accuracy_based_auto_calibration

功能说明

根据用户输入的模型、配置文件进行自动的校准过程，搜索得到一个满足目标精度的量化配置，输出可以在ONNX Runtime环境下做精度仿真的fake_quant模型，和可在昇腾AI处理器上做推理的deploy模型。

约束说明

无。

函数原型

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

参数说明

返回值说明

无。

函数输出

• 精度仿真模型文件：模型名中包含fake_quant，可以在ONNX执行框架ONNX Runtime进行精度仿真。
• 部署模型文件：模型名中包含deploy，经过ATC转换工具转换后可部署到在昇腾AI处理器。
• 量化因子记录文件：在接口中的record_file中写入量化因子。
• 敏感度信息文件：该文件记录了待量化层对于量化的敏感度信息，根据该信息进行量化回退层的选择。
• 自动量化回退历史记录文件：记录的回退层的信息。

调用示例

import amct_onnx as amct
from amct_onnx.common.auto_calibration import AutoCalibrationEvaluatorBase

# You need to implement the AutoCalibrationEvaluator's calibration(), evaluate() and metric_eval() funcs
class AutoCalibrationEvaluator(AutoCalibrationEvaluatorBase):
    """ subclass of AutoCalibrationEvaluatorBase"""
    def __init__(self, target_loss, batch_num):
        super(AutoCalibrationEvaluator, self).__init__()
        self.target_loss = target_loss
        self.batch_num = batch_num

    def calibration(self, model_file):
        """ implement the calibration function of AutoCalibrationEvaluatorBase
            calibration() need to finish the calibration inference procedure
            so the inference batch num need to >= the batch_num pass to create_quant_config
        """
        onnx_forward(onnx_model=model_file, batch_size=32, iterations=self.batch_num)

    def evaluate(self, model_file):
        """ implement the evaluate function of AutoCalibrationEvaluatorBase
            params: model_file in .onnx
            return: the accuracy of input model on the eval dataset, or other metric which
                    can describe the 'accuracy' of model
        """
        top1, top5 = onnx_forward(onnx_model=model_file, batch_size=32, iterations=5)
        return top1

    def metric_eval(self, original_metric, new_metric):
        """ implement the metric_eval function of AutoCalibrationEvaluatorBase
            params: original_metric: the returned accuracy of evaluate() on non quantized model
                    new_metric: the returned accuracy of evaluate() on fake quant model
            return:
                   [0]: whether the accuracy loss between non quantized model and fake quant model
                        can satisfy the requirement
                   [1]: the accuracy loss between non quantized model and fake quant model
        """
        loss = original_metric - new_metric
        if loss * 100 < self.target_loss:
            return True, loss
        return False, loss
   ...
 
    config_json_file = os.path.join(TMP, 'config.json')
    skip_layers = []
    batch_num = 1
    amct.create_quant_config(
        config_file=config_json_file, model_file=model_file, skip_layers=skip_layers, batch_num=batch_num,
        activation_offset=True, config_defination=None)

    # 1. step1 create quant config json file
    scale_offset_record_file = os.path.join(TMP, 'scale_offset_record.txt')
    result_path = os.path.join(PATH, 'results/mobilenet_v2')

    # 2. step2 construct the instance of AutoCalibrationEvaluator
    evaluator = AutoCalibrationEvaluator(target_loss=0.5, batch_num=batch_num)

    # 3. step3 using the accuracy_based_auto_calibration to quantized the model
    amct.accuracy_based_auto_calibration(
        model_file=model_file,
        model_evaluator=evaluator,
        config_file=config_json_file,
        record_file=scale_offset_record_file,
        save_dir=result_path,
        strategy='BinarySearch',
        sensitivity='CosineSimilarity'
    )