Define loss functions

nucleus

A lightweight library for importing, manipulating and visualizing images and datasets.

	def create_model(
	self,
	input_shape: Tuple = (None, None, 3),
	weights: Optional[Union[str, Path]] = 'imagenet',
	trainable: Union[bool, Sequence[bool]] = True,
	**kwargs
	) -> tf.keras.Model:
	r"""
	Creates a backbone model.

	Parameters
	----------
	input_shape

	alpha

	weights

	trainable

	kwargs


	Returns
	-------
	The backbone model.
	"""
	# Input layer
	inputs = tf.keras.Input(input_shape)

	# Define the backbone specific pre-processing layer
	x = self._preprocessing_layer(inputs)

	# Define the backbone network
	backbone: tf.keras.Model = self._backbone_fn(
	# layers=self._patch_keras_layers(),
	include_top=False,
	input_shape=input_shape,
	input_tensor=x,
	weights=weights,
	**kwargs
	)

	# Determine which layers should be trainable
	n_layers = len(backbone.layers)
	if isinstance(trainable, bool):
	trainable = [trainable for _ in range(n_layers)]
	elif isinstance(trainable, Sequence):
	trainable = list(trainable)
	trainable_length = len(trainable)
	assert trainable_length <= n_layers
	if trainable_length < n_layers:
	trainable = (
	[False for _ in range(n_layers - trainable_length)]
	+
	trainable
	)

	# Set the appropriate layers to trainable
	for layer, t in zip(backbone.layers, trainable):
	layer.trainable = t

	return tf.keras.Model(
	inputs=inputs,
	outputs=backbone.outputs,
	name=self.model_name
	)