""" #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"] = layerg2 = 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 = oldparamprint ('-----------------------------------------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"] = layerf.close()f2.close() old_model_path = './yolox_s.h5' new_model_path = './yolox_sq.h5' quantization16bit(old_model_path,new_model_path,8) # print (f['batch_normalization']['batch_normalization']['gamma:0'][:])

總結

以上是生活随笔為你收集整理的keras 模型量化的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。