TensorBoard(代码注释有一些对梯度计算,优化器,学习率,反向传播,损失这些基础概念的理解)

发布时间:2023年12月20日

1. 详细说明? ?进入pytorch官网-Tutorials-找到visuallizing Models,Data,and Training with TensorBoard.

tensorboard我的理解就是一个用来保存查看训练过程参数,训练情况的库,可以根据可视化的界面,理解模型的训练优化过程,并且可以利用得到的信息进行优化模型。

2. torch.utils.tensorboard.SummaryWriter("保存路径"),这里我运行在指定文件夹下有四个文件,应该是存的以下四种数据:

? ? ? ? add_graph()? 模型示意图

? ? ? ? add_scalar() 折线图

? ? ? ? add_figure() 预测图片在每一轮epoch的预测情况

? ? ? ? add_histogram()? 参数数量的直方图

3. 具体见代码和截图,打开tensorboard方法:进入命令行,激活相应环境(一定要进入相关环境才能用tensorboard.exe指令),进入SummaryWriter("保存路径")里的路径文件夹前一个文件夹,输入命令:

因为默认只显示10个epoch的预测情况,设置为50,才能显示完全。

tensorboard.exe --logdir=./ --samples_per_plugin=images=50  

? ??

训练完整代码:? 很多内容都在代码的注释里

train.py

import os
import math
import argparse

import torch
import torch.optim as optim
from torch.utils.tensorboard import SummaryWriter
from torchvision import transforms
import torch.optim.lr_scheduler as lr_scheduler

from model import resnet34
from my_dataset import MyDataSet
from data_utils import read_split_data, plot_class_preds
from train_eval_utils import train_one_epoch, evaluate


def main(args):
    device = torch.device(args.device if torch.cuda.is_available() else "cpu")

    print(args)
    print('Start Tensorboard with "tensorboard --logdir=runs", view at http://localhost:6006/')
    # 实例化SummaryWriter对象,log_dir是写入的目录
    tb_writer = SummaryWriter(log_dir="runs/flower_experiment")
    if os.path.exists("./weights") is False:
        os.makedirs("./weights")

    # 划分数据为训练集和验证集,训练集图片路径列表,对应训练集图片标签(这个标签是索引)
    train_images_path, train_images_label, val_images_path, val_images_label = read_split_data(args.data_path)

    # 定义训练以及预测时的预处理方法
    data_transform = {
        "train": transforms.Compose([transforms.RandomResizedCrop(224),
                                     transforms.RandomHorizontalFlip(),
                                     transforms.ToTensor(),
                                     transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])]),
        "val": transforms.Compose([transforms.Resize(256),
                                   transforms.CenterCrop(224),
                                   transforms.ToTensor(),
                                   transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])}

    # 实例化训练数据集
    train_data_set = MyDataSet(images_path=train_images_path,
                               images_class=train_images_label,
                               transform=data_transform["train"])

    # 实例化验证数据集
    val_data_set = MyDataSet(images_path=val_images_path,
                             images_class=val_images_label,
                             transform=data_transform["val"])

    batch_size = args.batch_size
    # 计算使用num_workers的数量 os.cpu_count() 是一个函数,用于获取系统中可用的 CPU 核心数量。
    nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, 8])  # number of workers
    print('Using {} dataloader workers every process'.format(nw))
    train_loader = torch.utils.data.DataLoader(train_data_set,
                                               batch_size=batch_size,
                                               shuffle=True,
                                               pin_memory=True, # 提高数据加载的效率
                                               num_workers=nw,
                                               collate_fn=train_data_set.collate_fn)

    val_loader = torch.utils.data.DataLoader(val_data_set,
                                             batch_size=batch_size,
                                             shuffle=False,
                                             pin_memory=True,
                                             num_workers=nw,
                                             collate_fn=val_data_set.collate_fn)

