DNNClassifier.fit_generator()

tf_oop_fit_generator.py

def fit_generator(self, generator, epochs, batches_per_epoch,
                  validation_data=None, callbacks=None):
    """
    Fits model with generator yielding batches of (x, y) pairs.

    The generator is expected to indefinitely generate samples from
    training set. Therefore, we need a "hint" how much times we would like
    to call generator's `next()` method during single epoch. For this
    purpose `batches_per_epoch` parameter is used.
    """
    if self._session is not None:
        self._session.close()

    graph = self._graph

    with graph.as_default():
        x, y, training, dropout = tf.get_collection('inputs')
        loss, accuracy = tf.get_collection('metrics')
        logits, training_op = tf.get_collection('training')
        init = tf.global_variables_initializer()
        self._saver = tf.train.Saver()

    self._session = session = tf.Session(graph=graph)
    monitor = CallbacksGroup(callbacks or [])
    monitor.set_model(self)
    monitor.on_start_training()

    init.run(session=session)
    for epoch in range(1, epochs + 1):
        if self.stop_training:
            break
        epoch_loss = 0.0
        for batch_index in range(batches_per_epoch):
            x_batch, y_batch = next(generator)
            feed = {x: x_batch, y: y_batch, training: True, dropout: 0.5}
            _, batch_loss = session.run([training_op, loss], feed)
            epoch_loss += batch_loss
            monitor.on_batch(epoch, batch_index)
        epoch_loss /= batches_per_epoch
        metrics = {'train_loss': epoch_loss}
        if validation_data:
            x_valid, y_valid = validation_data
            feed = {x: x_valid, y: y_valid, training: False, dropout: 0}
            val_acc, val_loss = session.run([accuracy, loss], feed)
            metrics['val_loss'] = val_loss
            metrics['val_acc'] = val_acc
        monitor.on_epoch(epoch, **metrics)
    monitor.on_end_training()