mobilenetv2网络结构图
时间: 2023-07-02 08:19:51 浏览: 43
以下是MobileNetV2的网络结构图:
```
Input: 224x224x3
1. Conv2D: 3x3x3x32, stride=2, padding='same', activation=ReLU6
2. Inverted Residual Block:
a. Depthwise Conv2D: 3x3x32x1, stride=1, padding='same', activation=ReLU6
b. Pointwise Conv2D: 1x1x32x16, stride=1, padding='same', activation=linear
c. Linear Bottleneck:
i. Depthwise Conv2D: 3x3x16x1, stride=2, padding='same', activation=ReLU6
ii. Pointwise Conv2D: 1x1x16x64, stride=1, padding='same', activation=linear
iii. Pointwise Conv2D: 1x1x64x16, stride=1, padding='same', activation=linear
d. Add with input
3. Inverted Residual Block x 1
4. Inverted Residual Block:
a. Depthwise Conv2D: 3x3x64x1, stride=2, padding='same', activation=ReLU6
b. Pointwise Conv2D: 1x1x64x32, stride=1, padding='same', activation=linear
c. Linear Bottleneck:
i. Depthwise Conv2D: 3x3x32x1, stride=1, padding='same', activation=ReLU6
ii. Pointwise Conv2D: 1x1x32x128, stride=1, padding='same', activation=linear
iii. Pointwise Conv2D: 1x1x128x32, stride=1, padding='same', activation=linear
d. Add with input
5. Inverted Residual Block x 2
6. Inverted Residual Block:
a. Depthwise Conv2D: 3x3x96x1, stride=2, padding='same', activation=ReLU6
b. Pointwise Conv2D: 1x1x96x40, stride=1, padding='same', activation=linear
c. Linear Bottleneck:
i. Depthwise Conv2D: 3x3x40x1, stride=1, padding='same', activation=ReLU6
ii. Pointwise Conv2D: 1x1x40x192, stride=1, padding='same', activation=linear
iii. Pointwise Conv2D: 1x1x192x40, stride=1, padding='same', activation=linear
d. Add with input
7. Inverted Residual Block x 3
8. Inverted Residual Block:
a. Depthwise Conv2D: 3x3x192x1, stride=2, padding='same', activation=ReLU6
b. Pointwise Conv2D: 1x1x192x80, stride=1, padding='same', activation=linear
c. Linear Bottleneck:
i. Depthwise Conv2D: 3x3x80x1, stride=1, padding='same', activation=ReLU6
ii. Pointwise Conv2D: 1x1x80x320, stride=1, padding='same', activation=linear
iii. Pointwise Conv2D: 1x1x320x80, stride=1, padding='same', activation=linear
d. Add with input
9. Conv2D: 1x1x80x1280, stride=1, padding='same', activation=ReLU6
10. Global Average Pooling
11. Fully Connected: 1280x1000, activation=softmax
```
注:ReLU6是ReLU函数和ReLU函数的上限为6的组合函数。Depthwise Conv2D是指使用单个卷积核对每个输入通道进行卷积,Pointwise Conv2D是指使用1x1卷积核对每个通道进行卷积。Linear Bottleneck是指使用两个Pointwise Conv2D层和一个Depthwise Conv2D层的组合。Inverted Residual Block是指使用Linear Bottleneck和残差连接的组合。