深度学习作业 - 食物分类(Food Classification)
题目
此次资料为网络上收集到的食物照片，共11类。
( 1 ) B r e a d , ( 2 ) D a i r y p r o d u c t , ( 3 ) D e s s e r t , ( 4 ) E g g , ( 5 ) F r i e d f o o d , (1)Bread,(2) Dairy product, (3)Dessert, (4)Egg, (5)Fried food, (1)Bread,(2)Dairyproduct,(3)Dessert,(4)Egg,(5)Friedfood,
( 6 ) M e a t , ( 7 ) N o o d l e s / P a s t a , ( 8 ) R i c e , ( 9 ) S e a f o o d , ( 10 ) S o u p , ( 11 ) V e g e t a b l e / F r u i t . (6)Meat, (7)Noodles/Pasta, (8)Rice, (9)Seafood, (10)Soup, (11)Vegetable/Fruit. (6)Meat,(7)Noodles/Pasta,(8)Rice,(9)Seafood,(10)Soup,(11)Vegetable/Fruit.
数据集：
- T r a i n i n g s e t Training set Trainingset: 9866 9866 9866张
- V a l i d a t i o n s e t Validation set Validationset: 3430 3430 3430张
- T e s t i n g s e t Testing set Testingset: 3347 3347 3347张
目标：
使用 C N N CNN CNN对图片进行分类
代码
（以下代码是参照官方代码手敲的，与官方代码只存在小部分不同）
一、导包
import os
import torch
import cv2
import numpy as np
import torchvision.transforms as transforms
from torch.utils.data import Dataset,DataLoader
import torch.nn as nn
import time
二、读取数据
读取数据的函数
def readfile(path,label):
'''
path:文件所在文件夹
labrl:bool类型，是否返回数据集标签
'''
#获取文件夹下的文件名
img_dir = sorted(os.listdir(path))
#装样本的x，len(img_dir)张128*128的彩色图片 (m,N_W,N_H,N_C)
x = np.zeros((len(img_dir),128,128,3),dtype=np.uint8)
y = np.zeros((len(img_dir)),dtype=np.uint8)
for i,file in enumerate(img_dir):
#把文件夹名和文件名拼在一起读取文件
img = cv2.imread(os.path.join(path,file))
x[i,:,:] = cv2.resize(img,(128,128))
if label:
y[i] = int(file.split("_")[0])#图片命名格式：分类_图片序号，如0_1.jpg 第0类的第1张图片
if label:
return x,y
else:
return x
读取数据
img_dir = 'Dataset/food-11'#文件夹位置
print('Reading data')
train_x,train_y = readfile(os.path.join(img_dir,"training"),True)
print("Size of training data={}".format(len(train_x)))
val_x,val_y = readfile(os.path.join(img_dir,"validation"),True)
print("Size of validation data={}".format(len(val_x)))
test_x = readfile(os.path.join(img_dir,"testing"),False)
print("Size of testing data={}".format(len(test_x)))
输出结果：
Reading data
Size of training data=9866
Size of validation data=3431
Size of testing data=3347
三、使用Dataset获取数据、DataLoader抽取数据
#定义transforms操作
#训练集数据的操作
train_trasform = transforms.Compose([
transforms.ToPILImage(),
#数据增强
transforms.RandomHorizontalFlip(),#水平翻转
transforms.RandomRotation(15),#随机旋转15°
transforms.ToTensor(),
])
#测试集数据不需要进行数据增强
test_transform = transforms.Compose([
transforms.ToPILImage(),
transforms.ToTensor(),
])
#继承Dataset来获取数据
class ImgDataset(Dataset):
def __init__(self,x,y=None,transform=None):
self.x = x
self.y = y
if y is not None:
self.y = torch.LongTensor(y)#转成tensor类型
self.transform = transform
def __len__(self):
return len(self.x)
#按下标获取数据
def __getitem__(self, item):
X = self.x[item]
if self.transform is not None:
X = self.transform(X)
if self.y is not None:
Y = self.y[item]
return X,Y
else:
return X
使用Dataloder进行抽样
batch_size = 128
train_set = ImgDataset(train_x,train_y,train_trasform)
val_set = ImgDataset(val_x,val_y,test_transform)
