Image classification using DNNs and Transfer Learning.

[codemzs]

• Rewired Microsoft.ML.Tensorflow to use Tensorflow .NET C# bindings.
• Introduces Microsoft.ML.Dnn package that contains transfer learning based image classification API using resnet and inception V3 models.
  • Contains API to retrain a DNN model using tensorflow.

All tensorflow tests are passing. The current test failures have to do with ONNX tests that seem to be using a higher version of protobuf since TF.NET also uses protobuf.

[CESARDELATORRE]

I'm really excited to try this with more elaborated samples as soon as we have the NuGet package. I provided some feedback with suggestions and possible improvements, but I consider this code could already be merged and keep evolving it from there taking my comments into account, since this still will be a very early preview.

The sooner we can create samples and code like what users would be creating, the better for being able to provide feedback and evolve accordingly. 👍

[Oceania2018]

@CESARDELATORRE We will add more other Model Architectures like MobileNet, ResNet and NASNet, which could give you some extra precision. I will implement it in TF.NET first. Hope it be true for ML.NET v1.4 release. #Resolved

[CESARDELATORRE]

This is great, I see you added InceptionV3 over the weekend! 👍 #Resolved

[codemzs]

@Oceania2018 This release we are not advertising GPU support. I want to first test it before we remove this dependency.

[codemzs]

Thanks, @CESARDELATORRE for reviewing.