9

I have a TF model that works well, constructed with Python and TFlearn. Is there a way to run this model on another system without installing Tensorflow on it? It is already pre-trained, so I just need to run data through it.

I am aware of tfcompile (Thread here), but it seems quite complex to set up. Are there any alternatives?

Zock77
  • 951
  • 8
  • 26
  • I don't think there's a ready-built way to transfer it without simply recoding the neural network matrix operations in whatever new environment you want to run it in. – David Parks Jan 26 '18 at 03:31

1 Answers1

3

 Is there a way to run this model on another system without installing Tensorflow on it? It is already pre-trained, so I just need to run data through it.

Yes

After you have your model trained . Use tf.python.tools.freeze_graph and tf.python.tools.optimize_for_inference_lib to freeze and optimize the model for inference on other devices like Android.

The output of the above will be

  1. Frozen graph protobuf file (.pb)
  2. Optimized graph protobuf file (.pb)

[These functions will converts all the Variables of the Model to Constant Operations and exports to a protobuf file]

Use the optimized graph protobuf file and load it using Inference methods available in Java and other Tensorflow APIs. Pass the data and get the output.

[ Note for this you didn't installed complete Tensorflow but you only needed the inference library]

A Simple example is demonstrated here :

https://omid.al/posts/2017-02-20-Tutorial-Build-Your-First-Tensorflow-Android-App.html

It is for Android but procedure should be same for Java.

For C++ :click here

coder3101
  • 3,920
  • 3
  • 24
  • 28