thosakwe/tensorflow.dart

Dart bindings for Tensorflow.

---

Tensorflow API for the Dart programming language.

The revised goal of this project to provide an API for running pre-built Tensorflow models/graphs, typically built in another language, like Python.

• Installation
  • Prerequisites
  • Building as a Dependency
  • Building by Itself
• Basic Usage
  • Importing Graphs
  • Low Level API

Installation

This library uses native bindings, which (currently) are not easily distributed using Dart's Pub package manager.

Building the bindings, however, is easy and quick, as the build uses pre-built Tensorflow binaries, rather than re-building per user.

The Pub package does include pre-built libraries, but your mileage may vary. It may make sense in your case to run the CMake build from within your ~/.pub-cache locally.

Prerequisites

You'll need the following installed to run Tensorflow for Dart:

• CMake
• Dart SDK ^2.0.0
• libtensorflow

Building as a Dependency

If you are using package:tensorflow as a dependency in a Dart project, then you may be able to get by with the pre-built dynamic libraries in the lib/ folder. If not, you will have to manually enter the ~/.pub-cache and run the build.

Building by Itself

If you are contributing to the project, you will certainly need to be able to build the project on the fly. Use the provided CMakeLists.txt script to build the project on-the-fly.

Basic Usage

Importing Graphs

This project supports loading and restoring models saved from other Tensorflow frontends, i.e. Python:

import 'package:tensorflow/tensorflow.dart' as tf;

void main() {
  // Using the `SavedModel` API:
  var model = SavedModelBundle('example/saved_models');
  model.restore('variables.index');

  // Or, you can import from a `GraphDef` protocol buffer:
  var graph = Graph.fromGraphDef(graphDef);
  graph['output'].run(feed: {'input': Tensor.from('Hello, world!')});
}

Low Level API

package:tensorflow/tensorflow.dart theoretically supports the entire low-level Tensorflow API. This can be used to perform a variety of complex mathematical operations, and also be used to compose higher-level functionality. These bindings are, for the most part, auto-generated, and there is not much incentive to properly unit test them, considering that the most common use of this package will be to execute existing Tensorflow models from other languages.

import 'package:tensorflow/tensorflow.dart' as tf;

void main() {
    var shape = tf.Shape(6, 6);

    var x = tf.getVariable(
      'x',
      shape: shape,
      initializer: tf.randomUniform(
        tf.constant(shape),
        dtype: tf.DataType.DT_FLOAT,
      ),
    );

    x = tf.matMul(x, x);

    print(x.run());
}

tensorflow.dart's Issues

Doesn't work

flutter pub get
Running "flutter pub get" in somemlproject...
The current Dart SDK version is 2.8.4.

Because every version of tensorflow depends on protobuf >=0.2.0 <=0.9.0 which requires SDK version >=0.7.5 <2.0.0-∞, tensorflow is forbidden.

So, because somemlproject depends on tensorflow ^0.0.0, version solving failed.
pub get failed (1; So, because somemlproject depends on tensorflow ^0.0.0, version solving failed.)
exit code 1

Because tensorflow depends on binary >=0.0.1 which requires SDK version >=1.8.0 <2.0.0, version solving failed

Re-use the Tensor class for a Dart Struct generator

The Tensor class is good for reading binary data in Dart.

Perhaps something like angel_serialize could be patched to read/write binary data?

Use ExternalProject_Add instead of downloader script

Instructions for using this library with libtensorflow.so and libtensorflow_framework.so

Hi @thosakwe thank's for your great library. I tried to upgrade the scripts package for Dart 2.0 for submitting a PR but there is an error with the id package. I don't know C yet and CMake, so I wonder how to use this package with native libraries (purpose : use it on cli / android / ios / desktop).