/bootstrap-macos.sh for examle cmake bootstrapping.

Build on Windows(Visual Studio 2017)

> vcsetup.bat

Solution file will be generated into build folder.

UI

Graph

• 'F' key to fit view
• right mouse drag to pan view

Supported format

Currently we only support JSON and weight generated by Chainer-TRT(https://github.com/pfnet-research/chainer-trt)

License

MIT license

TODO

• Support .tflite format(TensorFlow-Lite, Flatbuffers format)
• Support weight data in NPY(numpy) or NPZ(numpy zip compressed) format.
• Use nlohmann json.hpp or rapidjson for JSON schema validation.
• Better graph layout.

Third party licences

• imgui : Public domain
• imgui-node-editor : Copyright (c) 2019 Michał Cichoń. MIT license.
• glfw3 : zlib license
• nativefiledialog : zlib license
• roboto font : Apache 2.0 license
• json11 : MIT license. https://github.com/dropbox/json11