train_loder = DataLoader(train_set,batch_size=batch_size,shuffle=True)
val_loder = DataLoader(val_set,batch_size=batch_size,shuffle=False)
四、定义模型
通道数 N C N_C NC不断变多， N h 、 N w N_h、N_w Nh、Nw不断减少。
class Classifier(nn.Module):
def __init__(self) -> None:
super().__init__()
#input维度：[3,128,128] tensor类型是通道数在前的
self.cnn = nn.Sequential(
nn.Conv2d(in_channels=3,out_channels=64,kernel_size=3,stride=1,padding=1), #输出维度:[64,128,128]
nn.BatchNorm2d(64),
nn.ReLU(),
nn.MaxPool2d(2,2,0),#输出维度：[64，64，64]
nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, stride=1, padding=1), # 输出维度:[128,64,64]
nn.BatchNorm2d(128),
nn.ReLU(),
nn.MaxPool2d(2, 2, 0), # 输出维度：[128，32，32]
nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, stride=1, padding=1), # 输出维度:[256,32,32]
nn.BatchNorm2d(256),
nn.ReLU(),
nn.MaxPool2d(2, 2, 0), # 输出维度：[256，16，16]
nn.Conv2d(in_channels=256, out_channels=512, kernel_size=3, stride=1, padding=1), # 输出维度:[256,16,16]
nn.BatchNorm2d(512),
nn.ReLU(),
nn.MaxPool2d(2, 2, 0), # 输出维度：[512，8，8]
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1), # 输出维度:[512,8,8]
nn.BatchNorm2d(512),
nn.ReLU(),
nn.MaxPool2d(2, 2, 0), # 输出维度：[512，4，4]
)
self.fc = nn.Sequential(
nn.Linear(512*4*4,1024),
nn.ReLU(),
nn.Linear(1024,512),
nn.ReLU(),
nn.Linear(512,11)
)
def forward(self, x):
out = self.cnn(x)
out = out.view(out.size()[0],-1)#按照第一个维度平铺，即flatten()
return self.fc(out)
五、训练模型
model = Classifier()
#判断是否有cuda
if torch.cuda.is_available():
model = model.cuda()
loss = nn.CrossEntropyLoss()
if torch.cuda.is_available():
loss = loss.cuda()
#定义优化器
optimizer = torch.optim.Adam(model.parameters(),lr=0.001)
num_epoch = 30
for epoch in range(num_epoch):
epoch_start_time = time.time()
train_acc = 0.0
train_loss = 0.0
val_acc = 0.0
val_loss = 0.0
model.train()#在使用了Dropout、BatchNormal时不用省略，作用是确保是在训练模型
for data in train_loder:
img,targets = data
if torch.cuda.is_available():
img = img.cuda()
targets = targets.cuda()
optimizer.zero_grad()#梯度清零
train_pred = model(img)#获得输出
batch_loss = loss(train_pred,targets)#计算损失函数
batch_loss.backward()#反向传播
optimizer.step()#执行一次梯度下降
train_acc += (train_pred.argmax(1)==targets).sum()#计算准确率
train_loss += batch_loss.item()
model.eval()#表示开启测试，后续代码是进行验证集训练的(使用验证集选择参数)
with torch.no_grad():#确保梯度不会改变
for data in val_loder:
img, targets = data
if torch.cuda.is_available():
img = img.cuda()
targets = targets.cuda()
val_pred = model(img)
batch_loss = loss(val_pred,targets)
val_acc += (val_pred.argmax(1) == targets).sum()
val_loss += batch_loss.item()
print('[%03d/%03d] %2.2f sec(s) Train Acc: %3.6f Loss: %3.6f | Val Acc: %3.6f loss: %3.6f' % (epoch + 1, num_epoch, time.time() - epoch_start_time, train_acc / train_set.__len__(), train_loss / train_set.__len__(), val_acc / val_set.__len__(),val_loss / val_set.__len__()))
运行结果：
[001/030] 23.66 sec(s) Train Acc: 0.236063 Loss: 0.017809 | Val Acc: 0.213932 loss: 0.017797
[002/030] 23.69 sec(s) Train Acc: 0.347963 Loss: 0.014709 | Val Acc: 0.323521 loss: 0.015096