    # 实例化模型
    model = resnet34(num_classes=args.num_classes).to(device)

    # 将模型写入tensorboard, 用一个数据+模型就可以把模型写入tensorboatd
    init_img = torch.zeros((1, 3, 224, 224), device=device)
    tb_writer.add_graph(model, init_img)

    # 如果存在预训练权重则载入
    if os.path.exists(args.weights):
        weights_dict = torch.load(args.weights, map_location=device)
        load_weights_dict = {k: v for k, v in weights_dict.items()
                             # model.state_dict()[k].numel() 获取模型名k对应的参数的数量
                             if model.state_dict()[k].numel() == v.numel()}
        model.load_state_dict(load_weights_dict, strict=False)
    else:
        print("not using pretrain-weights.")

    # 是否冻结权重
    if args.freeze_layers:
        print("freeze layers except fc layer.")
        # model.named_parameters() 都是nn.Module里的方法,获取结构名和参数,
        for name, para in model.named_parameters():
            # 除最后的全连接层外,其他权重全部冻结
            if "fc" not in name:
                para.requires_grad_(False)

    # model.parameters() 获取需要训练的模型参数,p.requires_grad=True即需要梯度信息
    pg = [p for p in model.parameters() if p.requires_grad]
    # 学习率的动量参数,权重衰减系数定义参数优化器
    optimizer = optim.SGD(pg, lr=args.lr, momentum=0.9, weight_decay=0.005)
    # Scheduler https://arxiv.org/pdf/1812.01187.pdf
    # 定义学习率调度器
    lf = lambda x: ((1 + math.cos(x * math.pi / args.epochs)) / 2) * (1 - args.lrf) + args.lrf  # cosine
    # 将学习率调度器与优化器绑定,使得优化器能够根据学习率调度器的结果来更新学习率
    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)

    for epoch in range(args.epochs):
        # train
        mean_loss = train_one_epoch(model=model,
                                    optimizer=optimizer,
                                    data_loader=train_loader,
                                    device=device,
                                    epoch=epoch)
        # update learning rate
        # 每训练一个epoch,学习率进行一次更新,以便优化器根据新的学习率优化下一次epoch里的每个step的数据
        scheduler.step()

        # validate
        acc = evaluate(model=model,
                       data_loader=val_loader,
                       device=device)

        # add loss, acc and lr into tensorboard
        print("[epoch {}] accuracy: {}".format(epoch, round(acc, 3)))

        # 写入器 tb_writer  每个epoch写入一次,
        # 可以在 TensorBoard 的界面中查看、比较和分析标量值的变化趋势,帮助理解和监控训练过程中的指标。
        tags = ["train_loss", "accuracy", "learning_rate"]
        tb_writer.add_scalar(tags[0], mean_loss, epoch)
        tb_writer.add_scalar(tags[1], acc, epoch)
        tb_writer.add_scalar(tags[2], optimizer.param_groups[0]["lr"], epoch)

        # add figure into tensorboard
        # 用每个epoch的权重预测一遍
        fig = plot_class_preds(net=model,
                               images_dir="./plot_img",
                               transform=data_transform["val"],
                               num_plot=5,
                               device=device)
        if fig is not None:
            tb_writer.add_figure("predictions vs. actuals",
                                 figure=fig,
                                 global_step=epoch)

        # add conv1 weights into tensorboard   两个直方图
        tb_writer.add_histogram(tag="conv1",
                                values=model.conv1.weight,
                                global_step=epoch)
        tb_writer.add_histogram(tag="layer1/block0/conv1",
                                values=model.layer1[0].conv1.weight,
                                global_step=epoch)

        # save weights
        torch.save(model.state_dict(), "./weights/model-{}.pth".format(epoch))


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--num_classes', type=int, default=5)
    parser.add_argument('--epochs', type=int, default=30)
    parser.add_argument('--batch-size', type=int, default=16)
    # 学习率,控制参数更新步长的超参数
    parser.add_argument('--lr', type=float, default=0.001)
    parser.add_argument('--lrf', type=float, default=0.1)

