adding_new_layers_keras.md

Exporting and Importing layer - Keras

Follow the guide here to add your new layer to Fabrik's frontend. Then follow through this guide to add support for importing and exporting your layer for Keras.

Importing a layer

• Open keras_app/views/layers_import.py.

  • Add a function to process the new layer.

    • The function should take a Keras Layer object, called layer.
    • Get the layer parameters from layer, and create a json layer with the function jsonLayer.
    • Return the json layer you just created.

    def Dropout(layer):
        params = {}
        if (layer.rate is not None):
            params['rate'] = layer.rate
        if (layer.seed is not None):
            params['seed'] = layer.seed
        if (layer.trainable is not None):
            params['trainable'] = layer.trainable
        return jsonLayer('Dropout', params, layer)
    

• Open keras_app/views/import_json.py.

  • From layers_import, import the function you just defined.

    from layers_import import Dropout
    

  • Map your layer name to your function in layer_map.

    layer_map = {
        'InputLayer': Input,
        'Dense': Dense,
        ...
    +   'Dropout': Dropout,
    }

Exporting a layer

• Open keras_app/views/layers_export.py.

  • Add a function to process your layer that takes in the following arguments:

    • layer : a json string, same as the one created in layers_import.py
    • layer_in: a Tensor, output of the previous layer
    • layerId: a string id of layer
    • tensor: a bool

  • Inside the function, get layer parameters from layer and build a Keras layer using them.

  • If tensor==True, call the Keras layer on layer and return the output tensor. Else, return the layer itself.

  def dropout(layer, layer_in, layerId, tensor=True):
      out = {layerId: Dropout(0.5)}
      if tensor:
          out[layerId] = out[layerId](*layer_in)
      return out
  

• Open ide/tasks.py.

  • From layers_export, import the function you just defined.

  from keras_app.views.layers_export import dropout
  

  • Add your function to layer_map inside export_keras_json.

  layer_map = {
      'InputLayer': Input,
      'Dense': Dense,
          ...
  +   'Dropout': dropout,
  }

• Open keras_app/views/export_json.py

  • From layers_export, import the function you just defined.

  from layers_export import dropout
  

  • Add your function to layer_map.

  layer_map = {
      'InputLayer': Input,
      'Dense': Dense,
      ...
  +  'Dropout': dropout
  }

  Note : Fabrik exports models using the celery task export_keras_json, but we keep export_json.py updated to test the export code.