12.6深度学习_模型优化和迁移_整体流程梳理

七、整体流程梳理

1. 引入使用的包

用到什么包，临时引入就可以，不用太担心。

import time
import osimport numpy as np
import pandas as pd
import torch
import torch.nn as nn
import torch.optim as optim
import torchvision
import torchvision.transforms as transforms
from torch.utils.data import DataLoader
from torchvision.datasets import CIFAR10from torchvision.models import resnet18, ResNet18_Weights
import wandb
from torch.utils.tensorboard import SummaryWriter
from sklearn.metrics import *
import matplotlib.pyplot as plt

2. 数据

# 下面和以前就一样了
train_dataset = CIFAR10(root=datapath,train=True,download=True,transform=transform,
)
# 构建训练数据集
train_loader = DataLoader(#dataset=train_dataset,batch_size=batzh_size,shuffle=True,num_workers=2,
)

3. 模型

# 再次获取resnet18原始神经网络并对齐fc层进行调整
model = resnet18(weights=None)in_features = model.fc.in_features
# 重写FC：我们这里做的是10分类
model.fc = nn.Linear(in_features=in_features, out_features=10)# 需要对权重信息进行处理：要加载我们训练之后最新的权重文件
weights_default = torch.load(weightpath)
weights_default.pop("fc.weight")
weights_default.pop("fc.bias")# 把权重参数进行同步
new_state_dict = model.state_dict()
weights_default_process = {k: v for k, v in weights_default.items() if k in new_state_dict
}
new_state_dict.update(weights_default_process)
model.load_state_dict(new_state_dict)
model.to(device)

4. 训练

4.1 数据增强

为了防止过拟合，增加模型的泛化能力，我们会数据增强

transform = transforms.Compose([transforms.RandomRotation(45),  # 随机旋转，-45到45度之间随机选transforms.RandomCrop(32, padding=4),  # 随机裁剪transforms.RandomHorizontalFlip(p=0.5),  # 随机水平翻转 选择一个概率概率transforms.RandomVerticalFlip(p=0.5),  # 随机垂直翻转transforms.ToTensor(),transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2471, 0.2435, 0.2616)),]
)transformtest = transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2471, 0.2435, 0.2616)),]
)

4.2 开始训练

    # 损失函数和优化器loss_fn = nn.CrossEntropyLoss()optimizer = optim.Adam(model.parameters(), lr=lr)for epoch in range(epochs):# 开始时间start = time.time()# 总的损失值total_loss = 0.0# 样本数量：最后一次样本数量不是128samp_num = 0# 总的预测正确的分类correct = 0model.train()for i, (x, y) in enumerate(train_loader):x, y = x.to(device), y.to(device)# 累加样本数量samp_num += len(y)out = model(x)# 预测正确的样本数量correct += out.argmax(dim=1).eq(y).sum().item()loss = loss_fn(out, y)# 损失率累加total_loss += loss.item() * len(y)optimizer.zero_grad()loss.backward()optimizer.step()if i % 100 == 0:img_grid = torchvision.utils.make_grid(x)write1.add_image(f"r_m_{epoch}_{i}", img_grid, epoch * len(train_loader) + i)print("批次：%d 损失率：%.4f 准确率：%.4f 耗时：%.4f"% (epoch, total_loss / samp_num, correct / samp_num, time.time() - start))# log metrics to wandbwandb.log({"acc": correct / samp_num, "loss": total_loss / samp_num})

4.3 保存模型

torch.save(model.state_dict(), weightpath)

4.4 训练过程可视化

wandb

 # 训练过程可视化wandb.init(project="my-qianyi-project",config={"learning_rate": lr,"architecture": "CNN","dataset": "CIFAR-100","batch_size": batzh_size,"epochs": epochs,},)

tensorboard

write1 = SummaryWriter(log_dir=log_dir)
# 保存模型结构到tensorboard
write1.add_graph(model, input_to_model=torch.randn(1, 3, 32, 32).to(device=device))

