最近在學習pyTorch, 在閱讀pytorch教程的時候,發現有一個簡單的卷積神經網絡,之前搞明白過這個過程,時間太久,都忘的差不多了, 正好寫個筆記記錄總結一下
代碼如下:
#! usr/bin/env python3
# -*- coding:utf-8 -*-
"""
@Author:MaCan
@Time:2019/10/29 19:59
@File:torch_net.py
@Mail:[email protected]
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import torch
import torch.nn as nn
import torch.nn.functional as F
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.conv1 = nn.Conv2d(in_channels=1, out_channels=6, kernel_size=5)
self.conv2 = nn.Conv2d(6, 16, 5)
self.fc1 = nn.Linear(16 * 5 * 5, 128)
self.fc2 = nn.Linear(128, 64)
self.fc3 = nn.Linear(64, 10)
def forward(self, x):
x = self.conv1(x)
print('x1 {}'.format(x.size()))
x = F.max_pool2d(F.relu(x), (2, 2))
print('x2 {}'.format(x.size()))
x = self.conv2(x)
print('x3 {}'.format(x.size()))
x = F.max_pool2d(F.relu(x), 2)
print('x4: {}'.format(x.size()))
x = x.view(-1, self.num_flat_features(x))
print('x5: {}'.format(x.size()))
x = F.relu(self.fc1(x))
print('x6: {}'.format(x.size()))
x = F.relu(self.fc2(x))
print('x7: {}'.format(x.size()))
x = self.fc3(x)
print('x8: {}'.format(x.size()))
return x
def num_flat_features(self, x):
size = x.size()[1:] # 除了batch 外的其他緯度值
print('size: {}'.format(size))
num_features = 1
for s in size:
num_features *= s
return num_features
if __name__ == '__main__':
net = Net()
print(net)
input = torch.randn(1, 1, 32, 32)
out = net(input)
print(out)
運行這個代碼,輸出如下:
Net(
(conv1): Conv2d(1, 6, kernel_size=(5, 5), stride=(1, 1))
(conv2): Conv2d(6, 16, kernel_size=(5, 5), stride=(1, 1))
(fc1): Linear(in_features=400, out_features=128, bias=True)
(fc2): Linear(in_features=128, out_features=64, bias=True)
(fc3): Linear(in_features=64, out_features=10, bias=True)
)
x1 torch.Size([1, 6, 28, 28])
x2 torch.Size([1, 6, 14, 14])
x3 torch.Size([1, 16, 10, 10])
x4: torch.Size([1, 16, 5, 5])
size: torch.Size([16, 5, 5])
x5: torch.Size([1, 400])
x6: torch.Size([1, 128])
x7: torch.Size([1, 64])
x8: torch.Size([1, 10])
tensor([[ 0.0684, 0.0224, -0.0527, -0.1091, 0.0603, -0.0389, -0.0848, -0.0689,
0.0107, -0.0398]], grad_fn=<AddmmBackward>)
爲了觀察每個過程的維度變化,我寫了一些print操作,其實維度變化已經很明顯了,下面來具體計算一下每個維度是怎麼計算的到的
- 第0層:網絡的輸入[1, 1, 32, 32],其中第一維是batch_size, 第二維是input_channel,後面兩維是數據的大小,
- 第1層:卷積層(convolution layer).第一層使用的卷積核大小爲[5, 5],卷積核(filter)的輸出深度爲6(output_channel),使用不填充,步長爲1.不填充的情況下,輸出的矩陣大小爲32 -5 + 1 = 28.因此第一層卷積輸出的feature_map的大小爲[28, 28, 6].
- 第2層:池化層(pooling layer). 池化層過濾器大小[2, 2],步長爲2.第二層的輸出維度(28-2)/2 + 1(括號裏面的“2”是kernel的對應維度, “/2”的2是步長)。因此第二層的的大小爲[14,14,6].
- 後面的維度以此類推了。