AccelModel Class

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.

Inheritance

ABC

AccelModel

Constructor

AccelModel(model_base_path, model_folder_name, version, check_point_uri, save_name, is_frozen=False, weight_path=None)

Parameters

Name	Description
model_base_path Required
model_folder_name Required
version Required
check_point_uri Required
save_name Required
is_frozen	default value: False
weight_path	default value: None

Methods

get_default_classifier	Import a frozen, default Imagenet classifier for the model into the current graph.
get_input_dims	Get nth model input tensor dimensions.
get_output_dims	Get nth model output tensor dimensions.
import_graph_def	Import the graph definition corresponding to this model. Imports accelerated model into currently active graph.
restore_weights	Restore the weights of the model into the specific session.
save_weights	Save the weights of the model from a specific session into a specific path.

get_default_classifier

Import a frozen, default Imagenet classifier for the model into the current graph.

get_default_classifier(input_tensor, prefix='classifier')

Parameters

Name	Description
prefix	namespace to load classifier into. default value: classifier
input_tensor Required	The input feature tensor for the classifier. Expected to be [?, 2048]
model_dir Required	The directory to download the classifier into. Used as a cache locally.

get_input_dims

Get nth model input tensor dimensions.

abstract get_input_dims(index=0)

Parameters

Name	Description
index	default value: 0

get_output_dims

Get nth model output tensor dimensions.

abstract get_output_dims(index=0)

Parameters

Name	Description
index	default value: 0

import_graph_def

Import the graph definition corresponding to this model.

Imports accelerated model into currently active graph.

import_graph_def(input_tensor=None, is_training=True)

Parameters

Name	Description
input_tensor	Replace input tensor to accelerated model (must match expected shape and dtype) default value: None
is_training	Boolean indicating if the imported graph is intending for training. default value: True

Returns

Type	Description
	Either single output tensor or list of output tensors (if more than one).

restore_weights

Restore the weights of the model into the specific session.

restore_weights(session)

Parameters

Name	Description
session Required	<xref:tf.Session> The session to load the weights into.

save_weights

Save the weights of the model from a specific session into a specific path.

save_weights(path, session=None)

Parameters

Name	Description
path Required	Path of the checkpoint to save the weights into.
session	<xref:tf.Session> Session to save weights from. default value: None

Attributes

input_tensor_list

List of names of the input tensors of this model.

model_path

Path to directory that contains the model.

model_ref

Name that refers to the model - used for writing the model_def.

model_version

Model Version.

output_tensor_list

List of names of the output tensors of this model.