D20-based neural network architecture

Dependencies

• [2] TCMUFA6E Try some different hyperparameters
• [3] EB3DTD43 Model trainer: use ground altitude instead of crust mass distribution
• [4] 7ML3OFE7 Model trainer: initial train and visualize thread pair (total crust mass; todo: altitude instead)

Change contents

• edit in trainmodel/src/model.py at line 5
  [3.22]

  [4.1194]
  +

  + def icosahedral(inputs):

  + layer1 = keras.layers.Dense(20, activation=tf.math.asinh)(inputs)

  + layer2 = keras.layers.Dense(20, activation=tf.math.asinh)(layer1)

  + layer3 = keras.layers.Dense(20, activation=tf.math.asinh)(keras.layers.concatenate([layer1,layer2]))

  + layer4 = keras.layers.Dense(20, activation=tf.math.asinh)(keras.layers.concatenate([layer1,layer2]))

  + layer5 = keras.layers.Dense(20, activation=keras.activations.relu)(keras.layers.concatenate([layer2,layer3]))

  + layer6 = keras.layers.Dense(20, activation=keras.activations.relu)(keras.layers.concatenate([layer2,layer4,layer5]))

  + layer7 = keras.layers.Dense(20, activation=keras.activations.relu)(keras.layers.concatenate([layer5,layer6]))

  + layer8 = keras.layers.Dense(20, activation=tf.math.asinh)(keras.layers.concatenate([layer3,layer5,layer7]))

  + layer9 = keras.layers.Dense(20, activation=tf.math.asinh)(keras.layers.concatenate([layer4,layer6,layer7]))

  + layer10 = keras.layers.Dense(20, activation=tf.math.asinh)(keras.layers.concatenate([layer7,layer8,layer9]))

  + layer11 = keras.layers.Dense(20, activation=tf.math.asinh)(keras.layers.concatenate([layer1,layer3,layer8,layer10]))

  + layer12 = keras.layers.Dense(20, activation=tf.math.asinh)(keras.layers.concatenate([layer1,layer4,layer9,layer10,layer11]))

  + return keras.layers.concatenate([layer11,layer12])
• replacement in trainmodel/src/model.py at line 23
  [4.1245]→[4.1245:1450](∅→∅),[4.1450]→[2.0:138](∅→∅),[2.138]→[4.1450:1668](∅→∅),[4.1450]→[4.1450:1668](∅→∅)
  − layer = keras.layers.Dense(9, activation=tf.math.asinh)(inputs)

  − layer = keras.layers.Dense(24, activation=tf.math.asinh)(layer)

  − layer = keras.layers.Dense(240, activation=tf.math.asinh)(layer)

  − layer = keras.layers.Dense(560, activation=tf.math.asinh)(layer)

  − layer = keras.layers.Dense(560, activation=tf.math.asinh)(layer)

  − layer = keras.layers.Dense(240, activation=tf.math.asinh)(layer)

  − layer = keras.layers.Dense(24, activation=tf.math.asinh)(layer)

  − outputs = keras.layers.Dense(1, activation=keras.activations.sigmoid)(layer)
  [4.1245]

  [4.1668]
  + d1 = icosahedral(inputs)

  + d2 = icosahedral(inputs)

  + outputs = keras.layers.Dense(1, activation=keras.activations.sigmoid)(keras.layers.concatenate([d1,d2]))