MobileNetV2 leaning notes

最新推荐文章于 2024-08-12 18:49:54 发布
最新推荐文章于 2024-08-12 18:49:54 发布
阅读量210 收藏
点赞数
文章标签： 深度学习 网络
版权声明：本文为博主原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接和本声明。
本文链接：https://blog.csdn.net/qq_40776179/article/details/106085415
版权

Directory

Features and benefits of MobileNet

Synopsis

MobileNet is a compact and efficient CNN model proposed by Google, which makes a compromise between accuracy and latency. The model is suitable for use on mobile devices and meets the real-time requirement in daily life.

Features

There are two innovative points in the proposal of MobileNetV1:

  1. Depthwise Convolution.
  2. Superparameters α and β.

The characteristic of Depthwise convolution is that the size of the convolution kernel is 1, and the number of the input eigenmatrix is the number of the convolution kernel and the number of the output eigenmatrix. In general, Depthwise convolution convolved with Pointwise convolution. Pointwise convolution is actually ordinary convolution, but it adopts 1x1 convolution kernel. And that makes them 1/8 as much computation as normal convolution. What’s more, we can get a smaller model by adjusting the values of α and β.
在这里插入图片描述
MobileNetV2 makes two improvements over MobileNetV1:

  1. Inverted Resduals.
  2. Linear Bottlenecks.

MobileNetV1 uses the residual block, while MobileNetV2 uses the inverted residual block. This structure is just the opposite of the traditional residual block in which the dimension is first reduced and then amplified, so the shotcut also becomes the feature map connected with the dimension reduced. Linear Bottlenecks. In other words, the nonlinear activation layer behind the small dimension output layer was removed to ensure the expression ability of the model.
在这里插入图片描述

The structure of MobileNetV2

在这里插入图片描述
This is the basic structure of MobileNetV2 in the paper, and t is extended factor, and n is the bottleneck of repetitions. The inverted residual block, bottleneck, is as follows:
在这里插入图片描述
Among them, it is worth noting that stride of the bottleneck structure is only be set in the first time, the rest of the stride is 1. As for the applicable condition of shortcut, the paper only indicates the condition of stride = 1. However, when the code is implemented, I find that the result is inconsistent with the structure diagram. Therefore, I infer that this is an error in the paper. And then, after analysis, I find that if there is no shortcut, it will not affect the result. Therefore, it is inferred that the condition of n > 1 is still missing.

Using keras to implemented MobileNetV2

Code:

from keras.models import Model
from keras.layers import Input, Conv2D, GlobalAveragePooling2D, Dropout, DepthwiseConv2D
from keras.layers import Activation, BatchNormalization, Reshape, add
from keras import backend
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'


def relu6(x):
    relu_6 = backend.relu(x, max_value=6)

    return relu_6


def _conv_layer(inputs, filters, kernel, strides):
    channel_axis = 1 if backend.image_data_format() == 'channels_first' else -1

    x = Conv2D(filters, kernel, padding='same', strides=strides)(inputs)
    x = BatchNormalization(axis=channel_axis)(x)
    x = Activation(relu6)(x)

    return x


def _bottleneck(inputs, filters, kernel, t, strides, r=False):
    channel_axis = 1 if backend.image_data_format() == 'channels_first' else -1
    t_channel = backend.int_shape(inputs)[channel_axis] * t

    # input shape (h, w, k)
    # Convolution layer 1 output shape (h, w, tk)
    x = _conv_layer(inputs, t_channel, (1, 1), (1, 1))

    # DW convolution layer 2 output shape (h/s, w/s, tk)
    x = DepthwiseConv2D(kernel, strides=(strides, strides), depth_multiplier=1, padding='same')(x)
    x = BatchNormalization(axis=channel_axis)(x)
    x = Activation(relu6)(x)

    # Convolution layer 3 output shape (h/s, w/s, filters)
    x = Conv2D(filters, (1, 1), strides=(1, 1), padding='same')(x)
    x = BatchNormalization(axis=channel_axis)(x)
    x = Activation('linear')(x)

    if r:
        x = add([x, inputs])

    return x


def _bottleneck_layer(inputs, filters, kernel, t, strides, n):
    x = _bottleneck(inputs, filters, kernel, t, strides)

    for i in range(1, n):
        # In addition to the first layer, other bottleneck layers of strides is 1
        x = _bottleneck(x, filters, kernel, t, 1, True)

    return x


def MobileNetv2(input_shape, k):
    # Assume that the input shape is (224, 224, 3)
    inputs = Input(shape=input_shape)
    # Convolution layer 1 output shape (112, 112, 32)
    x = _conv_layer(inputs, 32, (3, 3), strides=(2, 2))

    # Bottleneck layer 2 output shape (112, 112, 16)
    x = _bottleneck_layer(x, 16, (3, 3), t=1, strides=1, n=1)
    # Bottleneck layer 3 output shape (56, 56, 24)
    x = _bottleneck_layer(x, 24, (3, 3), t=6, strides=2, n=2)
    # Bottleneck layer 4 output shape (28, 28, 32)
    x = _bottleneck_layer(x, 32, (3, 3), t=6, strides=2, n=3)
    # Bottleneck layer 5 output shape (14, 14, 64)
    x = _bottleneck_layer(x, 64, (3, 3), t=6, strides=2, n=4)
    # Bottleneck layer 6 output shape (14, 14, 96)
    x = _bottleneck_layer(x, 96, (3, 3), t=6, strides=1, n=3)
    # Bottleneck layer 7 output shape (7, 7, 160)
    x = _bottleneck_layer(x, 160, (3, 3), t=6, strides=2, n=3)
    # Bottleneck layer 8 output shape (7, 7, 320)
    x = _bottleneck_layer(x, 320, (3, 3), t=6, strides=1, n=1)

