• TensorFlow.NET
• • • Why TensorFlow.NET ?
    • How to use
    • More examples:
    • Contribute:
    • How to debug unit tests:
    • Git Knowhow for Contributors
    • Contact

TensorFlow.NET

TensorFlow.NET (TF.NET) provides a .NET Standard binding for TensorFlow. It aims to implement the complete Tensorflow API in CSharp which allows .NET developers to develop, train and deploy Machine Learning models with the cross-platform .NET Standard framework.

TF.NET is a member project of SciSharp STACK.

Why TensorFlow.NET ?

SciSharp STACK's mission is to bring popular data science technology into the .NET world and to provide .NET developers with a powerful Machine Learning tool set without reinventing the wheel. Scince the APIs are kept as similar as possible you can immediately adapt any existing Tensorflow code in C# with a zero learning curve. Take a look at a comparison picture and see how comfortably a Tensorflow/Python script translates into a C# program with TensorFlow.NET.

SciSharp's philosophy allows a large number of machine learning code written in Python to be quickly migrated to .NET, enabling .NET developers to use cutting edge machine learning models and access a vast number of Tensorflow resources which would not be possible without this project.

In comparison to other projects, like for instance TensorFlowSharp which only provide Tensorflow's low-level C++ API and can only run models that were built using Python, Tensorflow.NET also implements Tensorflow's high level API where all the magic happens. This computation graph building layer is still under active development. Once it is completely implemented you can build new Machine Learning models in C#.

How to use

Install TF.NET through NuGet.

PM> Install-Package TensorFlow.NET

Import TF.NET.

using Tensorflow;

Add two constants:

// Create a Constant op
var a = tf.constant(4.0f);
var b = tf.constant(5.0f);
var c = tf.add(a, b);

using (var sess = tf.Session())
{
    var o = sess.run(c);
}

Feed placeholder:

// Create a placeholder op
var a = tf.placeholder(tf.float32);
var b = tf.placeholder(tf.float32);
var c = tf.add(a, b);

using(var sess = tf.Session())
{
    var o = sess.run(c, new FeedItem(a, 3.0f), new FeedItem(b, 2.0f));
}

Linear Regression:

// We can set a fixed init value in order to debug
var W = tf.Variable(-0.06f, name: "weight");
var b = tf.Variable(-0.73f, name: "bias");

// Construct a linear model
var pred = tf.add(tf.multiply(X, W), b);

// Mean squared error
var cost = tf.reduce_sum(tf.pow(pred - Y, 2.0f)) / (2.0f * n_samples);

// Gradient descent
// Note, minimize() knows to modify W and b because Variable objects are trainable=True by default
var optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost);

// Initialize the variables (i.e. assign their default value)
var init = tf.global_variables_initializer();

// Start training
with(tf.Session(), sess => 
{
    // Run the initializer
    sess.run(init);
    
    // Fit all training data
    for (int epoch = 0; epoch < training_epochs; epoch++)
    {
        foreach (var (x, y) in zip<float>(train_X, train_Y))
            sess.run(optimizer, new FeedItem(X, x), new FeedItem(Y, y));
        
        // Display logs per epoch step
        if ((epoch + 1) % display_step == 0)
        {
            var c = sess.run(cost, new FeedItem(X, train_X), new FeedItem(Y, train_Y));
            Console.WriteLine($"Epoch: {epoch + 1} cost={c} " + $"W={sess.run(W)} b={sess.run(b)}");
        }
        
        Console.WriteLine("Optimization Finished!");
        var training_cost = sess.run(cost, new FeedItem(X, train_X), new FeedItem(Y, train_Y));
        Console.WriteLine($"Training cost={training_cost} W={sess.run(W)} b={sess.run(b)}");
        
        // Testing example
        var test_X = np.array(6.83f, 4.668f, 8.9f, 7.91f, 5.7f, 8.7f, 3.1f, 2.1f);
        var test_Y = np.array(1.84f, 2.273f, 3.2f, 2.831f, 2.92f, 3.24f, 1.35f, 1.03f);
        Console.WriteLine("Testing... (Mean square loss Comparison)");
        
        var testing_cost = sess.run(tf.reduce_sum(tf.pow(pred - Y, 2.0f)) / (2.0f * test_X.shape[0]), new FeedItem(X, test_X), new FeedItem(Y, test_Y));
        Console.WriteLine($"Testing cost={testing_cost}");
        
        var diff = Math.Abs((float)training_cost - (float)testing_cost);
        Console.WriteLine($"Absolute mean square loss difference: {diff}");
    }
});

Run this example in Jupyter Notebook.

Read the docs & book The Definitive Guide to Tensorflow.NET.

More examples:

Run specific example in shell:

dotnet TensorFlowNET.Examples.dll -ex "MNIST CNN"

Example runner will download all the required files like training data and model pb files.

• Hello World
• Basic Operations
• Linear Regression
• Logistic Regression
• Nearest Neighbor
• Naive Bayes Classification
• Full Connected Neural Network
• Image Processing
• K-means Clustering
• NN XOR
• Object Detection
• Text Classification
• CNN Text Classification
• MNIST CNN
• Named Entity Recognition
• Transfer Learning for Image Classification in InceptionV3

More troubleshooting of running example refer here.

Contribute:

Feel like contributing to one of the hottest projects in the Machine Learning field? Want to know how Tensorflow magically creates the computational graph? We appreciate every contribution however small. There are tasks for novices to experts alike, if everyone tackles only a small task the sum of contributions will be huge.

You can:

• Let everyone know about this project
• Port Tensorflow unit tests from Python to C#
• Port missing Tensorflow code from Python to C#
• Port Tensorflow examples to C# and raise issues if you come accross missing parts of the API
• Debug one of the unit tests that is marked as Ignored to get it to work
• Debug one of the not yet working examples and get it to work

How to debug unit tests:

The best way to find out why a unit test is failing is to single step it in C# and its pendant Python at the same time to see where the flow of execution digresses or where variables exhibit different values. Good Python IDEs like PyCharm let you single step into the tensorflow library code.

Git Knowhow for Contributors

Add SciSharp/TensorFlow.NET as upstream to your local repo ...

git remote add upstream git@github.com:SciSharp/TensorFlow.NET.git

Please make sure you keep your fork up to date by regularly pulling from upstream.

git pull upstream master

Contact

Feel free to star or raise issue on Github.

Follow us on Medium.

Join our chat on Gitter.

Scan QR code to join Tencent TIM group:

TensorFlow.NET is a part of SciSharp STACK