    # 数据集所在根目录
    # https://storage.googleapis.com/download.tensorflow.org/example_images/flower_photos.tgz
    img_root = "../../data_set/flower_data/flower_photos"
    parser.add_argument('--data-path', type=str, default=img_root)

    # resnet34 官方权重下载地址
    # https://download.pytorch.org/models/resnet34-333f7ec4.pth
    # parser.add_argument('--weights', type=str, default='resNet34.pth',
    #                     help='initial weights path')
    # 这个不用预训练权重,用预训练权重tensorboard上训练曲线效果不明显
    parser.add_argument('--weights', type=str, default='',
                        help='initial weights path')
    # freeze-layers 是否冻结除全连接层外的其他层
    parser.add_argument('--freeze-layers', type=bool, default=False)
    parser.add_argument('--device', default='cuda', help='device id (i.e. 0 or 0,1 or cpu)')

    # parse_args() 用于解析命令行参数,返回一个包含解析结果的命名空间对象,可以通过属性访问这些参数的值
    # 在命令行运行脚本时可以通过传递相应的选项和参数来解析
    opt = parser.parse_args()

    main(opt)

mdoel.py

import torch.nn as nn
import torch


class BasicBlock(nn.Module):
    expansion = 1

    def __init__(self, in_channel, out_channel, stride=1, downsample=None):
        super(BasicBlock, self).__init__()
        self.conv1 = nn.Conv2d(in_channels=in_channel, out_channels=out_channel,
                               kernel_size=3, stride=stride, padding=1, bias=False)
        self.bn1 = nn.BatchNorm2d(out_channel)
        self.relu = nn.ReLU()
        self.conv2 = nn.Conv2d(in_channels=out_channel, out_channels=out_channel,
                               kernel_size=3, stride=1, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(out_channel)
        self.downsample = downsample

    def forward(self, x):
        identity = x
        if self.downsample is not None:
            identity = self.downsample(x)

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)

        out += identity
        out = self.relu(out)

        return out


class Bottleneck(nn.Module):
    expansion = 4

    def __init__(self, in_channel, out_channel, stride=1, downsample=None):
        super(Bottleneck, self).__init__()
        self.conv1 = nn.Conv2d(in_channels=in_channel, out_channels=out_channel,
                               kernel_size=1, stride=1, bias=False)  # squeeze channels
        self.bn1 = nn.BatchNorm2d(out_channel)
        # -----------------------------------------
        self.conv2 = nn.Conv2d(in_channels=out_channel, out_channels=out_channel,
                               kernel_size=3, stride=stride, bias=False, padding=1)
        self.bn2 = nn.BatchNorm2d(out_channel)
        # -----------------------------------------
        self.conv3 = nn.Conv2d(in_channels=out_channel, out_channels=out_channel*self.expansion,
                               kernel_size=1, stride=1, bias=False)  # unsqueeze channels
        self.bn3 = nn.BatchNorm2d(out_channel*self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample

    def forward(self, x):
        identity = x
        if self.downsample is not None:
            identity = self.downsample(x)

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)

        out = self.conv3(out)
        out = self.bn3(out)

        out += identity
        out = self.relu(out)

        return out


class ResNet(nn.Module):

    def __init__(self, block, blocks_num, num_classes=1000, include_top=True):
        super(ResNet, self).__init__()
        self.include_top = include_top
        self.in_channel = 64

        self.conv1 = nn.Conv2d(3, self.in_channel, kernel_size=7, stride=2,
                               padding=3, bias=False)
        self.bn1 = nn.BatchNorm2d(self.in_channel)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        self.layer1 = self._make_layer(block, 64, blocks_num[0])
        self.layer2 = self._make_layer(block, 128, blocks_num[1], stride=2)
        self.layer3 = self._make_layer(block, 256, blocks_num[2], stride=2)
        self.layer4 = self._make_layer(block, 512, blocks_num[3], stride=2)
        if self.include_top:
            self.avgpool = nn.AdaptiveAvgPool2d((1, 1))  # output size = (1, 1)
            self.fc = nn.Linear(512 * block.expansion, num_classes)

