Extensor

Extensor provides Tensorflow bindings for inference in Elixir. This library can be used to execute computation graphs created in Tensorflow on the CPU or GPU. Extensor provides minimal abstractions over the Tensorflow C library and includes as little custom native code as possible. These NIFs have been extensively tested for memory leaks and paralellism safety so that the library can be relied on for production use.

Status

The API reference is available here.

Installation

Hex

def deps do
  [
    {:extensor, "~> 0.1"}
  ]
end

Dependencies

This project requires the Tensorflow C headers/libraries. For development, these can be installed from the steps here.

For docker deployment, see the sample dockerfiles in the docker directory. Docker for ubuntu can be tested with the following commands.

docker build -t extensor -f docker/ubuntu.dockerfile .
docker run --rm -it extensor mix test

If you have nvidia tools installed, you can test on the GPU by substituting nvidia-docker for docker above.

Usage

For a simple example, Extensor can be used to evaluate the Pythagorean identity (c² = a² + b²). The following python script can be used to create and save a graph that calculates the length of the hypotenuse.

import tensorflow as tf

a = tf.placeholder(tf.float32, name='a')
b = tf.placeholder(tf.float32, name='b')
c = tf.sqrt(tf.add(tf.square(a), tf.square(b)), name='c')

with tf.Session() as session:
    tf.train.write_graph(session.graph_def,
                         'test/data',
                         'pythagoras.pb',
                         as_text=False)

This model can then be used in Elixir to evaluate the compute graph and calculate a few hypotenuses. This model file is also available in the repo under test/data/pythagoras.pb.

session = Extensor.Session.load_frozen_graph!("test/data/pythagoras.pb")

input = %{
  "a" => Extensor.Tensor.from_list([3, 5]),
  "b" => Extensor.Tensor.from_list([4, 12])
}

output = Extensor.Session.run!(session, input, ["c"])

Extensor.Tensor.to_list(output["c"])

This block should output the list [5.0, 13.0], which corresponds to the lengths of the hypotenuses of the first two Pythagorean triples.

Model Formats

Extensor supports the frozen graph_def and saved_model serialization formats, via the load_frozen_graph and load_saved_model functions, respectively.

For example, the Google Inception model has its weights frozen to const tensors, so that it can be loaded directly from a protobuf.

However, the frozen graph approach doesn't work for models that contain non-freezable variables (like RNNs). For these models, Extensor supports the Tensorflow saved_model format, which is the format used by Tensorflow serving (TFS). The saved_model format is loaded from a directory path, which includes model metadata and initial variable weights.

Configuration

Extensor supports passing a ConfigProto object when creating a session for inference configuration. See the Tensorflow.ConfigProto module for more information on the configuration data structures.

config = %{
    Tensorflow.ConfigProto.new()
    | gpu_options: %{
        Tensorflow.GPUOptions.new()
        | allow_growth: true
    }
}

session = Extensor.Session.load_saved_model!("test/data/pythagoras", config)

Development

The Tensorflow protocol buffer wrappers were generated using the protobuf-elixir library using the following command, assuming Tensorflow is cloned in the ../tensorflow directory:

protoc --elixir_out=lib --proto_path=../tensorflow $(ls -1 ../tensorflow/tensorflow/core/framework/*.proto ../tensorflow/tensorflow/core/protobuf/*.proto)

License

Copyright 2018 Pylon, Inc.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

  http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.