keras模型量化

模型量化的本質就是將模型中的參數按照一定的規則 把32位或者64位浮點數 轉化位16位浮點數或者8位定點數。這里我用keras和numpy實現了16位和8位的量化，未考慮量化的科學合理性，僅僅是搞清楚量化本質的一次實驗。

量化


"""
#coding:utf-8
__project_ = 'TF2learning'
__file_name__ = 'quantization'
__author__ = 'qilibin'
__time__ = '2021/3/17 9:18'
__product_name = PyCharm
"""
import h5py
import pandas as pd
import numpy as np


'''
讀取原來的只包含權重的H5模型，按層遍歷，對每層的每個權重進行16位或8位量化，將量化后的權重數值重新保存在H5文件中
'''


def quantization16bit(old_model_path,new_model_path,bit_num):
    '''

    :param old_model_path: 未量化的模型路徑  模型是只保存了權重未保存網絡結構
    :param new_model_path: 量化過后的模型路徑
    :param bit_num: 量化位數
    :return:
    '''


    f = h5py.File(old_model_path,'r')
    f2 = h5py.File(new_model_path,'w')
    for  layer in f.keys():
        # layer ： 層的名稱

        print (layer)

        # # 每層里面的權重名稱 有的層沒有參數
        # name_of_weight_of_layer = f[layer].attrs['weight_names']
        # # 有的層是沒有參數的 比如 relu
        # length = len(name_of_weight_of_layer)

        length = len(list(f[layer].keys()))
        if length > 0:
            g1 = f2.create_group(layer)
            g1.attrs["weight_names"] = layer
            g2 = g1.create_group(layer)

            for weight in f[layer][layer].keys():
                print ("wieght name is :" + weight)
                oldparam = f[layer][layer][weight][:]
                print ('-----------------------------------------old-----------------------')
                print (oldparam)

                if type(oldparam) == np.ndarray:
                    if bit_num == 16:
                        newparam = np.float16(oldparam)
                    if bit_num == 8:
                        min_val = np.min(oldparam)
                        max_val = np.max(oldparam)
                        oldparam = np.round((oldparam - min_val) / (max_val - min_val) * 255)
                        newparam = np.uint8(oldparam)
                else:
                    newparam = oldparam
                print ('-----------------------------------------new-----------------------')
                #print (newparam)
                #f[key][key][weight_name][:] = newparam  在原來模型的基礎上修改 行不通
                if bit_num == 16:
                    d = g2.create_dataset(weight, data=newparam,dtype=np.float16)
                if bit_num == 8:
                    d = g2.create_dataset(weight, data=newparam, dtype=np.uint8)

        else:
            g1 = f2.create_group(layer)
            g1.attrs["weight_names"] = layer
    f.close()
    f2.close()
old_model_path = './model_0_.h5'
new_model_path = './new_model.h5'
quantization16bit(old_model_path,new_model_path,8)
# print (f['batch_normalization']['batch_normalization']['gamma:0'][:])

檢查量化后的文件

"""
#coding:utf-8
__project_ = 'TF2learning'
__file_name__ = 'readNewMoDel'
__author__ = 'qilibin'
__time__ = '2021/3/17 13:27'
__product_name = PyCharm
"""
'''
用來打印量化之后的模型 查看其各個權重的參數
'''
import h5py

modelpath = './new_model.h5'
#modelpath = './model_0_.h5'


f = h5py.File(modelpath,'r')

for layer in f.keys():
    # key ： 層的名稱

    print ("layer name is :"+layer)

    # 有些層是沒有參數的 比如relu
    length = len(list(f[layer].keys()))
    #print (length)
    if length > 0:
        for weight in f[layer][layer].keys():
            print("wieght name is :" + weight)
            param = f[layer][layer][weight][:]
            print(param)


f.close()

# print (f['batch_normalization']['batch_normalization']['gamma:0'][:])