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')

    def _make_layer(self, block, channel, block_num, stride=1):
        downsample = None
        if stride != 1 or self.in_channel != channel * block.expansion:
            downsample = nn.Sequential(
                nn.Conv2d(self.in_channel, channel * block.expansion, kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(channel * block.expansion))

        layers = []
        layers.append(block(self.in_channel, channel, downsample=downsample, stride=stride))
        self.in_channel = channel * block.expansion

        for _ in range(1, block_num):
            layers.append(block(self.in_channel, channel))

        return nn.Sequential(*layers)

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        if self.include_top:
            x = self.avgpool(x)
            x = torch.flatten(x, 1)
            x = self.fc(x)

        return x


def resnet34(num_classes=1000, include_top=True):
    return ResNet(BasicBlock, [3, 4, 6, 3], num_classes=num_classes, include_top=include_top)


def resnet101(num_classes=1000, include_top=True):
    return ResNet(Bottleneck, [3, 4, 23, 3], num_classes=num_classes, include_top=include_top)

my_dataset.py

from tqdm import tqdm
from PIL import Image
import torch
from torch.utils.data import Dataset


class MyDataSet(Dataset):
    """自定义数据集"""

    def __init__(self, images_path: list, images_class: list, transform=None):
        self.images_path = images_path
        self.images_class = images_class
        self.transform = transform

        delete_img = []
        for index, img_path in tqdm(enumerate(images_path)):
            img = Image.open(img_path)
            w, h = img.size
            ratio = w / h
            if ratio > 10 or ratio < 0.1:
                delete_img.append(index)
                # print(img_path, ratio)

        for index in delete_img[::-1]:
            self.images_path.pop(index)
            self.images_class.pop(index)

    def __len__(self):
        return len(self.images_path)

    def __getitem__(self, item):
        img = Image.open(self.images_path[item])
        # RGB为彩色图片,L为灰度图片
        if img.mode != 'RGB':
            raise ValueError("image: {} isn't RGB mode.".format(self.images_path[item]))
        label = self.images_class[item]

        if self.transform is not None:
            img = self.transform(img)

        return img, label

    @staticmethod
    def collate_fn(batch):
        # 官方实现的default_collate可以参考
        # https://github.com/pytorch/pytorch/blob/67b7e751e6b5931a9f45274653f4f653a4e6cdf6/torch/utils/data/_utils/collate.py
        images, labels = tuple(zip(*batch))

        images = torch.stack(images, dim=0)
        labels = torch.as_tensor(labels)
        return images, labels

data_utils.py

import os
import json
import pickle
import random

from PIL import Image
import torch
import numpy as np
import matplotlib.pyplot as plt


def read_split_data(root: str, val_rate: float = 0.2):
    random.seed(0)  # 保证随机结果可复现
    assert os.path.exists(root), "dataset root: {} does not exist.".format(root)

    # 遍历文件夹,一个文件夹对应一个类别
    flower_class = [cla for cla in os.listdir(root) if os.path.isdir(os.path.join(root, cla))]
    # 排序,保证顺序一致,列表排序
    flower_class.sort()
    # 生成类别名称以及对应的数字索引  {"daisy":0,"dandelion":1,"roses":2,"sunflowers":3,"tulips":4}
    class_indices = dict((k, v) for v, k in enumerate(flower_class))
    json_str = json.dumps(dict((val, key) for key, val in class_indices.items()), indent=4)
    with open('class_indices.json', 'w') as json_file:
        json_file.write(json_str)

