Heavy work in progress

NNPack

NNPack contains tools & standards for packaging and distributing Neural Networks. Today's Deep Learning frameworks are very powerful, but they their tools for persisting and distributing Machine Learning models are lacking. The NNPack specification provides a high-level structure for packaging up various types of models and their training data (Model Scaffolds).

NNPack Specification 1.0-DRAFT

Features

Coming soon: Easily push/pull models from the open nnpack.org repository
Developer-friendly packaging of models. Model meta-data lives in /nnpackage.json
Packaging and distributing training data using Model Scaffolds
Rich contextual meta-data for model, training data and labels

Coming Soon: Installing a Model

nnpack pull dominiek/scene_type_vision

Packaging a Model

In order to package a model, simply create an nnpackage.json in its root directory. All neural network state information needs to be moved into a state/ subfolder. Here's an example:

{
  "id": "58745c9fbd17c82dd4ff7c9c",
  "name": "Scene Types",
  "description": "Classify images into Indoor VS Outdoor scenes",
  "version": 0.1,
  "engines": {
    "tensorflow": ">=0.11"
  },
  "nodes": {
    "softmaxOutput": {
      "type": "tensor",
      "id": "retrained_layer:0"
    },
    "convolutionalRepresentation": {
      "type": "tensor",
      "id": "pool_3/_reshape:0"
    }
  },
  "author": {
    "name": "Dominiek Ter Heide",
    "email": "info@dominiek.com"
  },
  "labelsDefinitionFile": "labels.json",
  "stateDir": "state"
}

For each softmax output of the neural net, a labels.json provides context for what the predictions mean. Each node_id corresponds to a softmax output:

{
  "labels": [
    {"id": "58745e65bd17c82ec1545a64", "name": "Indoor", "node_id": 0},
    {"id": "58745e69bd17c82ec1545a65", "name": "Outdoor", "node_id": 1}
  ]
}

In the above two files a lot of information is contained on how to run the model, what engine is required, what the outputs mean and how transfer learning can occur.

Read Documentation

Coming Soon: Distributing a Model

nnpack init .
nnpack push dominiek/scene_type_vision

Todo