    # Convolution layer 9 output shape (7, 7, 1280)
    x = _conv_layer(x, 1280, (1, 1), strides=(1, 1))

    # Averagepooling layer 10 output shape (1, 1, 1280)
    x = GlobalAveragePooling2D()(x)
    x = Reshape((1, 1, 1280))(x)
    x = Dropout(0.3, name='Dropout')(x)

    # Convolution layer 10 output class k
    x = Conv2D(k, (1, 1), padding='same')(x)
    x = Activation('softmax', name='softmax')(x)
    outputs = Reshape((k,))(x)

    model = Model(inputs, outputs)

    return model


if __name__ == '__main__':
    # Take the ImageNet dataset as an example
    model = MobileNetv2((224, 224, 3), 1000)
    model.summary()

Result: Since the data set of ImageNet cannot be downloaded from the official website, the relevant experiment cannot be carried out temporarily.

Model: "model_1"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            (None, 224, 224, 3)  0                                            
__________________________________________________________________________________________________
conv2d_1 (Conv2D)               (None, 112, 112, 32) 896         input_1[0][0]                    
__________________________________________________________________________________________________
batch_normalization_1 (BatchNor (None, 112, 112, 32) 128         conv2d_1[0][0]                   
__________________________________________________________________________________________________
activation_1 (Activation)       (None, 112, 112, 32) 0           batch_normalization_1[0][0]      
__________________________________________________________________________________________________
conv2d_2 (Conv2D)               (None, 112, 112, 32) 1056        activation_1[0][0]               
__________________________________________________________________________________________________
batch_normalization_2 (BatchNor (None, 112, 112, 32) 128         conv2d_2[0][0]                   
__________________________________________________________________________________________________
activation_2 (Activation)       (None, 112, 112, 32) 0           batch_normalization_2[0][0]      
__________________________________________________________________________________________________
depthwise_conv2d_1 (DepthwiseCo (None, 112, 112, 32) 320         activation_2[0][0]               
__________________________________________________________________________________________________
batch_normalization_3 (BatchNor (None, 112, 112, 32) 128         depthwise_conv2d_1[0][0]         
__________________________________________________________________________________________________
activation_3 (Activation)       (None, 112, 112, 32) 0           batch_normalization_3[0][0]      
__________________________________________________________________________________________________
conv2d_3 (Conv2D)               (None, 112, 112, 16) 528         activation_3[0][0]               
__________________________________________________________________________________________________
batch_normalization_4 (BatchNor (None, 112, 112, 16) 64          conv2d_3[0][0]                   
__________________________________________________________________________________________________
activation_4 (Activation)       (None, 112, 112, 16) 0           batch_normalization_4[0][0]      
__________________________________________________________________________________________________
conv2d_4 (Conv2D)               (None, 112, 112, 96) 1632        activation_4[0][0]               
__________________________________________________________________________________________________
batch_normalization_5 (BatchNor (None, 112, 112, 96) 384         conv2d_4[0][0]                   
__________________________________________________________________________________________________
activation_5 (Activation)       (None, 112, 112, 96) 0           batch_normalization_5[0][0]      
__________________________________________________________________________________________________
depthwise_conv2d_2 (DepthwiseCo (None, 56, 56, 96)   960         activation_5[0][0]               
__________________________________________________________________________________________________
batch_normalization_6 (BatchNor (None, 56, 56, 96)   384         depthwise_conv2d_2[0][0]         
__________________________________________________________________________________________________
activation_6 (Activation)       (None, 56, 56, 96)   0           batch_normalization_6[0][0]      
__________________________________________________________________________________________________
conv2d_5 (Conv2D)               (None, 56, 56, 24)   2328        activation_6[0][0]               
__________________________________________________________________________________________________
batch_normalization_7 (BatchNor (None, 56, 56, 24)   96          conv2d_5[0][0]                   
__________________________________________________________________________________________________
activation_7 (Activation)       (None, 56, 56, 24)   0           batch_normalization_7[0][0]      
__________________________________________________________________________________________________
conv2d_6 (Conv2D)               (None, 56, 56, 144)  3600        activation_7[0][0]               
__________________________________________________________________________________________________
batch_normalization_8 (BatchNor (None, 56, 56, 144)  576         conv2d_6[0][0]                   
__________________________________________________________________________________________________
activation_8 (Activation)       (None, 56, 56, 144)  0           batch_normalization_8[0][0]      
__________________________________________________________________________________________________
depthwise_conv2d_3 (DepthwiseCo (None, 56, 56, 144)  1440        activation_8[0][0]               
__________________________________________________________________________________________________
batch_normalization_9 (BatchNor (None, 56, 56, 144)  576         depthwise_conv2d_3[0][0]         
__________________________________________________________________________________________________
activation_9 (Activation)       (None, 56, 56, 144)  0           batch_normalization_9[0][0]      
__________________________________________________________________________________________________
conv2d_7 (Conv2D)               (None, 56, 56, 24)   3480        activation_9[0][0]               
__________________________________________________________________________________________________
batch_normalization_10 (BatchNo (None, 56, 56, 24)   96          conv2d_7[0][0]                   
__________________________________________________________________________________________________
activation_10 (Activation)      (None, 56, 56, 24)   0           batch_normalization_10[0][0]     
__________________________________________________________________________________________________
add_1 (Add)                     (None, 56, 56, 24)   0           activation_10[0][0]              
                                                                 activation_7[0][0]               
__________________________________________________________________________________________________
conv2d_8 (Conv2D)               (None, 56, 56, 144)  3600        add_1[0][0]                      
