-
Notifications
You must be signed in to change notification settings - Fork 0
/
FC_NN.py
121 lines (107 loc) · 4.77 KB
/
FC_NN.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
# MINST数据集是0-9的手写数字，所以有十个类，图片大小是28*28也就是784
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import DataLoader
from torchvision import datasets
from torchvision.transforms import transforms
from codecarbon import EmissionsTracker
class NN(nn.Module):
def __init__(self, input_size, class_NUM):
super(NN, self).__init__()
# self.fc1=nn.Linear(input_size,20)
# self.fc2=nn.Linear(20,class_NUM)
self.block = nn.Sequential(
nn.Linear(input_size, 20,bias=False),
nn.Linear(20, class_NUM,bias=False),
)
# 注意最后要把我们的输出reshape一下，可以自己打印出来shape看看
def forward(self, x):
x = self.block(x)
# print(x.size())
x = x.view(x.size(0), -1)
return x
class FC(nn.Module):
def __init__(self, input_size, class_NUM):
super(FC, self).__init__()
# self.fc1=nn.Linear(input_size,20)
# self.fc2=nn.Linear(20,class_NUM)
# 一上来就是一个和输入图像等大的卷积核，把图像卷成【N,1,1,20】的大小，然后在用一个1*1的卷积核替代全连接，到输出层，大小为【B，CLASS】
self.block = nn.Sequential(
nn.Conv2d(in_channels=1, out_channels=20, kernel_size=(input_size, input_size), bias=False),
# nn.MaxPool2d(kernel_size=2),
nn.ReLU(inplace=True),
nn.Conv2d(20, class_NUM, 1, bias=False)
)
# 注意最后要把我们的输出reshape一下，可以自己打印出来shape看看
def forward(self, x):
x = self.block(x)
# print(x.size())
x = x.view(x.size(0), -1)
# print(x.size())
return x
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
learning_rate = 1e-3
input_size = 784
class_NUM = 10
epoch = 2
batch_size = 64
train_set = datasets.MNIST(root='dataset/', train=True, transform=transforms.ToTensor(), download=True)
train_loader = DataLoader(train_set, batch_size, shuffle=True)
test_set = datasets.MNIST(root='dataset/', train=False, transform=transforms.ToTensor(), download=True)
test_loader = DataLoader(test_set, batch_size, shuffle=True)
model = NN(input_size, class_NUM).to(device)
model_FC = FC(28, class_NUM).to(device)
loss_function = nn.CrossEntropyLoss()
scaler = torch.cuda.amp.GradScaler()
used_model = model_FC
optimizer = optim.Adam(used_model.parameters(), lr=learning_rate)
# 代码注释部分并非错误而是比基础还基础的写法，我改用稍复杂、更实用的结构来替换
with EmissionsTracker() as tracker:
print("parameter:", sum(p.numel() for p in used_model.parameters() if p.requires_grad))
for singleepoch in range(epoch):
for batch_idx, (img, label) in enumerate(train_loader):
img = img.to(device)
label = label.to(device)
if used_model._get_name() == 'NN':
img = img.reshape(img.shape[0], -1)
with torch.cuda.amp.autocast():
predictions = used_model(img)
loss = loss_function(predictions, label)
# print("img",img.shape)
# print("label",label.shape)
# x=model(img)
# loss=loss_function(x,label)
optimizer.zero_grad()
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
# optimizer.zero_grad()
# loss.backward()
# optimizer.step()
# 记得加上eval（）和train（），因为网络在这两种模式下可以决定有一些层要不要使用，比如dropout等，正确开关可以提高正确率
def check_acc(loader, model):
num_correct = 0
num_sample = 0
model.eval()
with torch.no_grad():
for (img, label) in loader:
img = img.to(device)
label = label.to(device)
if used_model._get_name() == 'NN':
img = img.reshape(img.shape[0], -1)
# x就是数据过model的结果，根据我们的网络架构我们知道一共有十个class（列）【0...9】,数据（行）有多少个呢，根据loader决定
x = model(img)
# 很有启发的写法，直接一步得出每一行最大值所属坐标，preds将会是batchsize长的list，里面记录着每一行最大值坐标
_, preds = x.max(1)
# print("x:", x.shape)
# print("preds:", preds.shape)
# 分类对为1，然后累加
num_correct += (preds == label).sum()
num_sample += preds.size(0)
print(f"acc{num_correct / num_sample}")
model.train()
return num_correct / num_sample
check_acc(test_loader, used_model)
check_acc(train_loader, used_model)