.. _FlagelNet:

Specifying the modified Flagel *et al.* CNN architecture
========================================================

The Python scripts containing code to train the modified Flagel *et al.* network can be found in the files
``train_*_flagel.py``. They can be run using the following command commands:

.. code-block:: bash

  # Here divergence_scaling is either 0.5, 1.0, or 2.0 coalescent units

  # Minimum dXY network
  python3 train_min_flagel.py --coal_units <divergence_scaling>

  # Mean dXY network
  python3 train_mean_flagel.py --coal_units <divergence_scaling>

TensorFlow imports
------------------

.. code-block:: python

  import tensorflow.keras
  from tensorflow.keras.models import Sequential
  from tensorflow.keras.layers import (
    Dense,
    Dropout,
    Flatten
  )
  from tensorflow.keras.layers import (
    Conv1D,
    AveragePooling1D
  )
  from tensorflow.keras.callbacks import (
    EarlyStopping,
    ModelCheckpoint
  )
  from tensorflow.keras.losses import categorical_crossentropy
  from tensorflow.keras.optimizers import Adam

Specifying the network architecture
-----------------------------------

.. code-block:: python

  model = Sequential()
  model.add(
      Conv1D(
          64, kernel_size=2,
          activation='relu',
          input_shape=(xtrain.shape[1],xtrain.shape[2])
      )
  )
  model.add(
      Conv1D(
          32, kernel_size=2,
          activation='relu'
      )
  )
  model.add(
      AveragePooling1D(
          pool_size=2
      )
  )
  model.add(Dropout(0.25))

  model.add(
      Conv1D(
          32, kernel_size=2,
          activation='relu'
      )
  )
  model.add(
      AveragePooling1D(
          pool_size=2
      )
  )
  model.add(Dropout(0.25))

  model.add(Flatten())
  model.add(Dense(32, activation='relu'))
  model.add(Dropout(0.5))
  model.add(Dense(32, activation='relu'))
  model.add(Dropout(0.5))
  model.add(Dense(4, activation='softmax'))
  model.compile(
      loss=categorical_crossentropy,
      optimizer=Adam(),
      metrics=['accuracy']
  )
  print(model.summary())

  callbacks = [
      EarlyStopping(monitor='val_loss'),
      ModelCheckpoint(
          filepath='hyde_flagel_mean_{}.mdl'.format(cu),
          monitor='val_loss',
          save_best_only=True
      )
  ]
  model.fit(
      xtrain, ytrain,
      batch_size=32,
      epochs=10,
      verbose=1,
      callbacks=callbacks,
      validation_data=(xval,yval)
  )

----

**References**

  - L Flagel, Y Brandvain, and DR Schrider. 2019. The Unreasonable Effectiveness of
    Convolutional Neural Networks in Population Genetic Inference. *Molecular
    Biology and Evolution* 36:220--238. https://doi.org/10.1093/molbev/msy224.