__________________________________________________________________________________________________
batch_normalization_11 (BatchNo (None, 56, 56, 144)  576         conv2d_8[0][0]                   
__________________________________________________________________________________________________
activation_11 (Activation)      (None, 56, 56, 144)  0           batch_normalization_11[0][0]     
__________________________________________________________________________________________________
depthwise_conv2d_4 (DepthwiseCo (None, 28, 28, 144)  1440        activation_11[0][0]              
__________________________________________________________________________________________________
batch_normalization_12 (BatchNo (None, 28, 28, 144)  576         depthwise_conv2d_4[0][0]         
__________________________________________________________________________________________________
activation_12 (Activation)      (None, 28, 28, 144)  0           batch_normalization_12[0][0]     
__________________________________________________________________________________________________
conv2d_9 (Conv2D)               (None, 28, 28, 32)   4640        activation_12[0][0]              
__________________________________________________________________________________________________
batch_normalization_13 (BatchNo (None, 28, 28, 32)   128         conv2d_9[0][0]                   
__________________________________________________________________________________________________
activation_13 (Activation)      (None, 28, 28, 32)   0           batch_normalization_13[0][0]     
__________________________________________________________________________________________________
conv2d_10 (Conv2D)              (None, 28, 28, 192)  6336        activation_13[0][0]              
__________________________________________________________________________________________________
batch_normalization_14 (BatchNo (None, 28, 28, 192)  768         conv2d_10[0][0]                  
__________________________________________________________________________________________________
activation_14 (Activation)      (None, 28, 28, 192)  0           batch_normalization_14[0][0]     
__________________________________________________________________________________________________
depthwise_conv2d_5 (DepthwiseCo (None, 28, 28, 192)  1920        activation_14[0][0]              
__________________________________________________________________________________________________
batch_normalization_15 (BatchNo (None, 28, 28, 192)  768         depthwise_conv2d_5[0][0]         
__________________________________________________________________________________________________
activation_15 (Activation)      (None, 28, 28, 192)  0           batch_normalization_15[0][0]     
__________________________________________________________________________________________________
conv2d_11 (Conv2D)              (None, 28, 28, 32)   6176        activation_15[0][0]              
__________________________________________________________________________________________________
batch_normalization_16 (BatchNo (None, 28, 28, 32)   128         conv2d_11[0][0]                  
__________________________________________________________________________________________________
activation_16 (Activation)      (None, 28, 28, 32)   0           batch_normalization_16[0][0]     
__________________________________________________________________________________________________
add_2 (Add)                     (None, 28, 28, 32)   0           activation_16[0][0]              
                                                                 activation_13[0][0]              
__________________________________________________________________________________________________
conv2d_12 (Conv2D)              (None, 28, 28, 192)  6336        add_2[0][0]                      
__________________________________________________________________________________________________
batch_normalization_17 (BatchNo (None, 28, 28, 192)  768         conv2d_12[0][0]                  
__________________________________________________________________________________________________
activation_17 (Activation)      (None, 28, 28, 192)  0           batch_normalization_17[0][0]     
__________________________________________________________________________________________________
depthwise_conv2d_6 (DepthwiseCo (None, 28, 28, 192)  1920        activation_17[0][0]              
__________________________________________________________________________________________________
batch_normalization_18 (BatchNo (None, 28, 28, 192)  768         depthwise_conv2d_6[0][0]         
__________________________________________________________________________________________________
activation_18 (Activation)      (None, 28, 28, 192)  0           batch_normalization_18[0][0]     
__________________________________________________________________________________________________
conv2d_13 (Conv2D)              (None, 28, 28, 32)   6176        activation_18[0][0]              
__________________________________________________________________________________________________
batch_normalization_19 (BatchNo (None, 28, 28, 32)   128         conv2d_13[0][0]                  
__________________________________________________________________________________________________
activation_19 (Activation)      (None, 28, 28, 32)   0           batch_normalization_19[0][0]     
__________________________________________________________________________________________________
add_3 (Add)                     (None, 28, 28, 32)   0           activation_19[0][0]              
                                                                 add_2[0][0]                      
__________________________________________________________________________________________________
conv2d_14 (Conv2D)              (None, 28, 28, 192)  6336        add_3[0][0]                      
__________________________________________________________________________________________________
batch_normalization_20 (BatchNo (None, 28, 28, 192)  768         conv2d_14[0][0]                  
__________________________________________________________________________________________________
activation_20 (Activation)      (None, 28, 28, 192)  0           batch_normalization_20[0][0]     
__________________________________________________________________________________________________
depthwise_conv2d_7 (DepthwiseCo (None, 14, 14, 192)  1920        activation_20[0][0]              
__________________________________________________________________________________________________
batch_normalization_21 (BatchNo (None, 14, 14, 192)  768         depthwise_conv2d_7[0][0]         
__________________________________________________________________________________________________
activation_21 (Activation)      (None, 14, 14, 192)  0           batch_normalization_21[0][0]     
__________________________________________________________________________________________________
conv2d_15 (Conv2D)              (None, 14, 14, 64)   12352       activation_21[0][0]              
__________________________________________________________________________________________________
batch_normalization_22 (BatchNo (None, 14, 14, 64)   256         conv2d_15[0][0]                  