    train_images_path = []  # 存储训练集的所有图片路径
    train_images_label = []  # 存储训练集图片对应索引信息
    val_images_path = []  # 存储验证集的所有图片路径
    val_images_label = []  # 存储验证集图片对应索引信息
    every_class_num = []  # 存储每个类别的样本总数
    supported = [".jpg", ".JPG", ".png", ".PNG"]  # 支持的文件后缀类型
    # 遍历每个文件夹下的文件
    for cla in flower_class:
        cla_path = os.path.join(root, cla)
        # 遍历获取supported支持的所有文件路径,os.path.join()可以拼接若干个字符串
        images = [os.path.join(root, cla, i) for i in os.listdir(cla_path)
                  if os.path.splitext(i)[-1] in supported]
        # 获取该类别对应的索引
        image_class = class_indices[cla]
        # 记录该类别的样本数量
        every_class_num.append(len(images))
        # 按比例随机采样验证样本
        val_path = random.sample(images, k=int(len(images) * val_rate))

        for img_path in images:
            if img_path in val_path:  # 如果该路径在采样的验证集样本中则存入验证集
                val_images_path.append(img_path)  # 验证图片的路径列表
                val_images_label.append(image_class) # 验证图片对应的标签列表
            else:  # 否则存入训练集
                train_images_path.append(img_path)
                train_images_label.append(image_class)

    print("{} images were found in the dataset.".format(sum(every_class_num)))
    print("{} images for training.".format(len(train_images_path)))
    print("{} images for validation.".format(len(val_images_path)))

    plot_image = False
    if plot_image:
        # 绘制每种类别个数柱状图
        plt.bar(range(len(flower_class)), every_class_num, align='center')
        # 将横坐标0,1,2,3,4替换为相应的类别名称
        plt.xticks(range(len(flower_class)), flower_class)
        # 在柱状图上添加数值标签
        for i, v in enumerate(every_class_num):
            plt.text(x=i, y=v + 5, s=str(v), ha='center')
        # 设置x坐标
        plt.xlabel('image class')
        # 设置y坐标
        plt.ylabel('number of images')
        # 设置柱状图的标题
        plt.title('flower class distribution')
        plt.show()

    return train_images_path, train_images_label, val_images_path, val_images_label


def plot_data_loader_image(data_loader):
    batch_size = data_loader.batch_size
    plot_num = min(batch_size, 4)

    json_path = './class_indices.json'
    assert os.path.exists(json_path), json_path + " does not exist."
    json_file = open(json_path, 'r')
    class_indices = json.load(json_file)

    for data in data_loader:
        images, labels = data
        for i in range(plot_num):
            # [C, H, W] -> [H, W, C]
            img = images[i].numpy().transpose(1, 2, 0)
            # 反Normalize操作
            img = (img * [0.229, 0.224, 0.225] + [0.485, 0.456, 0.406]) * 255
            label = labels[i].item()
            plt.subplot(1, plot_num, i+1)
            plt.xlabel(class_indices[str(label)])
            plt.xticks([])  # 去掉x轴的刻度
            plt.yticks([])  # 去掉y轴的刻度
            plt.imshow(img.astype('uint8'))
        plt.show()


def write_pickle(list_info: list, file_name: str):
    with open(file_name, 'wb') as f:
        pickle.dump(list_info, f)


def read_pickle(file_name: str) -> list:
    with open(file_name, 'rb') as f:
        info_list = pickle.load(f)
        return info_list


def plot_class_preds(net,
                     images_dir: str,
                     transform,
                     num_plot: int = 5,
                     device="cpu"):
    if not os.path.exists(images_dir):
        print("not found {} path, ignore add figure.".format(images_dir))
        return None

    label_path = os.path.join(images_dir, "label.txt")
    if not os.path.exists(label_path):
        print("not found {} file, ignore add figure".format(label_path))
        return None

    # read class_indict
    json_label_path = './class_indices.json'
    assert os.path.exists(json_label_path), "not found {}".format(json_label_path)
    json_file = open(json_label_path, 'r')
    # {"0": "daisy"}   json.load() 返回一个json文件里的字典
    flower_class = json.load(json_file)
    # {"daisy": "0"}
    class_indices = dict((v, k) for k, v in flower_class.items())

