AI Visualizer

This application allows visualizing the activation of different parts of an AI network as data is passed through it. This is based on the idea of AI models being built from combinations of layers that data is fed through. The outputs of each layer can then be inspected and displayed in a way that befits the layer type.

This does not train a model, but can take in an already created model saved from Tensorflow. This utilizes a utility in the fdeep library to convert the keras output format to an representation that the fdeep C++ library can ingest.

Requirements

(Testing is done with MINGW 64-bit on Windows 10)

• Qt 6 (tested on 6.6.2)
• Frugally Deep (fdeep)
• Eigen (Linear algebra library used by Fdeep)

Architecture

The project is broken into a few libraries and the main application. These are split into the following directories:

• application - GUI application and main executable for interacting and viewing the models
• viewmodels - Qt QML module containing QML views and the Qt mdoels that expose the underlying data to the QML
• interface - Classes defining the bridge between the underlying model representation and the front end viewmodels
• impl - Implementation of bridge to interact with underlying model implementations
  • fdeepbridge - Specific implementation to interact with Fdeep models
• common - Commonly included source files accross multiple of the previous items
• test - Integration level tests to full load of an AI model and creation of corresponding ViewModels

The purpose of the separation is to decouple the front end and back end implementations, and leave the opportunity to support other underlying model representations in the future. In addition, the fdeepbridge and viewmodels libraries use registries to be able to select the proper layer classes, models, and views such that they could be chosen dynamically. This way, adding or replacing layer representations and view can be done outside of the libraries (although may still need them as dependencies).

Current work in progress

This project is currently in progress, and this is roughly the work to complete before considering a "release"

• Implement dense layer
• Implement convolutional layer
• Create application
  • select and load model file
  • select and load a data file
  • run data and view activation of layers from the run data
• Clean up unused 3rd party modules
• Clean up extension of Fdeep to allow inspection of model. Currently it's just a few files that were copied and slightly modified to create extended versions of the needed classes. It would likely be cleaner to fork the repository and include the minor changes directly to the desired classes
• Create installer

Needed Improvements

• Optimization. Currently, the needs of the interface are being shaken out. As that settles it would be good to optimize how certain items are accessed to minimize data copying. The main trade off of over optimization would be inflexibility. It could be much faster to remove the bridge/interface to directly interact with the Eigen container classes used by Fdeep, but this would introduce heavy coupling. The downside is copying/converting the stored data to std::vectors for a more generic representation.
• More layer types need to be implemented. Currently only dense layers are supported.
• Package python as well and allow the application to directly select the keras format and do the fdeep conversion python step automatically.
• Fix 3rdparty submodules. Some are no-longer used and frugally deep should be built and installed separately
• Naming. Some items need updated names (like tst_annrepresentation), and some need consistency (i.e. FdeepBridge::DenseLayer and FdeepBridge::FdeepUnknownLayer)

Notes and Limitations

• This is a pet project so most items are minimally implemented, and progress is slow, and the git usage isn't to the level you can expect in industry.
• Eigen is a fairly large, header only library so compilation times can be long. MINGW requires the -mbig-obj option to support the size of certain generated objects.
• Does this support GPT? Not really. Some of the layer types GPT uses are implemented (mainly dense layers), however thought would have to be put in to determine how best to display layers such as the decode and "attention" layers. Additionally, many passes through the network may occur for a given answer. This format works well for looking at convolutional neural nets where filter layers and their outputs can be easily visually interpreted, as well as fully connected layers, which can be visualized as nodes "lighting up" as they are activated by a given input datum.