__________________________________________________________________________________________________
activation_22 (Activation)      (None, 14, 14, 64)   0           batch_normalization_22[0][0]     
__________________________________________________________________________________________________
conv2d_16 (Conv2D)              (None, 14, 14, 384)  24960       activation_22[0][0]              
__________________________________________________________________________________________________
batch_normalization_23 (BatchNo (None, 14, 14, 384)  1536        conv2d_16[0][0]                  
__________________________________________________________________________________________________
activation_23 (Activation)      (None, 14, 14, 384)  0           batch_normalization_23[0][0]     
__________________________________________________________________________________________________
depthwise_conv2d_8 (DepthwiseCo (None, 14, 14, 384)  3840        activation_23[0][0]              
__________________________________________________________________________________________________
batch_normalization_24 (BatchNo (None, 14, 14, 384)  1536        depthwise_conv2d_8[0][0]         
__________________________________________________________________________________________________
activation_24 (Activation)      (None, 14, 14, 384)  0           batch_normalization_24[0][0]     
__________________________________________________________________________________________________
conv2d_17 (Conv2D)              (None, 14, 14, 64)   24640       activation_24[0][0]              
__________________________________________________________________________________________________
batch_normalization_25 (BatchNo (None, 14, 14, 64)   256         conv2d_17[0][0]                  
__________________________________________________________________________________________________
activation_25 (Activation)      (None, 14, 14, 64)   0           batch_normalization_25[0][0]     
__________________________________________________________________________________________________
add_4 (Add)                     (None, 14, 14, 64)   0           activation_25[0][0]              
                                                                 activation_22[0][0]              
__________________________________________________________________________________________________
conv2d_18 (Conv2D)              (None, 14, 14, 384)  24960       add_4[0][0]                      
__________________________________________________________________________________________________
batch_normalization_26 (BatchNo (None, 14, 14, 384)  1536        conv2d_18[0][0]                  
__________________________________________________________________________________________________
activation_26 (Activation)      (None, 14, 14, 384)  0           batch_normalization_26[0][0]     
__________________________________________________________________________________________________
depthwise_conv2d_9 (DepthwiseCo (None, 14, 14, 384)  3840        activation_26[0][0]              
__________________________________________________________________________________________________
batch_normalization_27 (BatchNo (None, 14, 14, 384)  1536        depthwise_conv2d_9[0][0]         
__________________________________________________________________________________________________
activation_27 (Activation)      (None, 14, 14, 384)  0           batch_normalization_27[0][0]     
__________________________________________________________________________________________________
conv2d_19 (Conv2D)              (None, 14, 14, 64)   24640       activation_27[0][0]              
__________________________________________________________________________________________________
batch_normalization_28 (BatchNo (None, 14, 14, 64)   256         conv2d_19[0][0]                  
__________________________________________________________________________________________________
activation_28 (Activation)      (None, 14, 14, 64)   0           batch_normalization_28[0][0]     
__________________________________________________________________________________________________
add_5 (Add)                     (None, 14, 14, 64)   0           activation_28[0][0]              
                                                                 add_4[0][0]                      
__________________________________________________________________________________________________
conv2d_20 (Conv2D)              (None, 14, 14, 384)  24960       add_5[0][0]                      
__________________________________________________________________________________________________
batch_normalization_29 (BatchNo (None, 14, 14, 384)  1536        conv2d_20[0][0]                  
__________________________________________________________________________________________________
activation_29 (Activation)      (None, 14, 14, 384)  0           batch_normalization_29[0][0]     
__________________________________________________________________________________________________
depthwise_conv2d_10 (DepthwiseC (None, 14, 14, 384)  3840        activation_29[0][0]              
__________________________________________________________________________________________________
batch_normalization_30 (BatchNo (None, 14, 14, 384)  1536        depthwise_conv2d_10[0][0]        
__________________________________________________________________________________________________
activation_30 (Activation)      (None, 14, 14, 384)  0           batch_normalization_30[0][0]     
__________________________________________________________________________________________________
conv2d_21 (Conv2D)              (None, 14, 14, 64)   24640       activation_30[0][0]              
__________________________________________________________________________________________________
batch_normalization_31 (BatchNo (None, 14, 14, 64)   256         conv2d_21[0][0]                  
__________________________________________________________________________________________________
activation_31 (Activation)      (None, 14, 14, 64)   0           batch_normalization_31[0][0]     
__________________________________________________________________________________________________
add_6 (Add)                     (None, 14, 14, 64)   0           activation_31[0][0]              
                                                                 add_5[0][0]                      
__________________________________________________________________________________________________
conv2d_22 (Conv2D)              (None, 14, 14, 384)  24960       add_6[0][0]                      
__________________________________________________________________________________________________
batch_normalization_32 (BatchNo (None, 14, 14, 384)  1536        conv2d_22[0][0]                  
__________________________________________________________________________________________________
activation_32 (Activation)      (None, 14, 14, 384)  0           batch_normalization_32[0][0]     
__________________________________________________________________________________________________
depthwise_conv2d_11 (DepthwiseC (None, 14, 14, 384)  3840        activation_32[0][0]              
__________________________________________________________________________________________________
batch_normalization_33 (BatchNo (None, 14, 14, 384)  1536        depthwise_conv2d_11[0][0]        