[003/030] 23.48 sec(s) Train Acc: 0.397730 Loss: 0.013572 | Val Acc: 0.293209 loss: 0.019180
[004/030] 23.88 sec(s) Train Acc: 0.445064 Loss: 0.012603 | Val Acc: 0.445351 loss: 0.012793
[005/030] 23.58 sec(s) Train Acc: 0.488547 Loss: 0.011668 | Val Acc: 0.476246 loss: 0.011889
[006/030] 23.53 sec(s) Train Acc: 0.533448 Loss: 0.010752 | Val Acc: 0.495482 loss: 0.011747
[007/030] 23.54 sec(s) Train Acc: 0.557369 Loss: 0.010105 | Val Acc: 0.476246 loss: 0.012213
[008/030] 23.52 sec(s) Train Acc: 0.572066 Loss: 0.009716 | Val Acc: 0.519382 loss: 0.011405
[009/030] 23.55 sec(s) Train Acc: 0.595682 Loss: 0.009187 | Val Acc: 0.556398 loss: 0.010150
[010/030] 23.83 sec(s) Train Acc: 0.634502 Loss: 0.008349 | Val Acc: 0.521714 loss: 0.011477
[011/030] 23.56 sec(s) Train Acc: 0.641496 Loss: 0.008131 | Val Acc: 0.587292 loss: 0.010066
[012/030] 23.58 sec(s) Train Acc: 0.675451 Loss: 0.007482 | Val Acc: 0.489070 loss: 0.012889
[013/030] 23.54 sec(s) Train Acc: 0.681431 Loss: 0.007245 | Val Acc: 0.552026 loss: 0.011332
[014/030] 23.50 sec(s) Train Acc: 0.693493 Loss: 0.007033 | Val Acc: 0.575634 loss: 0.010282
[015/030] 23.59 sec(s) Train Acc: 0.701906 Loss: 0.006724 | Val Acc: 0.610318 loss: 0.009404
[016/030] 23.86 sec(s) Train Acc: 0.720454 Loss: 0.006281 | Val Acc: 0.572428 loss: 0.011218
[017/030] 23.58 sec(s) Train Acc: 0.744172 Loss: 0.005858 | Val Acc: 0.606529 loss: 0.010286
[018/030] 23.60 sec(s) Train Acc: 0.749443 Loss: 0.005562 | Val Acc: 0.602157 loss: 0.010116
[019/030] 23.53 sec(s) Train Acc: 0.771032 Loss: 0.005292 | Val Acc: 0.574177 loss: 0.011388
[020/030] 23.52 sec(s) Train Acc: 0.751976 Loss: 0.005733 | Val Acc: 0.615855 loss: 0.010243
[021/030] 23.87 sec(s) Train Acc: 0.751368 Loss: 0.005585 | Val Acc: 0.638298 loss: 0.009475
[022/030] 23.67 sec(s) Train Acc: 0.788364 Loss: 0.004805 | Val Acc: 0.637424 loss: 0.009800
[023/030] 23.61 sec(s) Train Acc: 0.802858 Loss: 0.004554 | Val Acc: 0.605946 loss: 0.011229
[024/030] 23.54 sec(s) Train Acc: 0.803061 Loss: 0.004387 | Val Acc: 0.663072 loss: 0.009186
[025/030] 23.52 sec(s) Train Acc: 0.828299 Loss: 0.003794 | Val Acc: 0.618770 loss: 0.011051
[026/030] 23.61 sec(s) Train Acc: 0.840969 Loss: 0.003588 | Val Acc: 0.657826 loss: 0.010324
[027/030] 23.83 sec(s) Train Acc: 0.843098 Loss: 0.003669 | Val Acc: 0.665112 loss: 0.009451
[028/030] 23.64 sec(s) Train Acc: 0.838334 Loss: 0.003664 | Val Acc: 0.659866 loss: 0.010017
[029/030] 23.57 sec(s) Train Acc: 0.867119 Loss: 0.003018 | Val Acc: 0.642087 loss: 0.010967
[030/030] 23.55 sec(s) Train Acc: 0.869653 Loss: 0.002892 | Val Acc: 0.659575 loss: 0.010865
六、验证集和训练集合并训练模型
合并数据集
train_val_x = np.concatenate((train_x,val_x),axis=0)
train_val_y = np.concatenate((train_y,val_y),axis=0)
train_val_set = ImgDataset(train_val_x,train_val_y,train_trasform)
train_val_lodaer = DataLoader(train_val_set,batch_size=batch_size,shuffle=True)
训练模型
model_best = Classifier()
if torch.cuda.is_available():
