[FIXED] How to make `model.fit` stop training immediately when called on a naive model that does not have trainable layers?

Issue

I am writing a training pipeline for time series forecasting models. I am using a naive seasonal model as baseline, which just outputs the last out_steps of the input.

class Naive(tf.keras.Model):
def __init__(self, out_steps: int,
             **kwargs):
    super().__init__(**kwargs)
    self.out_steps = out_steps

def call(self, inputs, training=None):
    features = inputs

    return features[:, -self.out_steps:, :]

I then can use the generic training stage:

def train_model(model_name, **model_params):
    model = instantiate_model(model_name, **model_params)
    model.compile(loss='mse', optimizer='adam')
    model.fit(train_dataset)

Is there a way to make fit understand that the model has no trainable layers and stop immediately?

Solution

Just a one-line wrapper:

if model.trainable_variables:
    model.fit(...)

model.trainable_variables contains the, well, trainable variables, for a model without such variables, this will be empty and the if condition is False.

NOTE though, that if you create models using Sequential, and did not provide an input shape in the first layer, and also did not call the build() function of the model, such a model will also not have any variables before calling fit! These will only be created once the model is called for the first time. So you have to be careful with this, or add something like this:

if model.trainable_variables:
    model.fit(...)
elif not model.built:
    raise SomeError  # or print a warning

Answered By - xdurch0