How to import a Keras model into a Vue.js application using Tensorflow.Js

In this Tensorflow.js tutorial I will show how to create a really simple model using Keras, then convert the model and import it into a Vue.js app using tensorflow.js. Above is the vid and below you will find some of the steps followed.

There are two projects as show in the video, one written in python that trains the model using Keras, Tensorflow and other libraries. In that project there is also a dockerfile for the tensorflowjs_converter that will be used for converting the model for later use inside the javascript application.
And then there is a vue.js project that was created using the Tensorflow.js + Vue.js Starter Template that already has TF.js configured (If you want to see how I created that application you can read the Creating a Tensorflow.js Vue.js Simple Application In Javascript post plus video!).
Below are the most important parts from the video:

1.
Pre-reqs
- You need Anaconda with Keras and Tensorflow installed
- Also node.js installed
  
  And docker installed (Optional!)

Keras/Tensorflow Python project

So first you need a model, in this case let's say that you created a model using Keras, you would then need to save it and finally transform it.

Create the model

  
#%% Create the model
def build_model(x_size, y_size):
    t_model = Sequential()
    t_model.add(Dense(x_size, input_shape=(x_size,)))

    t_model.compile(loss='mean_squared_error',
        optimizer='sgd',
        metrics=[metrics.mae])
    return(t_model)

model = build_model(X_train.shape[1], y_train.shape[1])

Train the model

  
history = model.fit(X_train, y_train,
    batch_size=batch_size,
    epochs=epochs,
    shuffle=True,
    verbose=2,
    validation_data=(X_test, y_test),
    callbacks=keras_callbacks)

And save the model
```
  
model.save('./model/model.h5')
  
```

Convert the model using the tfjs Converter

If you have Docker installed, first build the image
```
  
docker build -t tf-converter .
  
```

And finally use the tensorflow.js converter container giving it the path to the current keras model and the path for the converted model

  
docker run -it --rm --name tf-converter -v "$(pwd)":/workdir tf-converter
                                      --input_format keras \
                                      ./model/-scaled-categorical/model.h5 \
                                      ./shared/model

If you don't have docker :(, you first need to install tensorflow.js
```
  
pip install tensorflowjs
  
```

And then as before use the tensorflowjs_converter command

  
tensorflowjs_converter --input_format keras \
                       ./model/-scaled-categorical/model.h5 \
                       ./shared/model

Vue.js + Tensorflow Javascript project

Load the model

  
that.model = await tf.loadModel('/shared/model/model.json');

And then you can just take the inputs, transform them into a tensor and then predict the output

  

    predictValue(inputs) {
      const tfarray = tf.tensor2d(inputs, [1, inputs.length]);
      const prediction = this.model.predict(tfarray);

      prediction.print();

      return prediction.get(0, 0);
    }

5.

Resources

Github link for the code

Tensorflow.js Vue.js Starter template

Keras

Tensorflow.js Converter

Loss functions

Metrics functions

Creating checkpoints

Early stopping

DLightHouse

Buscar este blog