计算神经网络内存占用

水深无声 2022-05-08 01:33 322阅读 0赞

如果layer.count\_params()或者model.summary()权重参数个数为负数的可以参考:  
[彻底解决keras model.summary()或者layer.count\_params()权重参数个数为负数问题][keras model.summary_layer.count_params]  
举两个例子，一张(1024,1024,3)的图片，分别使用VGGNet,DenseNet网络。

一.VGGNet内存消耗计算

''' Created on 2018年9月30日 ''' from keras import applications import numpy as np import cv2 image = cv2.imread("D:\\xxxx\\hashiqi.jpg") image = cv2.resize(image,(1024,1024),interpolation = cv2.INTER_CUBIC) x_train = np.expand_dims(image,axis=0) y_train = np.array([0]) print(image.shape) # (1024, 1024, 3) model = applications.VGG16(input_shape=(1024,1024,3),include_top=False,weights=None) print("无全连接层总参数量:",model.count_params()) model = applications.VGG16(input_shape=(1024,1024,3),include_top=True,weights=None) print("有全连接层总参数量:",model.count_params()) # 无全连接层总参数量: 14714688 # 有全连接层总参数量: 2183080744 可见权重都基本占用在全连接层 all_params_memory = 0 all_feature_memory = 0 for layer in model.layers: #训练权重w占用的内存 params_memory = layer.count_params()/(1024*1024) * 4 print("训练权重w占用的内存:",layer.name,layer.count_params(),str(params_memory)+" M") all_params_memory = all_params_memory + params_memory #特征图占用内存 feature_shape = layer.output_shape feature_size = 1 for i in range(1,len(feature_shape)): feature_size = feature_size*feature_shape[i] feature_memory = feature_size/(1024*1024) * 4 print("特征图占用内存:",feature_shape,feature_size,str(feature_memory)+" M") all_feature_memory = all_feature_memory + feature_memory # 特征图占用内存: (None, 1024, 1024, 3) 3145728 12.0 M # 训练权重w占用的内存: block1_conv1 1792 0.0068359375 M # 特征图占用内存: (None, 1024, 1024, 64) 67108864 256.0 M # 训练权重w占用的内存: block1_conv2 36928 0.140869140625 M # 特征图占用内存: (None, 1024, 1024, 64) 67108864 256.0 M # 训练权重w占用的内存: block1_pool 0 0.0 M # 特征图占用内存: (None, 512, 512, 64) 16777216 64.0 M # 训练权重w占用的内存: block2_conv1 73856 0.28173828125 M # 特征图占用内存: (None, 512, 512, 128) 33554432 128.0 M # 训练权重w占用的内存: block2_conv2 147584 0.56298828125 M # 特征图占用内存: (None, 512, 512, 128) 33554432 128.0 M # 训练权重w占用的内存: block2_pool 0 0.0 M # 特征图占用内存: (None, 256, 256, 128) 8388608 32.0 M # 训练权重w占用的内存: block3_conv1 295168 1.1259765625 M # 特征图占用内存: (None, 256, 256, 256) 16777216 64.0 M # 训练权重w占用的内存: block3_conv2 590080 2.2509765625 M # 特征图占用内存: (None, 256, 256, 256) 16777216 64.0 M # 训练权重w占用的内存: block3_conv3 590080 2.2509765625 M # 特征图占用内存: (None, 256, 256, 256) 16777216 64.0 M # 训练权重w占用的内存: block3_pool 0 0.0 M # 特征图占用内存: (None, 128, 128, 256) 4194304 16.0 M # 训练权重w占用的内存: block4_conv1 1180160 4.501953125 M # 特征图占用内存: (None, 128, 128, 512) 8388608 32.0 M # 训练权重w占用的内存: block4_conv2 2359808 9.001953125 M # 特征图占用内存: (None, 128, 128, 512) 8388608 32.0 M # 训练权重w占用的内存: block4_conv3 2359808 9.001953125 M # 特征图占用内存: (None, 128, 128, 512) 8388608 32.0 M # 训练权重w占用的内存: block4_pool 0 0.0 M # 特征图占用内存: (None, 64, 64, 512) 2097152 8.0 M # 训练权重w占用的内存: block5_conv1 2359808 9.001953125 M # 特征图占用内存: (None, 64, 64, 512) 2097152 8.0 M # 训练权重w占用的内存: block5_conv2 2359808 9.001953125 M # 特征图占用内存: (None, 64, 64, 512) 2097152 8.0 M # 训练权重w占用的内存: block5_conv3 2359808 9.001953125 M # 特征图占用内存: (None, 64, 64, 512) 2097152 8.0 M # 训练权重w占用的内存: block5_pool 0 0.0 M # 特征图占用内存: (None, 32, 32, 512) 524288 2.0 M # 训练权重w占用的内存: flatten 0 0.0 M # 特征图占用内存: (None, 524288) 524288 2.0 M # 训练权重w占用的内存: fc1 2147487744 8192.015625 M # 特征图占用内存: (None, 4096) 4096 0.015625 M # 训练权重w占用的内存: fc2 16781312 64 .015625 M # 特征图占用内存: (None, 4096) 4096 0.015625 M # 训练权重w占用的内存: predictions 4097000 15.628814697265625 M # 特征图占用内存: (None, 1000) 1000 0.003814697265625 M print("网络权重W占用总内存:",str(all_params_memory)+" M") print("网络特征图占用总内存:",str(all_feature_memory)+" M") print("网络总消耗内存:",str(all_params_memory+all_feature_memory)+" M") # 网络权重W占用总内存: 8327.79214477539 M # 网络特征图占用总内存: 1216.0350647 M # 网络总消耗内存: 9543.82720947 M

