How to add more layers to LSTM Autoencoder primitive

[baji-shaik01]

Hi Team,

We are using the Orion for our internal project, current version of LSTM autoencoder primitive support one encoder layer with 60 neurons and one decoder layer with 60 neurons. We would like to add one more layer in encoder and decoder. Can you please let us know how can we add more layers?.

Thanks
Baji

[sarahmish]

Hi Baji, thanks for the question! You can change the number of layers through setting the hyperparameters of the primitive. For example, you can alter the list of layers provided in default "layers" parameter of the "keras.Sequential.LSTMSeq2Seq" primitive to your liking:

from orion import Orion

hyperparameters = {
  'keras.Sequential.LSTMSeq2Seq#1': {
    'layers': [
      {
        "class": "tensorflow.keras.layers.LSTM",
        "parameters": {
          "input_shape": "input_shape",
          "units": 60
        }
      },
      {
        "class": "tensorflow.keras.layers.RepeatVector",
        "parameters": {
          "n": "window_size"
        }
      },
      {
        "class": "tensorflow.keras.layers.LSTM",
        "parameters": {
          "units": 60,
          "return_sequences": true
        }
      },
      {
        "class": "tensorflow.keras.layers.TimeDistributed",
        "parameters": {
          "layer": {
            "class": "tensorflow.keras.layers.Dense",
            "parameters": {
              "units": 1
            }
          }
        }
      },
      {
        "class": "tensorflow.keras.layers.Reshape",
        "parameters": {
          "target_shape": "target_shape"
        }
      }
    ]
  }
}


orion = Orion(
    pipeline='lstm_autoencoder',
    hyperparameters=hyperparameters
)

orion.fit(train_data)

To increase the number of layers, you can double stack multiple LSTM layers:

      {
        "class": "tensorflow.keras.layers.LSTM",
        "parameters": {
          "input_shape": "input_shape",
          "units": 60
        }
      },
      {
        "class": "tensorflow.keras.layers.LSTM",
        "parameters": {
          "units": 60
        }
      }