auto_channel_prune_search

Applicability

Product	Supported
Atlas 350 Accelerator Card	√
Atlas A3 training product/Atlas A3 inference product	√
Atlas A2 training product/Atlas A2 inference product	√
Atlas 200I/500 A2 inference product	√
Atlas inference product	√
Atlas training product	√

Description

Calculates the sparsity sensitivity (affecting accuracy) and sparsity gain (affecting performance) of each channel based on the user model. Then, the search policy searches for the optimal layer-wise channel sparsity ratio based on the input to balance accuracy and performance. Finally, a configuration file is generated.

Prototype

auto_channel_prune_search(graph, output_nodes, config, input_data, output_cfg, sensitivity, search_alg)

Parameters

Parameter	Input/Output	Description
graph	Input	tf.Graph training graph to be sparsified, which contains automatic differentiation. A tf.Graph.
output_nodes	Input	Name of the model output node. A list of strings.
config	Input	Path of the configuration file for auto channel pruning search. The simplified configuration file is generated based on AutoChannelPruneConfig in the basic_info.proto file. The .proto file is stored in /amct_tensorflow/proto/* under the AMCT installation directory. For details about the parameters in the *.proto file and the generated auto channel pruning search configuration file, see Simplified Configuration File for Auto Channel Pruning Search. A string.
input_data	Input	Calibration data provided by the user. A list. The content is the corresponding feed_dict data.
output_cfg	Input	Path of the output channel pruning configuration file. A string.
sensitivity	Input	Sensitivity calculation method. A string or subclass of SensitivityBase. The string is an existing method of AMCT, which can be set to TaylorLossSensitivity. The subclass of SensitivityBase can be instantiated and defined by the user. Defaults to TaylorLossSensitivity.
search_alg	Input	Method of searching for channels to be sparsified. A string or subclass of SearchChannelBase. The string is an existing method of AMCT, which can be set to GreedySearch. The subclass of SearchChannelBase can be instantiated and defined by the user. The default value is GreedySearch.

Returns

None

Example

import amct_tensorflow as amct
# Construct feed_dict data.
input_data = np.random.uniform(-10, 10, (2, 14, 14, 64)).astype(np.str_)
feed_dict = [{'input:0': input_data}]

amct.auto_channel_prune_search(
     graph=graph, 
     output_nodes=[operation_name_1, operation_name_2],
     config='./tmp/sample.cfg', 
     input_data=feed_dict,  
     output_cfg='./tmp/output.cfg', 
     sensitivity='TaylorLossSensitivity', 
     search_alg='GreedySearch')

Flush file: auto channel pruning configuration file, which needs to be passed to the channel pruning API for subsequent services.

Parent topic: Auto Channel Pruning Search API