model_best = model_best.cuda()
loss = nn.CrossEntropyLoss()
if torch.cuda.is_available():
loss = loss.cuda()
optimizer = torch.optim.Adam(model_best.parameters(),lr=0.001)
num_epoch = 30
for epoch in range(num_epoch):
epoch_start_time = time.time()
train_acc = 0.0
train_loss = 0.0
for data in train_val_lodaer:
img,targets = data
if torch.cuda.is_available():
img = img.cuda()
targets = targets.cuda()
optimizer.zero_grad()
train_pred = model_best(img)
batch_loss = loss(train_pred,targets)
batch_loss.backward()
optimizer.step()
train_acc += (train_pred.argmax(1) == targets).sum()
train_loss += batch_loss.item()
print('[%03d/%03d] %2.2f sec(s) Train Acc: %3.6f Loss: %3.6f' % (epoch + 1, num_epoch, time.time()-epoch_start_time, train_acc/train_val_set.__len__(), train_loss/train_val_set.__len__()))
输出结果：
[001/030] 28.60 sec(s) Train Acc: 0.264345 Loss: 0.016747
[002/030] 28.16 sec(s) Train Acc: 0.385952 Loss: 0.013655
[003/030] 28.05 sec(s) Train Acc: 0.442055 Loss: 0.012396
[004/030] 28.14 sec(s) Train Acc: 0.500338 Loss: 0.011158
[005/030] 28.14 sec(s) Train Acc: 0.542228 Loss: 0.010296
[006/030] 28.39 sec(s) Train Acc: 0.587877 Loss: 0.009280
[007/030] 28.09 sec(s) Train Acc: 0.610062 Loss: 0.008759
[008/030] 28.00 sec(s) Train Acc: 0.633602 Loss: 0.008285
[009/030] 28.09 sec(s) Train Acc: 0.666316 Loss: 0.007510
[010/030] 28.10 sec(s) Train Acc: 0.689554 Loss: 0.006984
[011/030] 28.34 sec(s) Train Acc: 0.708957 Loss: 0.006621
[012/030] 28.10 sec(s) Train Acc: 0.717906 Loss: 0.006296
[013/030] 28.11 sec(s) Train Acc: 0.743927 Loss: 0.005802
[014/030] 28.04 sec(s) Train Acc: 0.747913 Loss: 0.005656
[015/030] 28.07 sec(s) Train Acc: 0.770775 Loss: 0.005209
[016/030] 28.39 sec(s) Train Acc: 0.784538 Loss: 0.004807
[017/030] 28.03 sec(s) Train Acc: 0.790629 Loss: 0.004639
[018/030] 28.09 sec(s) Train Acc: 0.815673 Loss: 0.004161
[019/030] 28.04 sec(s) Train Acc: 0.821915 Loss: 0.004001
[020/030] 28.05 sec(s) Train Acc: 0.833496 Loss: 0.003724
[021/030] 28.40 sec(s) Train Acc: 0.838535 Loss: 0.003553
[022/030] 28.06 sec(s) Train Acc: 0.850718 Loss: 0.003346
[023/030] 28.03 sec(s) Train Acc: 0.867639 Loss: 0.002950
[024/030] 28.06 sec(s) Train Acc: 0.875611 Loss: 0.002728
[025/030] 28.02 sec(s) Train Acc: 0.884260 Loss: 0.002583
[026/030] 28.34 sec(s) Train Acc: 0.892758 Loss: 0.002402
[027/030] 28.06 sec(s) Train Acc: 0.897646 Loss: 0.002194
[028/030] 28.01 sec(s) Train Acc: 0.912762 Loss: 0.001959
[029/030] 27.98 sec(s) Train Acc: 0.921486 Loss: 0.001751
[030/030] 28.33 sec(s) Train Acc: 0.917951 Loss: 0.001820
七、模型参数保存
torch.save(model_best.state_dict(),'model_best.pth')
八、使用测试集图片预测
pred_dict = {
0:'Bread',1:'Dairy product', 2:'Dessert',3:'Egg',4:'Fried food',5:'Meat',6:'Noodles/Pasta',7:'Rice',8:'Seafood',9:'Soup',10:'Vegetable/Fruit'
}
img = test_x[10]
cv2_imshow(img)
img = test_transform(img).unsqueeze(0)#进行transform操作并添加一个维度
img = img.cuda()
test_pred = model_best(img)
