【深度学习】快速制作图像标签数据集以及训练

快速制作图像标签数据集以及训练

制作DataSet

• 先从网络收集十张图片 每种十张
  

• 定义dataSet和dataloader

import glob
import torch
from torch.utils import data
from PIL import Image
import numpy as np
from torchvision import transforms
import matplotlib.pyplot as plt# 通过创建data.Dataset子类Mydataset来创建输入
class Mydataset(data.Dataset):# init() 初始化方法，传入数据文件夹路径def __init__(self, root):self.imgs_path = root# getitem() 切片方法，根据索引下标，获得相应的图片def __getitem__(self, index):img_path = self.imgs_path[index]# len() 计算长度方法，返回整个数据文件夹下所有文件的个数def __len__(self):return len(self.imgs_path)# 使用glob方法来获取数据图片的所有路径
all_imgs_path = glob.glob(r"./Data/*/*.jpg")  # 数据文件夹路径# 利用自定义类Mydataset创建对象brake_dataset
# 将所有的路径塞进dataset  使用每张图片的路径进行索引图片
brake_dataset = Mydataset(all_imgs_path)
# print("图片总数:{}".format(len(brake_dataset)))  # 返回文件夹中图片总个数# 制作dataloader
brake_dataloader = torch.utils.data.DataLoader(brake_dataset, batch_size=2)  # 每次迭代时返回4个数据
# print(next(iter(break_dataloader)))

制作标签

# 为每张图片制作对应标签
species = ['sun', 'rain', 'cloud']
species_to_id = dict((c, i) for i, c in enumerate(species))
# print(species_to_id)id_to_species = dict((v, k) for k, v in species_to_id.items())
# print(id_to_species)# 对所有图片路径进行迭代
all_labels = []
for img in all_imgs_path:# 区分出每个img，应该属于什么类别for i, c in enumerate(species):if c in img:all_labels.append(i)
# print(all_labels)

制作数据和标签一起的dataset和dataloader

• 上面的dataset不够完善

# 将数据转换为张量数据
# 对数据进行转换处理
transform = transforms.Compose([transforms.Resize((256, 256)),  # 做的第一步转换transforms.ToTensor()  # 第二步转换，作用：第一转换成Tensor，第二将图片取值范围转换成0-1之间，第三会将channel置前
])class Mydatasetpro(data.Dataset):def __init__(self, img_paths, labels, transform):self.imgs = img_pathsself.labels = labelsself.transforms = transform# 进行切片def __getitem__(self, index):img = self.imgs[index]label = self.labels[index]pil_img = Image.open(img)  # pip install pillowpil_img = pil_img.convert('RGB')data = self.transforms(pil_img)return data, label# 返回长度def __len__(self):return len(self.imgs)BATCH_SIZE = 4
brake_dataset = Mydatasetpro(all_imgs_path, all_labels, transform)
brake_dataloader = data.DataLoader(brake_dataset,batch_size=BATCH_SIZE,shuffle=True
)

imgs_batch, labels_batch = next(iter(brake_dataloader))# 4 X 3 X 256 X 256
print(imgs_batch.shape)plt.figure(figsize=(12, 8))
for i, (img, label) in enumerate(zip(imgs_batch[:10], labels_batch[:10])):img = img.permute(1, 2, 0).numpy()plt.subplot(2, 3, i + 1)plt.title(id_to_species.get(label.item()))plt.imshow(img)
plt.show()  # 展示图片

制作训练集和测试集

# 划分数据集和测试集
index = np.random.permutation(len(all_imgs_path))#  打乱所有图片的索引
print(index)# 根据索引获取所有图片的路径
all_imgs_path = np.array(all_imgs_path)[index]
all_labels = np.array(all_labels)[index]print("打乱顺序之后的所有图片路径{}".format(all_imgs_path))
print("打乱顺序之后的所有图片索引{}".format(all_labels))# 80%做训练集
s = int(len(all_imgs_path) * 0.8)
# print(s)train_imgs = all_imgs_path[:s]
# print(train_imgs)
train_labels = all_labels[:s]
test_imgs = all_imgs_path[s:]
test_labels = all_labels[s:]# 将训练集和标签 制作dataset 需要转换为张量
train_ds = Mydatasetpro(train_imgs, train_labels, transform)  # TrainSet TensorData
test_ds = Mydatasetpro(test_imgs, test_labels, transform)  # TestSet TensorData
# print(train_ds)
# print(test_ds)
print("**********")
# 制作trainLoader
train_dl = data.DataLoader(train_ds, batch_size=BATCH_SIZE, shuffle=True)  # TrainSet Labels
test_dl = data.DataLoader(train_ds, batch_size=BATCH_SIZE, shuffle=True)  # TestSet Labels