__________________________________________________________________________________________________
activation_33 (Activation)      (None, 14, 14, 384)  0           batch_normalization_33[0][0]     
__________________________________________________________________________________________________
conv2d_23 (Conv2D)              (None, 14, 14, 96)   36960       activation_33[0][0]              
__________________________________________________________________________________________________
batch_normalization_34 (BatchNo (None, 14, 14, 96)   384         conv2d_23[0][0]                  
__________________________________________________________________________________________________
activation_34 (Activation)      (None, 14, 14, 96)   0           batch_normalization_34[0][0]     
__________________________________________________________________________________________________
conv2d_24 (Conv2D)              (None, 14, 14, 576)  55872       activation_34[0][0]              
__________________________________________________________________________________________________
batch_normalization_35 (BatchNo (None, 14, 14, 576)  2304        conv2d_24[0][0]                  
__________________________________________________________________________________________________
activation_35 (Activation)      (None, 14, 14, 576)  0           batch_normalization_35[0][0]     
__________________________________________________________________________________________________
depthwise_conv2d_12 (DepthwiseC (None, 14, 14, 576)  5760        activation_35[0][0]              
__________________________________________________________________________________________________
batch_normalization_36 (BatchNo (None, 14, 14, 576)  2304        depthwise_conv2d_12[0][0]        
__________________________________________________________________________________________________
activation_36 (Activation)      (None, 14, 14, 576)  0           batch_normalization_36[0][0]     
__________________________________________________________________________________________________
conv2d_25 (Conv2D)              (None, 14, 14, 96)   55392       activation_36[0][0]              
__________________________________________________________________________________________________
batch_normalization_37 (BatchNo (None, 14, 14, 96)   384         conv2d_25[0][0]                  
__________________________________________________________________________________________________
activation_37 (Activation)      (None, 14, 14, 96)   0           batch_normalization_37[0][0]     
__________________________________________________________________________________________________
add_7 (Add)                     (None, 14, 14, 96)   0           activation_37[0][0]              
                                                                 activation_34[0][0]              
__________________________________________________________________________________________________
conv2d_26 (Conv2D)              (None, 14, 14, 576)  55872       add_7[0][0]                      
__________________________________________________________________________________________________
batch_normalization_38 (BatchNo (None, 14, 14, 576)  2304        conv2d_26[0][0]                  
__________________________________________________________________________________________________
activation_38 (Activation)      (None, 14, 14, 576)  0           batch_normalization_38[0][0]     
__________________________________________________________________________________________________
depthwise_conv2d_13 (DepthwiseC (None, 14, 14, 576)  5760        activation_38[0][0]              
__________________________________________________________________________________________________
batch_normalization_39 (BatchNo (None, 14, 14, 576)  2304        depthwise_conv2d_13[0][0]        
__________________________________________________________________________________________________
activation_39 (Activation)      (None, 14, 14, 576)  0           batch_normalization_39[0][0]     
__________________________________________________________________________________________________
conv2d_27 (Conv2D)              (None, 14, 14, 96)   55392       activation_39[0][0]              
__________________________________________________________________________________________________
batch_normalization_40 (BatchNo (None, 14, 14, 96)   384         conv2d_27[0][0]                  
__________________________________________________________________________________________________
activation_40 (Activation)      (None, 14, 14, 96)   0           batch_normalization_40[0][0]     
__________________________________________________________________________________________________
add_8 (Add)                     (None, 14, 14, 96)   0           activation_40[0][0]              
                                                                 add_7[0][0]                      
__________________________________________________________________________________________________
conv2d_28 (Conv2D)              (None, 14, 14, 576)  55872       add_8[0][0]                      
__________________________________________________________________________________________________
batch_normalization_41 (BatchNo (None, 14, 14, 576)  2304        conv2d_28[0][0]                  
__________________________________________________________________________________________________
activation_41 (Activation)      (None, 14, 14, 576)  0           batch_normalization_41[0][0]     
__________________________________________________________________________________________________
depthwise_conv2d_14 (DepthwiseC (None, 7, 7, 576)    5760        activation_41[0][0]              
__________________________________________________________________________________________________
batch_normalization_42 (BatchNo (None, 7, 7, 576)    2304        depthwise_conv2d_14[0][0]        
__________________________________________________________________________________________________
activation_42 (Activation)      (None, 7, 7, 576)    0           batch_normalization_42[0][0]     
__________________________________________________________________________________________________
conv2d_29 (Conv2D)              (None, 7, 7, 160)    92320       activation_42[0][0]              
__________________________________________________________________________________________________
batch_normalization_43 (BatchNo (None, 7, 7, 160)    640         conv2d_29[0][0]                  
__________________________________________________________________________________________________
activation_43 (Activation)      (None, 7, 7, 160)    0           batch_normalization_43[0][0]     
__________________________________________________________________________________________________
conv2d_30 (Conv2D)              (None, 7, 7, 960)    154560      activation_43[0][0]              
__________________________________________________________________________________________________
batch_normalization_44 (BatchNo (None, 7, 7, 960)    3840        conv2d_30[0][0]                  
__________________________________________________________________________________________________
activation_44 (Activation)      (None, 7, 7, 960)    0           batch_normalization_44[0][0]     
