accel_model Module

Module with abstract base class of HW accelerated models.

Classes

AccelModel

Abstract base class for accel models. Accelerated models are neural networks that can be accelerated using dedicated hardware. Abstract base class for anaccel model. To add a new model, implement a subclass - model folder_name, version, checkpoint_uri and save_name should all be given by the user. Weight path and is frozen should be exposed at least for quantized versions. :param model_base_path: The base path to store all models in. Generally given by the user. :param model_folder_name: The path on disk to store all versions of this model. :param version: The version of this model. :param check_point_uri: The URI where the model is downloaded from if they don't have it on disk. :param save_name: The name the checkpoint is saved under, used to load metagraph. :param is_frozen: If the model should be frozen when it is loaded. This freezes the graph by removing the variables from tf.GraphKeys.TRAINABLE_VARIABLES. :param weight_path: A custom path to load weights from, instead of the default path on disk. Used in retraining scenarios.