1.数据集的路径,结构
dataset.py
目的:
输入:没有输入,路径是写死了的。
输出:返回的是一个对象,里面有self.data。self.data是一个列表,里面是(图片路径.jpg,标签)
-data[item]返回的是(img_tensor , one-hot编码)。one-hot编码是[0,1]或者[1,0]
import glob import os.path import cv2 import torch from torch.utils.data import Dataset from torchvision import transforms class DtataAndLabel(Dataset): def __init__(self,path='fruits',is_train=True): self.tran=transforms.Compose([ transforms.ToTensor(), transforms.Resize(size=(88,88)) ]) is_train='train' if True else 'test' self.data=[] path=os.path.join(path,is_train) print('path=',path) print(os.path.join(path, '*', '*')) img_paths=glob.glob(os.path.join(path,'*','*')) for img_path in img_paths: label=0 if img_path.split('\\')[-2]=='blueberry' else 1 self.data.append((img_path,label)) def __getitem__(self, idx): #每一张图片返回一个img_tensor,one_hot img_path,label =self.data[idx] img=cv2.imread(img_path) # img_gray=cv2.cvtColor(img,cv2.COLOR_BGR2GRAY) img_tensor=self.tran(img) img_tensor=img_tensor/255 img_tensor=torch.flatten(img_tensor) one_hot=torch.zeros(2) one_hot[label]=1 return img_tensor,one_hot def __len__(self): return len(self.data) if __name__ == '__main__': # 测试 data=DtataAndLabel() print(data[1][0].shape) print(data[1][1])
net.py
目的:将输入维度(k(k是加载进去的图片数),88,88,3)三通道的宽高是88,88,通过网络变化为(k,2)。
import torch.nn import torch.nn as nn class Net(nn.Module): def __init__(self): super().__init__() self.model = nn.Sequential( nn.Linear(88*88*3, 800), nn.ReLU(), nn.Linear(800, 500), nn.ReLU(), nn.Linear(500, 800), nn.ReLU(), nn.Linear(800, 200), nn.ReLU(), nn.Linear(200, 2), ) self.softmax=nn.Softmax(dim=1) def forward(self,x): x=self.model(x) x=self.softmax(x) return x if __name__ == '__main__': net=Net() #测试一下 x=torch.randn(1,100*100) out=net(x) print(out.shape)
test_train.py
目的:将图像丢进模型,然后训练出最优模型
步骤:
1.定义初始化
-定义拿到data对象
-定义加载器分批加载,这里可以变换维度
-定义初始化网络
-定义损失函数,这里采用了均方差函数
-定义优化器
2.实现训练
-将每一批数据丢给网络,此时维度发生了变化,产生了升维
-使用优化器
---自动梯度清0
---自动求导更新参数
-计算损失值和准确度
·~自己建一个文件夹
import torch.optim from torch import nn from torch.utils.data import DataLoader from torch.utils.tensorboard import SummaryWriter from tqdm import tqdm from net import Net from dataset import DtataAndLabel import torch.nn as nn class TrainAndTest(): def __init__(self): self.writer = SummaryWriter("logs") self.train_data=DtataAndLabel(is_train=True) self.test_data=DtataAndLabel(is_train=False) #使用加载器分批加载 self.train_loader=DataLoader(self.train_data,batch_size=10,shuffle=True) self.test_loader=DataLoader(self.test_data,batch_size=10,shuffle=True) #初始化网络 #损失函数 #优化器 net=Net() self.net=net self.loss=nn.MSELoss() self.opt=torch.optim.Adam(net.parameters(),lr=0.001) self.min_loss=100.0 self.weight_path='weight/best.pt' def train(self,epoch): sum_loss = 0 sum_acc = 0 for img_tensors, targets in tqdm(self.train_loader, desc="train...", total=len(self.train_loader)): out = self.net(img_tensors) loss = self.loss(out, targets) self.opt.zero_grad() loss.backward() self.opt.step() sum_loss += loss.item() pred_cls = torch.argmax(out, dim=1) target_cls = torch.argmax(targets, dim=1) accuracy = torch.mean(torch.eq(pred_cls, target_cls).to(torch.float32)) sum_acc += accuracy.item() avg_loss = sum_loss / len(self.train_loader) avg_acc = sum_acc / len(self.train_loader) print(f'train:loss{round(avg_loss, 3)} acc:{round(avg_acc, 3)}') self.writer.add_scalars("loss", {"train_avg_loss": avg_loss}, epoch) self.writer.add_scalars("acc", {"train_avg_acc": avg_acc}, epoch) def test(self,epoch): sum_loss = 0 sum_acc = 0 for img_tensors, targets in tqdm(self.test_loader, desc="test...", total=len(self.test_loader)): out = self.net(img_tensors) loss = self.loss(out, targets) sum_loss += loss.item() pred_cls = torch.argmax(out, dim=1) target_cls = torch.argmax(targets, dim=1) accuracy = torch.mean(torch.eq(pred_cls, target_cls).to(torch.float32)) sum_acc += accuracy.item() avg_loss = sum_loss / len(self.test_loader) avg_acc = sum_acc / len(self.test_loader) print(f'test:loss{round(avg_loss, 3)} acc:{round(avg_acc, 3)}') self.writer.add_scalars("loss", {"test_avg_loss": avg_loss}, epoch) self.writer.add_scalars("acc", {"test_avg_acc": avg_acc}, epoch) if avg_loss<self.min_loss: self.min_loss=min(self.min_loss,avg_loss) torch.save(self.net.state_dict(), self.weight_path) def run(self): for epo in range(100): self.train(epo) self.test(epo) if __name__ == '__main__': trainer=TrainAndTest() trainer.run()
精度的计算:
比如通过网络出现的维度是(1,2),其数值是[[0.9 , 0.1]](0.9与0.1表示预测的两个类别的概率)。我们通过maxarg取到其中最大的索引0,与之前真实的标签0或者1做比较。从而可以得出结果