accuracy_based_auto_calibration

Applicability

Product	Supported
Atlas A3 training series products/Atlas A3 inference series products	√
Atlas A2 training products/Atlas A2 inference products	√
Atlas 200I/500 A2 inference product	√
Atlas inference series products	√
Atlas training products	√

Description

Calibrates the input model based on the input configuration file to search for a quantization configuration that meets accuracy requirements, and outputs a model suitable for both accuracy simulation in the TensorFlow environment and inference deployment on the Ascend AI Processor.

Prototype

accuracy_based_auto_calibration(model_file,outputs,record_file,config_file,save_dir,evaluator,strategy='BinarySearch',sensitivity='CosineSimilarity')

Parameters

Parameter	Input/Output	Description
model_file	Input	Definition file (.pb) of the dequantized TensorFlow model. A string.
outputs	Input	String list of the output node. A string.
config_file	Input	Quantization configuration file generated by the user. A string.
record_file	Input	Path of the quantization factor record file. The existing file (if any) in the path will be overwritten. A string.
save_dir	Input	Model save path. Must include the prefix of the model name, for example, *./quantized_model/model**. A string.
evaluator	Input	Python instance for automatic quantization calibration and accuracy evaluation. A Python instance.
strategy	Input	Policy for searching for the quantization configuration that meets the accuracy requirements. The dichotomy policy is used by default. A string or a Python instance. Default: BinarySearch
sensitivity	Input	Metric used to evaluate how quantization-sensible each layer to be quantized is. By default, the cosine similarity metric is used. A string or a Python instance. Default: CosineSimilarity

Returns

None

Example

import amct_tensorflow as amct
from amct_tensorflow.accuracy_based_auto_calibration import accuracy_based_auto_calibration

def main():
    args_check(args)
    outputs = [PREDICTIONS]
    record_file = os.path.join(RESULT_DIR, 'record.txt')
    config_file = os.path.join(RESULT_DIR, 'config.json')
    with tf.io.gfile.GFile(args.model, mode='rb') as model:
        graph_def = tf.compat.v1.GraphDef()
        graph_def.ParseFromString(model.read())
    tf.import_graph_def(graph_def, name='')
    graph = tf.compat.v1.get_default_graph()
    amct.create_quant_config(config_file, graph)
    save_dir = os.path.join(RESULT_DIR, 'MobileNetV2')
    evaluator = MobileNetV2Evaluator(args.dataset, args.keyword, args.num_parallel_reads, args.batch_size)
    accuracy_based_auto_calibration(args.model, outputs, record_file, config_file, save_dir, evaluator)

Flush files:

A .pb model file that can be used for accuracy simulation in the TensorFlow environment or inference on the Ascend AI Processor.
A quantization factor record file, a quantization configuration file, a layer similarity result file, and an automatic unquantization history file.

Parent topic: PTQ APIs