    # reading label.txt file
    label_info = []
    with open(label_path, "r") as rd:
        for line in rd.readlines():
            line = line.strip()
            if len(line) > 0:
                # 读取到的txt里的一行以空格分格,返回一个这行元素的列表
                split_info = [i for i in line.split(" ") if len(i) > 0]
                # 这行元素生成的列表一定是两个元素,一个是图片名,一个是类别名称
                assert len(split_info) == 2, "label format error, expect file_name and class_name"
                image_name, class_name = split_info
                image_path = os.path.join(images_dir, image_name)
                # 如果文件不存在,则跳过
                if not os.path.exists(image_path):
                    print("not found {}, skip.".format(image_path))
                    continue
                # 如果读取的类别不在给定的类别内,则跳过
                if class_name not in class_indices.keys():
                    print("unrecognized category {}, skip".format(class_name))
                    continue
                # label_info 是获取到的图片名和对应类型的二维列表,因为把每行的两个元素都加进去了
                label_info.append([image_path, class_name])

    if len(label_info) == 0:
        return None

    # get first num_plot info  num_plot是画预测图的数量限制,如果txt文件过多,只取前num_plot个
    if len(label_info) > num_plot:
        label_info = label_info[:num_plot]

    num_imgs = len(label_info)
    images = []
    labels = []
    for img_path, class_name in label_info:
        # read img
        img = Image.open(img_path).convert("RGB")
        label_index = int(class_indices[class_name])

        # preprocessing
        img = transform(img)
        images.append(img)
        labels.append(label_index)

    # batching images  添加维度,把要预测的经过预处理后的图片张量打包成一批,生成images张量
    images = torch.stack(images, dim=0).to(device)

    # inference
    with torch.no_grad():
        # output: [batch_pred,num_cls]
        output = net(images)
        # softmax 函数常用于将模型输出转换为概率分布,用于多分类任务的概率预测
        # torch.softmax(output, dim=1) 将在维度 dim=1 上计算张量 output 的 softmax 函数值。
        # 这意味着将对 output 张量中的每一行应用 softmax 函数,并返回具有相同形状的张量,其中每一行的元素之和等于 1
        # torch.softmax(output, dim=1)返回一个具有相同形状的张量。
        # torch.max() 函数在这个 softmax 张量上再次使用 dim=1,来获取每一行的最大值和最大值的索引
        # 这个索引就是预测类别preds,probs长是预测batch的一维张量,preds长也是预测batch的一维张量,batch就是batch_pred
        # 一个batch有多少张图片,预测矩阵张量就有多少行,预测的有多少类,就有多少列
        probs, preds = torch.max(torch.softmax(output, dim=1), dim=1)
        # 通过调用 cpu() 方法,将它们从 GPU 设备上移动到 CPU 设备上。
        """
        为什么要移动到cpu上?
        数据后处理:有时候,在模型的输出上需要进行一些后处理操作,例如计算指标、可视化结果或保存结果到磁盘。
            这些操作通常在 CPU 上执行,因为 CPU 上的计算能力足够处理这些任务。
        与其他库的兼容性:某些库或工具可能对 CPU 上的数据更加友好,而不是 GPU 上的数据。
            例如,一些常用的数据处理、可视化或存储库可能更容易与 CPU 上的数据进行交互,所以需要将 GPU 上的数据转换为 CPU 上的数据。
        内存限制:在处理大规模数据集或模型时,GPU 的显存可能会受到限制。
            如果计算结果的大小超过 GPU 的显存容量,那么需要将其移动到 CPU 上进行处理。
        模型部署:在将训练好的模型部署到生产环境中时,可能需要将模型的输出从 GPU 移动到 CPU 上,
            因为生产环境通常没有 GPU 或者 GPU 的数量有限。
        """
        probs = probs.cpu().numpy()
        preds = preds.cpu().numpy()

