# (1) Run the aggregation server with configuration using .toml
cfg = toml.load("${configPATH}.toml")
modalic.run_server(cfg)

# (2) Construct the client logic.
import modalic

# Define a FLClient object that implements all the ML logic and will
# used as an input to an internal modalic client which enables the 
# program to connect to the server an perform training in distributed fashion.
class FLClient(modalic.Client):

  def __init__(self, dataset, ...):
    self.model = Net()
    self.dataset = torch.utils.data.DataLoader(dataset, batch_size=32)
    ...

  def train(self):
    for epoch in range(0, self.epochs):
        for images, labels in self.trainloader:
            ...

    return self.model

  def serialize_local_model(self, model):
      return modalic.serialize_torch_model(model)

  def deserialize_global_model(self, global_model):
      self.model = modalic.deserialize_torch_model(
          self.model, global_model, self._get_model_shape()
      )

  def get_model_shape(self):
      return modalic.get_torch_model_shape(self.model)

  def get_model_dtype(self):
      ...

# Construct the client layer..
client = FLClient(...)

# (3) Run training for single client.
modalic.run_client(client)

# (2) Construct the client logic.

class FLClient(modalic.Client):

  def __init__(self, dataset, ...):
    # Initialize & compile the MobileNetV2 model.
    self.model = tf.keras.applications.MobileNetV2((32, 32, 3), classes=10, weights=None)
    # Load the CIFAR-10 dataset using tf.keras.
    (self.x_train, self.y_train), (_, _) = tf.keras.datasets.cifar10.load_data()
    ...

  def train(self):
    ...
    self.model.fit(self.x_train, self.y_train, batch_size=self.batch_size, epochs=self.epochs)
    ...
    return self.model

  def serialize_local_model(self, model):
      return modalic.serialize_tf_keras_model(model)

  def deserialize_global_model(self, global_model):
      self.model = modalic.deserialize_tf_keras_model(
          self.model, global_model, self._get_model_shape()
      )

  def get_model_shape(self):
      return modalic.get_tf_keras_model_shape(self.model)

  def get_model_dtype(self):
      ...

# Construct the client layer..
client = FLClient(...)

# (3) Run training for single client.
modalic.run_client(client)