quantize_model_ascend

Function Usage

Quantizes the input graph structure to be quantized based on the given quantization configuration file, inserts quantization operators into the input graph structure, generates the quantization factor record file record_file, and returns the list of new quantization operators.

Prototype

calibration_graph, calibration_outputs=quantize_model_ascend(graph, outputs, config_file, record_file)

Command-Line Options

Option	Input/Return	Meaning	Restriction
graph	Input	A tf.Graph of the model to be quantized.	A tf.Graph. Constraints: The graph must be an inference graph and cannot contain operators in training mode. For example, is_training of the FusedBatchNormV3 operator must be False. The trained weights must be loaded in the graph.
outputs	Input	List of output operators of the graph.	A list of strings.
config_file	Input	User-defined quantization configuration file, which specifies the configuration of each layer to be quantized in the tf.Graph.	A string
record_file	Input	Directory of the quantization factor record file, including the file name.	A string
calibration_graph	Returns	The quantization operator is inserted into the graph modified by the tool.	A tf.Graph.
calibration_outputs	Return	Output operator list of calibration_graph.	A list of strings.

Return Value

List of quantized layers on the network.

quantize_model_ascend performs BN fusion on the graph. If the outputs of the network model contain the BN layer and the BN layer is also fused, the output node of the network changes. For example, Conv+BN (or Conv+BiasAdd+BN) is fused into Conv+BiasAdd, and the output node equivalent to BN is BiasAdd.

Example

import amct_tensorflow as amct
# Build a network to be quantized.
network = build_network()

# Insert the quantization API.
calibration_graph, calibration_outputs = amct.quantize_model_ascend(
      graph=tf.get_default_graph(),
      config_file="./configs/config.json",
      record_file="./record_scale_offset.txt")

Parent topic: PTQ APIs