-
Notifications
You must be signed in to change notification settings - Fork 0
/
my_main.py
157 lines (133 loc) · 5.07 KB
/
my_main.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
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
import sys
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, Dataset
import numpy as np
import pandas as pd
from PIL import Image
class MnistDataset(Dataset):
def __init__(self, file_path):
self.meta_data = pd.read_csv(file_path)
self.imgs = self.meta_data.iloc[:, 0]
self.labels = self.meta_data.iloc[:, 1]
def __getitem__(self, item):
img_path = self.imgs[item]
label = self.labels[item]
try:
img = np.array(Image.open(img_path))
except Exception as e:
img = None
print(e, img_path)
img = np.expand_dims(img, axis=0)
img = torch.from_numpy(img).float()
return img, label
def __len__(self):
return len(self.meta_data)
class lstm(nn.Module):
def __init__(self, in_dim=28, hidden_dim=64, num_class=10, num_layers=2):
super(lstm, self).__init__()
self.in_dim = in_dim
self.hidden_dim = hidden_dim
self.encoder = nn.LSTM(self.in_dim, self.hidden_dim, num_layers)
self.classifier = nn.Linear(self.hidden_dim, num_class)
def forward(self, x):
x = x.squeeze().permute(2,0,1)
features, _ = self.encoder(x)
out = self.classifier(features[-1, :, :])
return out
class ConvBlock(nn.Module):
def __init__(self, indim, outdim, padding = 1):
super(ConvBlock, self).__init__()
self.indim = indim
self.outdim = outdim
# self.C =
# self.BN =
# self.ReLu =
# self.P =
self.layers = [nn.Conv2d(indim, outdim, 3, padding=padding),
nn.BatchNorm2d(outdim),
nn.ReLU(inplace=True),
nn.MaxPool2d(2)]
self.block = nn.Sequential(*self.layers)
def forward(self, x):
out = self.block(x)
return out
class Conv4(nn.Module):
def __init__(self, num_fea=64, num_class=10, depth=4):
super(Conv4, self).__init__()
blocks = []
self.final_feat_dim = 64
self.num_fea = num_fea
self.num_class = num_class
for i in range(depth):
indim = 1 if i == 0 else num_fea
outdim = 64
b = ConvBlock(indim, outdim)
blocks.append(b)
self.embeding = nn.Sequential(*blocks)
self.classifier = nn.Linear(self.final_feat_dim, num_class)
def forward(self, x):
feature = self.embeding(x)
feature = feature.view(feature.shape[0], -1)
out = self.classifier(feature)
return out
def get_dataloader(file_csv, batch_size, shuffle, num_works):
print("读取数据:", file_csv)
data_set = MnistDataset(file_csv)
data_loader = DataLoader(data_set, batch_size=batch_size, shuffle=shuffle, num_workers=num_works)
return data_loader
def train_epoch(model, device, train_loader, optimizer, epoch):
avg_loss = 0
loss_func = nn.CrossEntropyLoss()
for batch_idx, (data, target) in enumerate(train_loader):
target = torch.squeeze(target)
data, target = data.to(device), target.to(device)
output = model(data)
loss = loss_func(output, target)
avg_loss += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
if (batch_idx % 5 == 1):
print("===> Epoch[{}]({}/{}): Loss: {:.4f}".format(epoch, batch_idx, len(train_loader), loss.item()))
print("===> Epoch[{}]: AvgLoss: {:.4f}".format(epoch, avg_loss/len(train_loader)))
def val_epoch(model, device, val_dataloader):
avg_loss, eval_acc = 0, 0
loss_func = nn.CrossEntropyLoss()
data_count = 0
for batch_idx, (data, target) in enumerate(val_dataloader):
target = torch.squeeze(target)
data, target = data.to(device), target.to(device)
output = model(data)
loss = loss_func(output, target)
avg_loss += loss.item()
_, pred = torch.max(output, 1)
num_correct = (pred == target).sum()
eval_acc += num_correct.data.item()
data_count += len(data)
print("===> Test: eval_acc: {:.4f}".format(eval_acc / data_count))
return eval_acc / data_count
def train(train_csv, val_csv):
train_dataloader = get_dataloader(train_csv, batch_size=256, shuffle=True, num_works=4)
val_dataloader = get_dataloader(val_csv, batch_size=256, shuffle=False, num_works=4)
# choose model CNN or LSTM
model = lstm()
# model = Conv4()
optimizer = optim.Adam(model.parameters(), lr=0.01)
device = torch.device("cuda:3")
# model = nn.DataParallel(model, device_ids=[1,2])
model.to(device)
print("开始训练")
max_acc = 0
for epoch in range(31):
model.train()
train_epoch(model, device, train_dataloader, optimizer, epoch)
if epoch % 5 ==0:
cur_acc = val_epoch(model, device, val_dataloader)
max_acc = max(max_acc, cur_acc)
print("===> Test: max_acc: {:.4f}".format(max_acc))
if __name__ == '__main__':
train_csv = "/data/gaoshan/MNIST/train.csv"
val_csv = "/data/gaoshan/MNIST/test.csv"
train(train_csv, val_csv)