__________________________________________________________________________________________________
depthwise_conv2d_15 (DepthwiseC (None, 7, 7, 960)    9600        activation_44[0][0]              
__________________________________________________________________________________________________
batch_normalization_45 (BatchNo (None, 7, 7, 960)    3840        depthwise_conv2d_15[0][0]        
__________________________________________________________________________________________________
activation_45 (Activation)      (None, 7, 7, 960)    0           batch_normalization_45[0][0]     
__________________________________________________________________________________________________
conv2d_31 (Conv2D)              (None, 7, 7, 160)    153760      activation_45[0][0]              
__________________________________________________________________________________________________
batch_normalization_46 (BatchNo (None, 7, 7, 160)    640         conv2d_31[0][0]                  
__________________________________________________________________________________________________
activation_46 (Activation)      (None, 7, 7, 160)    0           batch_normalization_46[0][0]     
__________________________________________________________________________________________________
add_9 (Add)                     (None, 7, 7, 160)    0           activation_46[0][0]              
                                                                 activation_43[0][0]              
__________________________________________________________________________________________________
conv2d_32 (Conv2D)              (None, 7, 7, 960)    154560      add_9[0][0]                      
__________________________________________________________________________________________________
batch_normalization_47 (BatchNo (None, 7, 7, 960)    3840        conv2d_32[0][0]                  
__________________________________________________________________________________________________
activation_47 (Activation)      (None, 7, 7, 960)    0           batch_normalization_47[0][0]     
__________________________________________________________________________________________________
depthwise_conv2d_16 (DepthwiseC (None, 7, 7, 960)    9600        activation_47[0][0]              
__________________________________________________________________________________________________
batch_normalization_48 (BatchNo (None, 7, 7, 960)    3840        depthwise_conv2d_16[0][0]        
__________________________________________________________________________________________________
activation_48 (Activation)      (None, 7, 7, 960)    0           batch_normalization_48[0][0]     
__________________________________________________________________________________________________
conv2d_33 (Conv2D)              (None, 7, 7, 160)    153760      activation_48[0][0]              
__________________________________________________________________________________________________
batch_normalization_49 (BatchNo (None, 7, 7, 160)    640         conv2d_33[0][0]                  
__________________________________________________________________________________________________
activation_49 (Activation)      (None, 7, 7, 160)    0           batch_normalization_49[0][0]     
__________________________________________________________________________________________________
add_10 (Add)                    (None, 7, 7, 160)    0           activation_49[0][0]              
                                                                 add_9[0][0]                      
__________________________________________________________________________________________________
conv2d_34 (Conv2D)              (None, 7, 7, 960)    154560      add_10[0][0]                     
__________________________________________________________________________________________________
batch_normalization_50 (BatchNo (None, 7, 7, 960)    3840        conv2d_34[0][0]                  
__________________________________________________________________________________________________
activation_50 (Activation)      (None, 7, 7, 960)    0           batch_normalization_50[0][0]     
__________________________________________________________________________________________________
depthwise_conv2d_17 (DepthwiseC (None, 7, 7, 960)    9600        activation_50[0][0]              
__________________________________________________________________________________________________
batch_normalization_51 (BatchNo (None, 7, 7, 960)    3840        depthwise_conv2d_17[0][0]        
__________________________________________________________________________________________________
activation_51 (Activation)      (None, 7, 7, 960)    0           batch_normalization_51[0][0]     
__________________________________________________________________________________________________
conv2d_35 (Conv2D)              (None, 7, 7, 320)    307520      activation_51[0][0]              
__________________________________________________________________________________________________
batch_normalization_52 (BatchNo (None, 7, 7, 320)    1280        conv2d_35[0][0]                  
__________________________________________________________________________________________________
activation_52 (Activation)      (None, 7, 7, 320)    0           batch_normalization_52[0][0]     
__________________________________________________________________________________________________
conv2d_36 (Conv2D)              (None, 7, 7, 1280)   410880      activation_52[0][0]              
__________________________________________________________________________________________________
batch_normalization_53 (BatchNo (None, 7, 7, 1280)   5120        conv2d_36[0][0]                  
__________________________________________________________________________________________________
activation_53 (Activation)      (None, 7, 7, 1280)   0           batch_normalization_53[0][0]     
__________________________________________________________________________________________________
global_average_pooling2d_1 (Glo (None, 1280)         0           activation_53[0][0]              
__________________________________________________________________________________________________
reshape_1 (Reshape)             (None, 1, 1, 1280)   0           global_average_pooling2d_1[0][0] 
__________________________________________________________________________________________________
Dropout (Dropout)               (None, 1, 1, 1280)   0           reshape_1[0][0]                  
__________________________________________________________________________________________________
conv2d_37 (Conv2D)              (None, 1, 1, 1000)   1281000     Dropout[0][0]                    
__________________________________________________________________________________________________
softmax (Activation)            (None, 1, 1, 1000)   0           conv2d_37[0][0]                  
__________________________________________________________________________________________________
reshape_2 (Reshape)             (None, 1000)         0           softmax[0][0]                    
==================================================================================================
Total params: 3,557,224
Trainable params: 3,523,048
Non-trainable params: 34,176
__________________________________________________________________________________________________

