An open API service indexing awesome lists of open source software.

https://github.com/sjkaliski/infer

🔮 Use TensorFlow models in Go to evaluate Images (and more soon!)
https://github.com/sjkaliski/infer

go golang inference machine-learning prediction tensorflow

Last synced: over 1 year ago
JSON representation

🔮 Use TensorFlow models in Go to evaluate Images (and more soon!)

Awesome Lists containing this project

README

          

# infer

Infer is a Go package for running predicitions in TensorFlow models.

[![Build Status](https://travis-ci.org/sjkaliski/infer.png)](https://travis-ci.org/sjkaliski/infer)
[![GoDoc](https://godoc.org/github.com/sjkaliski/infer?status.svg)](https://godoc.org/github.com/sjkaliski/infer)
[![Go Report Card](https://goreportcard.com/badge/github.com/sjkaliski/infer)](https://goreportcard.com/report/github.com/sjkaliski/infer)

## Overview

This package provides abstractions for running inferences in TensorFlow models for common types. At the moment it only has methods for images, however in the future it can certainly support more.

## Getting Started

The easiest way to get going is looking at some examples, two have been provided:

1. [Image Recognition API using Inception](examples/inception).
2. [MNIST](examples/mnist)

## Setup

This package requires

- [Go](https://golang.org/dl/)
- [TensorFlow for Go](https://godoc.org/github.com/tensorflow/tensorflow/tensorflow/go)

Installation instructions can be found [here](https://www.tensorflow.org/install/install_go). Additionally, a [Dockerfile](examples/Dockerfile) has been included which can be used to run the [examples](examples).

## Usage

### Overview

To use infer, a TensorFlow [Graph](https://www.tensorflow.org/programmers_guide/graphs) is required, as well as a defined Input and Output.

Classes may also be included, a slice of possible values. It's assumed the results of any model execution refer to these classes, in order. (e.g. 0 -> mountain, 1 -> cat, 2 -> apple).

```go
m, _ = infer.New(&infer.Model{
Graph: graph,
Classes: classes,
Input: &infer.Input{
Key: "input",
Dimensions: []int32{100, 100},
},
Output: &infer.Output{
Key: "output",
},
})
```

Once a new model is defined, inferences can be executed.

```go
predictions, _ := m.FromImage(file, &infer.ImageOptions{})
```

Predictions are returned sorted by score (most accurate first). A `Prediction` looks like

```go
Prediction{
Class: "mountain",
Score: 0.97,
}
```

### Graph

An `infer.Model` requires a `tf.Graph`. The Graph defines the computations required to determine an output based on a provided input. The Graph can be included in two ways:

1. Create the Graph using Go in your application.
2. Load an existing model.

For the latter, an existing model (containing the graph and weights) can be loaded using Go:

```go
model, _ := ioutil.ReadFile("/path/to/model.pb")
graph := tf.NewGraph()
graph.Import(model, "")
```

For more information on TensorFlow model files, [see here](https://www.tensorflow.org/extend/tool_developers/).

### Input & Output

`infer.Input` and `infer.Output` describe [TensorFlow layers](https://www.tensorflow.org/tutorials/layers). In practice this is the layer the input data should be fed to and the layer from which to fetch results.

Each require a `Key`. This is the unique identifier (name) in the TensorFlow graph for that layer. To see a list of layers and type, the following can be run:

```go
ops := graph.Operations()
for _, o := range ops {
log.Println(o.Name(), o.Type())
}
```

If you're not using a pre-trained model, the layers can be named, which can ease in identifying the appropriate layers.

### Analyzing Results

In the [MNIST example](examples/mnist), we can execute a prediction and inspect results as so:

```go
predictions, err := m.FromImage(img, opts)
if err != nil {
panic(err)
}

// predictions[0].Class -> 8
```