NeuralNetwork -- for the fine-tunning it is impossible to pick a default embedding/vocabulary.

[nicolay-r]

We need there provide a check when the target path is None.
The same is for vocabulary.
Need to perfrorm refactoring at least for a better logic understanding.
Keep the single-time declared logic both for vocabulary and embedding.

AREkit/arekit/contrib/utils/model_io/tf_networks.py

Lines 148 to 159 in 4667a54

	def __get_term_embedding_source(self):
	""" It is possible to load a predefined embedding from another experiment
	using the related filepath provided by model_io.
	"""
	model_io = self._exp_ctx.ModelIO
	if model_io is None:
	return self.__get_default_embedding_filepath()
	assert(isinstance(model_io, NeuralNetworkModelIO))
	return model_io.get_model_embedding_filepath() if self.__model_is_pretrained_state_provided(model_io) \
	else self.__get_default_embedding_filepath()