二.DenseNet内存消耗计算

''' Created on 2018年9月30日 ''' from keras import applications import numpy as np import cv2 image = cv2.imread("D:\\xxx\\hashiqi.jpg") image = cv2.resize(image,(1024,1024),interpolation = cv2.INTER_CUBIC) x_train = np.expand_dims(image,axis=0) y_train = np.array([0]) print(image.shape) # (1024, 1024, 3) model = applications.DenseNet201(input_shape=(1024,1024,3),include_top=False,weights=None) print("无全连接层总参数量:",model.count_params()) model = applications.DenseNet201(input_shape=(1024,1024,3),include_top=True,weights=None) print("有全连接层总参数量:",model.count_params()) # 无全连接层总参数量: 18321984 # 有全连接层总参数量: 20242984 因为使用了GlobalAveragePooling2D，使得全连接参数少了很多 all_params_memory = 0 all_feature_memory = 0 for layer in model.layers: #训练权重w占用的内存 params_memory = layer.count_params()/(1024*1024) * 4 print("训练权重w占用的内存:",layer.name,layer.count_params(),str(params_memory)+" M") all_params_memory = all_params_memory + params_memory #特征图占用内存 feature_shape = layer.output_shape feature_size = 1 for i in range(1,len(feature_shape)): feature_size = feature_size*feature_shape[i] feature_memory = feature_size/(1024*1024) * 4 print("特征图占用内存:",feature_shape,feature_size,str(feature_memory)+" M") all_feature_memory = all_feature_memory + feature_memory # 训练权重w占用的内存: input_2 0 0.0 M # 特征图占用内存: (None, 1024, 1024, 3) 3145728 12.0 M # 训练权重w占用的内存: zero_padding2d_3 0 0.0 M # 特征图占用内存: (None, 1030, 1030, 3) 3182700 12.141036987304688 M # 训练权重w占用的内存: conv1/conv 9408 0.035888671875 M # 特征图占用内存: (None, 512, 512, 64) 16777216 64.0 M # 训练权重w占用的内存: conv1/bn 256 0.0009765625 M # 特征图占用内存: (None, 512, 512, 64) 16777216 64.0 M # 训练权重w占用的内存: conv1/relu 0 0.0 M # 特征图占用内存: (None, 512, 512, 64) 16777216 64.0 M # ......... print("网络权重W占用总内存:",str(all_params_memory)+" M") print("网络特征图占用总内存:",str(all_feature_memory)+" M") print("网络总消耗内存:",str(all_params_memory+all_feature_memory)+" M") # 网络权重W占用总内存: 77.22085571289062 M # 网络特征图占用总内存: 6151.65315246582 M # 网络总消耗内存: 6228.874008178711 M 内存消耗基本在特征图上

[keras model.summary_layer.count_params]: https://blog.csdn.net/u011311291/article/details/82968945