极智AI | 再谈昇腾CANN量化

# (1) Anacona3安装，用conda管理python环境，略过
# (2) conda创建amct环境，python版本3.7.5
conda create -n amct_py375 python=3.7.5
# (3) 激活conda环境
conda activate amct_py375
# (4) 安装onnx=1.8.0，onnxruntime=1.6.0，protobuf=3.11.3
pip install -i https://pypi.douban.com/simple onnx=1.8.0 onnxruntime=1.6.0 protobuf=3.11.3
# (5) 下载Ascend-cann-amct_{software version}_{arch}.tar.gz并解压
tar -zxvf Ascend-cann-amct_{software version}_{arch}.tar.gz
# (6) 安装amct_onnx-{version}-py3-none-linux_{arch}.whl ==> 以onnx为例
pip install amct_onnx-{version}-py3-none-linux_{arch}.whl
# (7) 编译并安装自定义算子包
tar -zxvf amct_onnx_op.tar.gz
## 解压后目录
## amct_onnx_op
## |--inc
## |--src
## |--setup.py
cd amct_onnx_op
python setup.py build
# (8) 验证amct_onnx是否安装成功
amct_onnx

import os
import numpy as np
import cv2
PATH = os.path.realpath('./')
IMAGE_PATH = os.path.join(PATH, './images')
BIN_PATH = os.path.join(PATH, './calibration')
BIN_FILE = os.path.join(BIN_PATH, 'calibration.bin')
# 前处理参数
MEAN = [0.485, 0.456, 0.406] 
STD = [0.229, 0.224, 0.225]
CALIBRATION_SIZE = 16
def get_labels_from_txt(label_file):
    """Read all images' name and label from label_file"""
    image_names = []
    labels = []
    label_file = os.path.realpath(label_file)
    with open(label_file, 'r') as fid:
        lines = fid.readlines()
        for line in lines:
            image_names.append(line.split(' ')[0])
            labels.append(int(line.split(' ')[1]))
    return image_names, labels
# h、w可自定义
def prepare_image_input(images, height=256, width=256, crop_size=224):
    """Read image files to blobs [batch_size, 3, 224, 224]"""
    input_array = np.zeros((len(images), 3, crop_size, crop_size), np.float32)
    mean = MEAN
    std = STD
    imgs = np.zeros((len(images), 3, height, width), np.float32)
    for index, im_file in enumerate(images):
        im_data = cv2.imread(im_file)
        im_data = cv2.resize(
            im_data, (256, 256), interpolation=cv2.INTER_CUBIC)
        cv2.cvtColor(im_data, cv2.COLOR_BGR2RGB)
        imgs[index] = im_data.transpose(2, 0, 1).astype(np.float32)
    h_off = int((height - crop_size) / 2)
    w_off = int((width - crop_size) / 2)
    input_array = imgs[:, :, h_off:(h_off + crop_size),
                        w_off:(w_off + crop_size)]
    # trans uint8 image data to float
    input_array /= 255
    # do channel-wise reduce mean value
    for channel in range(input_array.shape[1]):
        input_array[:, channel, :, :] -= mean[channel]
    # do channel-wise divide std
    for channel in range(input_array.shape[1]):
        input_array[:, channel, :, :] /= std[channel]
    return input_array
def process_data():
    """process data"""
    # prepare cur batch data
    image_names, labels = get_labels_from_txt(
        os.path.join(IMAGE_PATH, 'image_label.txt'))
    if len(labels) < CALIBRATION_SIZE:
        raise RuntimeError(
            'num of image in {} is less than total_num{}'
            .format(IMAGE_PATH, CALIBRATION_SIZE))
    labels = labels[0:CALIBRATION_SIZE]
    image_names = image_names[0:CALIBRATION_SIZE]
    image_names = [
        os.path.join(IMAGE_PATH, image_name) for image_name in image_names
    ]
    input_array = prepare_image_input(image_names)
    return input_array
def main():
    """process image and save it to bin"""
    input_array = process_data()
    if not os.path.exists(BIN_PATH):
        os.mkdir(BIN_PATH)
    input_array.tofile(BIN_FILE)
if __name__ == '__main__':
    main()

amct_onnx calibration --model {model}.onnx --input_shape "input:1,3,224,224" --data_type "float32" --data_dir ./calibration/ --save_path ./results