NiN.py

#!/usr/bin/env python
# -*- coding: UTF-8 -*-
'''
@Project :Awesome-DL-Models 
@File    :NiN.py
@Author  :JackHCC
@Date    :2022/3/13 14:19 
@Desc    :

'''
import torch
import torch.nn as nn


def nin_block(in_channels, out_channels, kernel_size, strides, padding):
    return nn.Sequential(
        nn.Conv2d(in_channels, out_channels, kernel_size, strides, padding),
        nn.ReLU(),
        nn.Conv2d(out_channels, out_channels, kernel_size=1),
        nn.ReLU(),
        nn.Conv2d(out_channels, out_channels, kernel_size=1),
        nn.ReLU()
    )


NiN = nn.Sequential(
    nin_block(1, 96, kernel_size=11, strides=4, padding=0),
    nn.MaxPool2d(3, stride=2),
    nin_block(96, 256, kernel_size=5, strides=1, padding=2),
    nn.MaxPool2d(3, stride=2),
    nin_block(256, 384, kernel_size=3, strides=1, padding=1),
    nn.MaxPool2d(3, stride=2),
    nn.Dropout(0.5),
    # 标签类别数是10
    nin_block(384, 10, kernel_size=3, strides=1, padding=1),
    # 自适应平均池化，根据给定输出尺寸自己池化
    nn.AdaptiveAvgPool2d((1, 1)),
    # 将四维的输出转成二维的输出，其形状为(批量大小,10)
    nn.Flatten()
)


X = torch.rand(size=(1, 1, 224, 224))
for layer in NiN:
    X = layer(X)
    print(layer.__class__.__name__,'output shape:\t', X.shape)