    # width, height  每张图片显示宽是250,高是300,共num_imgs张图片
    fig = plt.figure(figsize=(num_imgs * 2.5, 3), dpi=100)
    for i in range(num_imgs):
        # 1:子图共1行,num_imgs:子图共num_imgs列,当前绘制第i+1个子图 xticks=[], yticks=[]清空坐标刻度
        ax = fig.add_subplot(1, num_imgs, i+1, xticks=[], yticks=[])

        # CHW -> HWC
        npimg = images[i].cpu().numpy().transpose(1, 2, 0)

        # 将图像还原至标准化之前
        # mean:[0.485, 0.456, 0.406], std:[0.229, 0.224, 0.225]
        npimg = (npimg * [0.229, 0.224, 0.225] + [0.485, 0.456, 0.406]) * 255
        plt.imshow(npimg.astype('uint8'))

        title = "{}, {:.2f}%\n(label: {})".format(
            flower_class[str(preds[i])],  # predict class
            probs[i] * 100,  # predict probability
            flower_class[str(labels[i])]  # true class
        )
        ax.set_title(title, color=("green" if preds[i] == labels[i] else "red"))

    return fig


train_eval_utils.py

import sys

from tqdm import tqdm
import torch


def train_one_epoch(model, optimizer, data_loader, device, epoch):
    """
    在深度学习中,model.train()是一个用于设置模型为训练模式的方法。调用该方法后,模型会进行以下操作:
    激活训练模式:设置模型的train属性为True,表示将进入训练模式。这是为了告诉模型在训练期间需要执行训练相关的操作,例如启用Dropout和Batch Normalization层的训练行为。
    梯度计算:启用梯度计算,以便在反向传播过程中计算和累积参数的梯度。这是为了在训练过程中更新模型的权重和偏置。
    批归一化更新:在训练模式下,批归一化层会根据当前批次的输入数据更新其内部统计信息,例如均值和方差。这是为了保证每个批次的数据都能够获得相对一致的归一化效果。
    Dropout激活:在训练模式下,Dropout层将以一定的概率随机丢弃部分神经元的输出。这有助于提高模型的泛化能力和防止过拟合。
    总的来说,model.train()方法的调用是为了将模型设置为训练模式,以便进行参数更新和训练相关的操作。

    要先在反向传播里计算参数的梯度,才有用优化器根据学习率调度规则来优化训练的参数
    """
    model.train()
    loss_function = torch.nn.CrossEntropyLoss()
    # 用于累积每个批次的损失值
    mean_loss = torch.zeros(1).to(device)
    # 使用optimizer.zero_grad()将优化器中的梯度信息清零。这是为了确保在每个批次开始时,梯度不会累积之前的批次的梯度信息。
    optimizer.zero_grad()

    # 将 data_loader 中的迭代过程包装在一个进度条中,并将进度信息输出到标准输出(stdout)
    data_loader = tqdm(data_loader, file=sys.stdout)
    for step, data in enumerate(data_loader):
        images, labels = data
        pred = model(images.to(device))

        """
        softmax 归一化处理,确保每个样本的预测概率之和等于1。这样做是为了将预测输出转换为概率分布,使其表示模型对每个类别的预测概率。
        """
        # 具体损失函数做了什么,pred是一个 batch_size * num_cls 的二维张量
        # labels是一个存有类别索引的 长度为 batch_size 的一维张量
        # 交叉熵损失函数先把labels根据索引类别变成 batch_size * num_cls 的二维独热编码张量
        # 然后根据真实标签的概率和预测的相应位置上的概率计算损失
        # 最后,对这些交叉熵值进行平均或求和操作,得到一个表示整个批次的损失值
        # 通常情况下,我们会将这个损失值作为模型在当前批次数据上的损失。
        loss = loss_function(pred, labels.to(device))
        # 损失的反向传播
        """
        backward() 方法的作用是计算损失函数关于模型参数的梯度。具体来说,它执行以下操作:
        从损失值开始,通过链式法则计算梯度:PyTorch 中的自动求导机制会根据计算图中的节点,自动计算每个节点相对于损失的梯度。这是通过应用链式法则来实现的,从损失节点开始,逐层向后传播计算梯度。
        累积梯度:梯度计算得到的梯度值会被累积到每个参数的 .grad 属性中。
        梯度计算完成:一旦完成梯度计算,可以使用得到的梯度值来执行参数更新和优化算法,例如使用梯度下降算法来更新模型的参数。
        """
        loss.backward()
        """
        loss.detach() 方法执行以下操作:
        分离张量:该方法创建一个新的张量,其值与原始张量相同,但不再与计算图关联。
        断开梯度计算:分离的张量将不再保留梯度信息。这意味着对分离张量进行任何操作或计算梯度时,梯度将不会回传到原始张量。
        """
        # 计算迭代到当前批次的前面所有批次和这批的平均损失,这是一个批次平均损失,每个批次的损失里包括每个图片张量的损失,可以是一批图片
        # 损失的平均值或和
        mean_loss = (mean_loss * step + loss.detach()) / (step + 1)  # update mean losses