print(pred_dict[test_pred.argmax(axis=1).item()])
实验结果：
完整代码
import os
import torch
import cv2
import numpy as np
import torchvision.transforms as transforms
from torch.utils.data import Dataset,DataLoader
import torch.nn as nn
import time
def readfile(path,label):
'''
path:文件所在文件夹
labrl:bool类型，是否返回数据集标签
'''
#获取文件夹下的文件名
img_dir = sorted(os.listdir(path))
#装样本的x，len(img_dir)张128*128的彩色图片 (m,N_W,N_H,N_C)
x = np.zeros((len(img_dir),128,128,3),dtype=np.uint8)
y = np.zeros((len(img_dir)),dtype=np.uint8)
for i,file in enumerate(img_dir):
#把文件夹名和文件名拼在一起读取文件
img = cv2.imread(os.path.join(path,file))
x[i,:,:] = cv2.resize(img,(128,128))
if label:
y[i] = int(file.split("_")[0])
if label:
return x,y
else:
return x
img_dir = '/content/gdrive/MyDrive/ColabNotebooks/Dataset/food-11'#文件夹位置
print('Reading data')
train_x,train_y = readfile(os.path.join(img_dir,"training"),True)
print("Size of training data={}".format(len(train_x)))
val_x,val_y = readfile(os.path.join(img_dir,"validation"),True)
print("Size of validation data={}".format(len(val_x)))
test_x = readfile(os.path.join(img_dir,"testing"),False)
print("Size of testing data={}".format(len(test_x)))
train_trasform = transforms.Compose([
transforms.ToPILImage(),
#数据增强
transforms.RandomHorizontalFlip(),#水平翻转
transforms.RandomRotation(15),#随机旋转15°
transforms.ToTensor(),
])
#测试集数据不需要进行数据增强
test_transform = transforms.Compose([
transforms.ToPILImage(),
transforms.ToTensor(),
])
#继承Dataset来获取数据
class ImgDataset(Dataset):
def __init__(self,x,y=None,transform=None):
self.x = x
self.y = y
if y is not None:
self.y = torch.LongTensor(y)#转成tensor类型
self.transform = transform
def __len__(self):
return len(self.x)
def __getitem__(self, item):
X = self.x[item]
if self.transform is not None:
X = self.transform(X)
if self.y is not None:
Y = self.y[item]
return X,Y
else:
return X
batch_size = 128
train_set = ImgDataset(train_x,train_y,train_trasform)
val_set = ImgDataset(val_x,val_y,test_transform)
train_loder = DataLoader(train_set,batch_size=batch_size,shuffle=True)
val_loder = DataLoader(val_set,batch_size=batch_size,shuffle=False)
class Classifier(nn.Module):
def __init__(self) -> None:
super().__init__()
# input维度：[3,128,128] tensor类型是通道数在前的
self.cnn = nn.Sequential(
nn.Conv2d(in_channels=3, out_channels=64, kernel_size=3, stride=1, padding=1), # 输出维度:[64,128,128]
nn.BatchNorm2d(64),
nn.ReLU(),
nn.MaxPool2d(2, 2, 0), # 输出维度：[64，64，64]
nn.Conv2d(in_channels=64, out_channels=128, kernel_size=3, stride=1, padding=1), # 输出维度:[128,64,64]
nn.BatchNorm2d(128),
nn.ReLU(),
nn.MaxPool2d(2, 2, 0), # 输出维度：[128，32，32]
nn.Conv2d(in_channels=128, out_channels=256, kernel_size=3, stride=1, padding=1), # 输出维度:[256,32,32]
nn.BatchNorm2d(256),
nn.ReLU(),
nn.MaxPool2d(2, 2, 0), # 输出维度：[256，16，16]
nn.Conv2d(in_channels=256, out_channels=512, kernel_size=3, stride=1, padding=1), # 输出维度:[256,16,16]
nn.BatchNorm2d(512),
nn.ReLU(),
nn.MaxPool2d(2, 2, 0), # 输出维度：[512，8，8]