训练代码

import torch
import torchvision.models as models
from torch import nn
from torch import optim
from DataSetMake import brake_dataloader
from DataSetMake import train_dl, test_dl# 判断是否使用GPU
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")#  使用resnet 训练
model_ft = models.resnet50(pretrained=True)  # 使用迁移学习，加载预训练权in_features = model_ft.fc.in_features
model_ft.fc = nn.Sequential(nn.Linear(in_features, 256),nn.ReLU(),# nn.Dropout(0, 4),nn.Linear(256, 4),nn.LogSoftmax(dim=1))model_ft = model_ft.to(DEVICE)  # 将模型迁移到gpu# 优化器
loss_fn = nn.CrossEntropyLoss()
loss_fn = loss_fn.to(DEVICE)  # 将loss_fn迁移到GPU# Adam损失函数
optimizer = optim.Adam(model_ft.fc.parameters(), lr=0.003)epochs = 50  # 迭代次数
steps = 0
running_loss = 0
print_every = 10
train_losses, test_losses = [], []for epoch in range(epochs):model_ft.train()# 遍历训练集数据for imgs, labels in brake_dataloader:steps += 1# 标签转换为 tensorlabels = torch.tensor(labels, dtype=torch.long)# 将图片和标签 放到设备上imgs, labels = imgs.to(DEVICE), labels.to(DEVICE)optimizer.zero_grad()  # 梯度归零#  前向推理outputs = model_ft(imgs)# 计算lossloss = loss_fn(outputs, labels)loss.backward()  # 反向传播计算梯度optimizer.step()  # 梯度优化# 累加lossrunning_loss += loss.item()if steps % print_every == 0:test_loss = 0accuracy = 0# 验证模式model_ft.eval()# 测试集 不需要计算梯度with torch.no_grad():# 遍历测试集数据for imgs, labels in test_dl:#  转换为tensorlabels = torch.tensor(labels, dtype=torch.long)#  数据标签 部署到gpuimgs, labels = imgs.to(DEVICE), labels.to(DEVICE)#  前向推理outputs = model_ft(imgs)#  计算损失loss = loss_fn(outputs, labels)# 累加测试机的损失test_loss += loss.item()ps = torch.exp(outputs)top_p, top_class = ps.topk(1, dim=1)equals = top_class == labels.view(*top_class.shape)accuracy += torch.mean(equals.type(torch.FloatTensor)).item()train_losses.append(running_loss / len(train_dl))test_losses.append(test_loss / len(test_dl))print(f"Epoch {epoch + 1}/{epochs}.. "f"Train loss: {running_loss / print_every:.3f}.. "f"Test loss: {test_loss / len(test_dl):.3f}.. "f"Test accuracy: {accuracy / len(test_dl):.3f}")#  回到训练模式 训练误差清0running_loss = 0model_ft.train()
torch.save(model_ft, "aerialmodel.pth")

预测代码

import os
import torch
from PIL import Image
from torch import nn
from torchvision import transforms, modelsi = 0  # 识别图片计数
# 这里最好新建一个test_data文件随机放一些上面整理好的图片进去
root_path = r"D:\CODE\ImageClassify\Test"  # 待测试文件夹
names = os.listdir(root_path)for name in names:print(name)i = i + 1data_class = ['sun', 'rain', 'cloud']  # 按文件索引顺序排列#  找出文件夹中的所有图片image_path = os.path.join(root_path, name)image = Image.open(image_path)print(image)#  张量定义格式transform = transforms.Compose([transforms.Resize((256, 256)),transforms.ToTensor()])# 图片转换为张量image = transform(image)print(image.shape)#  定义resnet模型model_ft = models.resnet50()# 模型结构in_features = model_ft.fc.in_featuresmodel_ft.fc = nn.Sequential(nn.Linear(in_features, 256),nn.ReLU(),# nn.Dropout(0, 4),nn.Linear(256, 4),nn.LogSoftmax(dim=1))# 加载已经训练好的模型参数model = torch.load("aerialmodel.pth", map_location=torch.device("cpu"))# 将每张图片 调整维度image = torch.reshape(image, (1, 3, 256, 256))  # 修改待预测图片尺寸，需要与训练时一致model.eval()#  速出预测结果with torch.no_grad():output = model(image)print(output)  # 输出预测结果# print(int(output.argmax(1)))# 对结果进行处理，使直接显示出预测的种类  根据索引判别是哪一类print("第{}张图片预测为：{}".format(i, data_class[int(output.argmax(1))]))

工程目录结构