5. 验证阶段

5.1 数据验证

weights_default = torch.load(weightpath)# 再次获取resnet18原始神经网络并对齐fc层进行调整model = resnet18(pretrained=False)in_features = model.fc.in_features# 重写FC：我们这里做的是10分类model.fc = nn.Linear(in_features=in_features, out_features=10)model.load_state_dict(weights_default)model.to(device)model.eval()samp_num = 0correct = 0data2csv = np.empty(shape=(0, 13))for x, y in vaild_loader:x = x.to(device)y = y.to(device)# 累加样本数量samp_num += len(y)# 模型运算out = model(x)# 数组的合并data2csv = np.concatenate((data2csv, outdata_softmax), axis=0)# 预测正确的样本数量correct += out.argmax(dim=1).eq(y).sum().item()print("准确率：%.4f" % (correct / samp_num))

5.2 验证结果可视化

验证数据保存到Excel

data2csv = np.empty(shape=(0, 13))#数据整理
out = model(x)
outdata = out.cpu().detach()
outdata_softmax = torch.softmax(outdata, dim=1)
# 合并目标值到样本  [5, 7,9,0,1,,1,2,3,4,3,4]
outdata_softmax = np.concatenate((# 本身预测的值outdata_softmax.numpy(),# 真正的目标值y.cpu().numpy().reshape(-1, 1),# 预测值outdata_softmax.argmax(dim=1).reshape(-1, 1),# 分类名称np.array([vaild_dataset.classes[i] for i in y.cpu().numpy()]).reshape(-1, 1),),axis=1,
)
# 数组的合并
data2csv = np.concatenate((data2csv, outdata_softmax), axis=0)#写入CSV
columns = np.concatenate((vaild_dataset.classes, ["target", "prep", "分类"]))
pddata = pd.DataFrame(data2csv, columns=columns)
pddata.to_csv(csvpath, encoding="GB2312")

指标分析：可视化

def analy():# 读取csv数据data1 = pd.read_csv(csvpath, encoding="GB2312")print(type(data1))# 整体数据分析报告report = classification_report(y_true=data1["target"].values,y_pred=data1["prep"].values,)print(report)# 准确度 Accprint("准确度Acc：",accuracy_score(y_true=data1["target"].values,y_pred=data1["prep"].values,),)# 精确度print("精确度Precision：",precision_score(y_true=data1["target"].values, y_pred=data1["prep"].values, average="macro"),)# 召回率print("召回率Recall：",recall_score(# 100y_true=data1["target"].values,y_pred=data1["prep"].values,average="macro",),)# F1 Scoreprint("F1 Score：",f1_score(y_true=data1["target"].values,y_pred=data1["prep"].values,average="macro",),)passdef matrix():# 读取csv数据data1 = pd.read_csv(csvpath, encoding="GB2312", index_col=0)confusion = confusion_matrix(# 0y_true=data1["target"].values,y_pred=data1["prep"].values,# labels=data1.columns[0:10].values,)print(confusion)# 绘制混淆矩阵plt.rcParams["font.sans-serif"] = ["SimHei"]plt.rcParams["axes.unicode_minus"] = Falseplt.matshow(confusion, cmap=plt.cm.Greens)plt.colorbar()for i in range(confusion.shape[0]):for j in range(confusion.shape[1]):plt.text(j, i, confusion[i, j], ha="center", va="center", color="b")plt.title("验证数据混淆矩阵")plt.xlabel("Predicted label")plt.xticks(range(10), data1.columns[0:10].values, rotation=45)plt.ylabel("True label")plt.yticks(range(10), data1.columns[0:10].values)plt.show()

6. 使用

def app():dir = os.path.dirname(__file__)imgpath = os.path.join("./write", "6.png")# 读取图像文件 '8.png'img = cv2.imread(imgpath)# 将图像转换为灰度图img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)# 对灰度图进行二值化处理，采用OTSU自适应阈值方法，并反转颜色ret, img = cv2.threshold(img, 0, 255, cv2.THRESH_BINARY_INV | cv2.THRESH_OTSU)plt.imshow(img)plt.show()# img = cv2.resize(img, (32, 32))img = torch.Tensor(img).unsqueeze(0)transform = transforms.Compose([transforms.Resize((32, 32)),  # 调整输入图像大小为32x32transforms.ToTensor(),transforms.Normalize((0.1307,), (0.3081,)),])img = transform(img).unsqueeze(0)# 加载我们的模型net = LeNet5()net.load_state_dict(torch.load(modelpath))# 预测outputs = net(img)print(outputs)print(outputs.argmax(axis=1))