nn.Conv2d(in_channels=512, out_channels=512, kernel_size=3, stride=1, padding=1), # 输出维度:[512,8,8]
nn.BatchNorm2d(512),
nn.ReLU(),
nn.MaxPool2d(2, 2, 0), # 输出维度：[512，4，4]
)
self.fc = nn.Sequential(
nn.Linear(512 * 4 * 4, 1024),
nn.ReLU(),
nn.Linear(1024, 512),
nn.ReLU(),
nn.Linear(512, 11)
)
def forward(self, x):
out = self.cnn(x)
out = out.view(out.size()[0], -1)
return self.fc(out)
model = Classifier()
if torch.cuda.is_available():
model = model.cuda()
loss = nn.CrossEntropyLoss()
if torch.cuda.is_available():
loss = loss.cuda()
optimizer = torch.optim.Adam(model.parameters(),lr=0.001)
num_epoch = 30
for epoch in range(num_epoch):
epoch_start_time = time.time()
train_acc = 0.0
train_loss = 0.0
val_acc = 0.0
val_loss = 0.0
model.train()#在使用了Dropout、BatchNormal时不用省略，作用是确保是在训练模型
for data in train_loder:
img,targets = data
if torch.cuda.is_available():
img = img.cuda()
targets = targets.cuda()
optimizer.zero_grad()#梯度清零
train_pred = model(img)#获得输出
batch_loss = loss(train_pred,targets)
batch_loss.backward()
optimizer.step()#执行梯度下降
train_acc += (train_pred.argmax(1)==targets).sum()
train_loss += batch_loss.item()
model.eval()#表示开启测试，后续代码是进行验证集训练的
with torch.no_grad():
for data in val_loder:
img, targets = data
if torch.cuda.is_available():
img = img.cuda()
targets = targets.cuda()
val_pred = model(img)
batch_loss = loss(val_pred,targets)
val_acc += (val_pred.argmax(1) == targets).sum()
val_loss += batch_loss.item()
print('[%03d/%03d] %2.2f sec(s) Train Acc: %3.6f Loss: %3.6f | Val Acc: %3.6f loss: %3.6f' % (epoch + 1, num_epoch, time.time() - epoch_start_time, train_acc / train_set.__len__(), train_loss / train_set.__len__(), val_acc / val_set.__len__(),val_loss / val_set.__len__()))
train_val_x = np.concatenate((train_x,val_x),axis=0)
train_val_y = np.concatenate((train_y,val_y),axis=0)
train_val_set = ImgDataset(train_val_x,train_val_y,train_trasform)
train_val_lodaer = DataLoader(train_val_set,batch_size=batch_size,shuffle=True)
model_best = Classifier()
if torch.cuda.is_available():
model_best = model_best.cuda()
loss = nn.CrossEntropyLoss()
if torch.cuda.is_available():
loss = loss.cuda()
optimizer = torch.optim.Adam(model_best.parameters(),lr=0.001)
num_epoch = 30
for epoch in range(num_epoch):
epoch_start_time = time.time()
train_acc = 0.0
train_loss = 0.0
for data in train_val_lodaer:
img,targets = data
if torch.cuda.is_available():
img = img.cuda()
targets = targets.cuda()
optimizer.zero_grad()
train_pred = model_best(img)
batch_loss = loss(train_pred,targets)
batch_loss.backward()
optimizer.step()
train_acc += (train_pred.argmax(1) == targets).sum()
train_loss += batch_loss.item()
print('[%03d/%03d] %2.2f sec(s) Train Acc: %3.6f Loss: %3.6f' % (epoch + 1, num_epoch, time.time()-epoch_start_time, train_acc/train_val_set.__len__(), train_loss/train_val_set.__len__()))
torch.save(model_best.state_dict(),'model_best.pth')
总结
- 数据增强：
对图片数据进行旋转、翻转等操作，通过增加训练集的方式进行数据增强；
本次实验的旋转和翻转都是随机进行的，所有每次训练的训练集数量上一样，但是图片内容都是不同的；所以迭代次数越多数据增强效果越明显。