Conclusion

MobileNetV2 can train a model that is similar in accuracy to ordinary CNN, but lighter than ordinary CNN. Next, I will try to apply the mobilenet network to the detection of fire and smoke.

Reference

  1. MobileNet网络详解
  2. 徐梓涵,刘军,张苏沛,肖澳文,杜壮.一种基于MobileNet的火灾烟雾检测方法[J].武汉工程大学学报,2019,41(06):580-585.
    XU Zihan,LIUJun,ZHANG Supei,XIAO Aowen,DU Zhuang.Fire and Smoke Detection Method Based on MobileNet[J]. Journal of Wuhan Institute of Technology,2019,41(06):580-585.
grid_vision
关注
  • 0
    点赞
  • 踩
  • 0
    收藏
    觉得还不错? 一键收藏
  • 0
    评论
倒置残差与线性瓶颈层从原理到实现
weixin_42108090的博客
11-17 5613
对比 MobileNet V1 和 V2 的网络结构和计算量 MobileNet V1 网络结构和计算量 MobileNet V2 的网络结构和计算量 总结对比
MobileNetV2: Inverted Residuals and Linear Bottlenecks
12-03
Abstract In this paper we describe a new mobile architecture, MobileNetV2, that improves the state of the art perfor- mance of mobile models on multiple tasks and bench- marks as well as across a spectrum of different model sizes. We also describe efficient ways of applying these mobile models to object detection in a novel framework we call SSDLite. Additionally, we demonstrate how to build mobile semantic segmentation models through a reduced form of DeepLabv3 which we call Mobile DeepLabv3. is based on an inverted residual structure where the shortcut connections are between the thin bottle- neck layers. The intermediate expansion layer uses lightweight depthwise convolutions to filter features as a source of non-linearity. Additionally, we find that it is important to remove non-linearities in the narrow layers in order to maintain representational power. We demon- strate that this improves performance and provide an in- tuition that led to this design. Finally, our approach allows decoupling of the in- put/output domains from the expressiveness of the trans- formation, which provides a convenient framework for further analysis. We measure our performance on ImageNet [1] classification, COCO object detection [2], VOC image segmentation [3]. We evaluate the trade-offs between accuracy, and number of operations measured by multiply-adds (MAdd), as well as actual latency, and the number of parameters.
深度学习 —— 个人学习笔记14（ResNet、DenseNet）
Springer的博客
08-06 443
本文章为个人学习使用，版面观感若有不适请谅解，文中知识仅代表个人观点，若出现错误，欢迎各位批评指正。文中部分知识参考：B 站 —— 跟李沐学AI；
【深度学习】【语音】TTS，Matcha-TTS，测试效果，训练中文，chinese
q742971636的博客
08-08 185
即使使用较小的数据集（特别是如果这是一个录音棚录制的朗读语料库），也可以有效。然而，在这种情况下，微调通常会比从头开始训练效果更好，因此，你可以先在更大的数据集上进行训练，然后对你的特定数据集进行微调。一般来说，我更倾向于将所有音频文件重新采样到 22050 采样率，而不是更改音频参数，因为这样可以解决寻找不同声码器的问题。（同样，我没有在大多数其他数据集上进行训练的经验，但你可以在这里更改音素化语言。中添加的步骤为你的数据集生成均值和标准差（以便更好的标准化）。的默认值，并覆盖你特定数据集所需的部分。
深度学习与图像修复：ADetailer插件在Stable Diffusion中的应用
Why_does_it_work的博客
08-06 207
我们可以在这里进行下载，After Detailer插件的主界面设计简洁直观，为用户提供了便捷的操作体验。该插件具备强大的多通道处理能力，能够同时在一张照片中对脸部、手部等多个区域进行精细修复，极大地提升了图片处理的效率和灵活性。启用After Detailer，只需轻轻一勾，即可轻松开启您的图片修复之旅。在模型选择方面，After Detailer提供了丰富多样的模型选项，以满足不同场景下的修复需求。尤其是脸部修复功能，其效果堪称完美，赢得了用户的一致好评。
深度学习 - 数据存储形式对比(pkl/CSV/JSON等)
weixin_47552266的博客
08-09 733
在讨论深度学习中 .pkl 格式与其他常见文件格式（如 CSV、JSON、HDF5、Parquet 等）的区别时，重点可以放在它们如何处理数据的序列化、存储和读取。这些区别直接影响到文件格式的性能、存储效率和使用场景。
循环神经网络五-使用LSTM完成文本情感分类
m0_73426548的博客
08-09 277
2.构建dataset的过程，将数据转化为2分类，前面是十分类。pos类为1，neg为0，因为25000个要本做十分类数据量太小了。在前面使用word embedding去实现了文本情感分类。那么现在在这个模型中添加上LSTM层，为了达到更好的效果，做一下修改。3.在实例化LSTM的时候，使用dropout=0.5，在模型评估过程中，dropout自动会为0。二.完成训练和测试代码。
【扒代码】CCFF跨尺度特征融合
2301_77549977的博客
08-08 530
类提供了一种灵活的方式来构建和部署高效的卷积块，它通过融合卷积和归一化层来简化模型结构，同时保持性能。这种设计在实际应用中有助于提高模型的推理速度和减少内存占用。类实现了一个灵活的特征融合模块，可以用于深度学习模型中的各种任务，如特征提取、特征融合等。通过调整参数，可以适应不同的网络结构和应用需求。
7-1 深度学习硬件：CPU和GPU
Alkali!的博客
08-07 593
为cache，数据只有进入寄存器的时候，才能开始运算。（此图错误，一个矩阵应该是按。
【机器学习之深度学习】神经元、层的基本概念，神经网络以及神经元与线性回归和逻辑回归的相似之处
m0_49243785的博客
08-09 938
神经元是深度学习模型的基本单元，它模拟人脑中的神经元。在数学上，神经元接收输入，通过权重和偏置进行转换，并应用一个激活函数来产生输出层是由多个神经元组成的集合，它在神经网络中执行特定的变换神经元和层是深度学习模型的核心组成部分，它们共同工作，使得深度学习能够处理复杂的数据和任务已构建了一个非常简单的神经网络已经探索神经元与线性回归和逻辑回归的相似之处
Angiopep-2；脑靶向多肽；TFFYGGSRGKRNNFKTEEY；CAS:906480-05-5
Apeptide008的博客
08-09 1063
Angiopep-2是一种由19个氨基酸组成的新型多肽，能够通过血脑屏障上的低密度脂蛋白受体相关蛋白（LRP）受体介导内吞进入脑组织。
【深度学习】TTS，LibriTTS数据集
q742971636的博客
08-07 137
LibriTTS 是一个多说话人的英文语料库，包含大约 585 小时的朗读英语语音，采样率为 24kHz，由 Heiga Zen 在 Google Speech 和 Google Brain 团队成员的协助下准备。LibriTTS 语料库专为 TTS（文本到语音）研究设计。它源自 LibriSpeech 语料库的原始材料（LibriVox 的 mp3 音频文件和 Project Gutenberg 的文本文件）。这个数据集包括高质量的录音和对应的文本转录，可以帮助开发者构建和优化 TTS 模型。
基于深度学习的图像伪造检测
weixin_42605076的博客
08-09 490
基于深度学习的图像伪造检测主要利用深度学习技术来识别和检测伪造的图像内容，尤其是在生成对抗网络（GAN）等技术发展的背景下，伪造图像的逼真程度大大提升。图像伪造检测在信息安全、隐私保护、司法鉴定等领域具有重要意义。
多头注意力用单元矩阵实现以及原因
一休
08-09 1073
MHA(多头注意力）共享矩阵和独立矩阵实现，为什么采用共享矩阵，并用实例说明
深度学习中的梯度消失问题
最新发布
FeelTouch Labs
08-12 637
梯度消失问题是深度学习中的一个重要挑战，尤其在训练深层神经网络时尤为突出。当网络层数较多时，反向传播过程中，梯度可能会逐层减小，最终导致前面的层几乎无法更新，从而阻碍网络的学习。这种现象会导致模型性能下降，训练过程变得非常缓慢或甚至停滞。：例如，ReLU（修正线性单元）及其变体，能够有效缓解梯度消失问题。：选择合适的权重初始化方法，如 Xavier 初始化或 He 初始化，能够提高训练的稳定性。：在每一层引入批归一化，可以帮助保持激活值的分布稳定，从而加速训练，并减轻梯度消失的问题。
深度学习 —— 个人学习笔记16（目标检测和边界框、目标检测数据集）
Springer的博客
08-09 510
本文章为个人学习使用，版面观感若有不适请谅解，文中知识仅代表个人观点，若出现错误，欢迎各位批评指正。文中部分知识参考：B 站 —— 跟李沐学AI；
【深度学习】TTS，CosyVoice，推理部署的代码原理讲解分享
q742971636的博客
08-07 171
spk2info：存储和提供说话者嵌入，用于生成特定说话者的个性化语音。：从输入语音中提取特征并生成，作为语音处理任务的中间表示。它们在语音合成过程中相互补充，spk2info提供了说话者特征，而提供了语音信号的中间表示，从而共同实现高质量和个性化的语音生成。在语音处理过程中扮演着提取说话人特征的重要角色。通过生成说话人嵌入，可以在语音合成和识别任务中实现个性化和高质量的语音处理。
深度学习之NVIDIA驱动程序550的手动安装和配置（成功版）
weixin_53036073的博客
08-09 620
在训练数据集时遇到了驱动程序和cuda-11.8版本不匹配导致训练失败，检查驱动程序显示没有，重新下载驱动程序发现电脑上的驱动程序自动下载安装的话只能到384，所以这篇文章主要是教大家如何手动安装和配置NVIDIA-550版本。手动安装高版本的 NVIDIA 驱动程序（如 550版本）到电脑上通常可以正常使用，但需要确保安装过程正确且与现有系统和 CUDA 版本兼容。以下是官网查询到的版本兼容性列表：驱动程序本身并不决定计算能力，而是显卡的架构和型号决定的。
Bayes-Transformer-LSTM多变量时序预测 Matlab代码
small_ccb的博客
08-07 406
【独家原创】Bayes-Transformer-LSTM多变量时序预测Matlab代码 基于贝叶斯算法(Bayes)优化Transformer结合长短期记忆神经网络(LSTM)的数据多变量时序预测，Matlab代码，可直接运行，适合小白新手。采用 Transformer 编码器对光伏、负荷数据特征间的复杂关系以及时间序列中的长短期依赖关系进行挖掘，可以提高光伏功率、负荷预测的准确性。2️⃣、评价指标包括:R2、MAE、MSE、RPD、RMSE等，图很多，符合您的需要。数据格式为excel！
rbf q-leaning
11-30
RBF Q-learning是一种基于径向基函数的增强学习算法，它结合了Q-learning的基本原理和径向基函数的非线性映射特性。在传统的Q-learning算法中，学习过程是基于状态和动作的简单线性组合，这种方法在处理复杂的状态空间和动作空间时存在一定的局限性。而RBF Q-learning通过引入径向基函数，能够对状态空间和动作空间进行非线性映射，从而更好地适应复杂的环境。 具体来讲，RBF Q-learning的基本原理是利用径向基函数对状态空间进行映射，将复杂的状态表示为一组简单的基函数的线性组合。这样可以大大降低状态空间的维度，减少了学习参数的数量，降低了计算的复杂度。在每个基函数的基础上，RBF Q-learning利用Q-learning的奖励更新规则，不断优化动作价值函数，从而实现针对复杂状态空间的强化学习。 RBF Q-learning的应用领域非常广泛，特别适合处理具有高维状态空间和大规模动作空间的问题。例如，在机器人控制、自动驾驶、游戏策略等领域，RBF Q-learning都能够发挥出色的性能。同时，RBF Q-learning也为研究者提供了一个新的思路，可以结合深度学习的方法，进一步提高强化学习算法在复杂环境下的表现。 总之，RBF Q-learning是一种结合了Q-learning和径向基函数的增强学习算法，通过非线性映射和奖励更新规则的探索，能够更好地适应复杂的状态空间和动作空间，具有广泛的应用前景和研究价值。
评论
添加红包

请填写红包祝福语或标题

个

红包个数最小为10个

元

红包金额最低5元

当前余额3.43元 前往充值 >
需支付：10.00元
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付元
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明：

1.余额是钱包充值的虚拟货币，按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载，可以购买VIP、付费专栏及课程。

余额充值