        # 打印平均loss
        # mean_loss.item() 是一个张量的方法,用于获取张量的数值表示
        # round(mean_loss.item(), 3) 将损失值保留三位小数
        # 当使用 tqdm 来迭代一个可迭代对象时,可以通过设置 desc 参数来为进度条添加描述信息。
        # 每迭代一次data_loader,都要更新一次下面的进度条描述信息
        data_loader.desc = "[epoch {}] mean loss {}".format(epoch, round(mean_loss.item(), 3))

        # torch.isfinite(loss) 用于检查损失值是否为有限值。
        #
        if not torch.isfinite(loss):
            print('WARNING: non-finite loss, ending training ', loss)
            sys.exit(1)

        # 参数更新:如果损失值是有限值,代码会执行 optimizer.step() 来更新模型的参数。
        # 这一步是优化算法的核心,根据计算得到的梯度来调整模型参数,以最小化损失函数。
        # 清空梯度:在参数更新之后,为了避免梯度的累积影响后续的迭代,
        # 代码会执行 optimizer.zero_grad() 来清空之前计算的梯度值,为下一次迭代做准备。
        # 定义优化器时已经传入了需要进行训练的参数(梯度计算的参数),
        # 通过para.requires_grad_(False)也设置了模型里不需要求梯度的参数,loss.backward()也不会求损失对应相关参数的梯度
        # 这样optimizer.step()优化器进行用定义优化器时传入的学习率更新策略优化时就只根据loss.backward()求得的梯度优化传入优化器的需要优化的参数
        optimizer.step()
        optimizer.zero_grad()

    return mean_loss.item()


@torch.no_grad()
def evaluate(model, data_loader, device):
    model.eval()

    # 用于存储预测正确的样本个数   torch.zeros(1)创建一个1*n的元素全是0的张量,其实就是一个一维张量,一维是线,点是0维
    sum_num = torch.zeros(1).to(device)
    # 统计验证集样本总数目,这个总数目不是批次数,而是验证集的所有图片数
    num_samples = len(data_loader.dataset)

    # 打印验证进度
    data_loader = tqdm(data_loader, desc="validation...", file=sys.stdout)

    for step, data in enumerate(data_loader):
        images, labels = data
        pred = model(images.to(device))
        # dim等于1表示取每行最大值及其索引,dim=0表示取每一列上最大值及其索引,
        # torch.max(pred, dim=1)返回的是两个一维张量,第一个是每一行最大值,第二个是对应的索引
        # [1] 返回的是第二个输出,即每行上最大值对应的索引,其实就是类别,预测的类别
        pred = torch.max(pred, dim=1)[1]
        # pred是一个一维张量,包含一个批次的每张图片的索引,共batch_size个元素,
        # torch.eq(pred, labels.to(device)) 对比预测的一维张量和从dataset中获取的labels这个标签一维张量有几个位置的元素相同
        # 即有几张图片预测正确了,sum()求得一个批次中预测正确的图片数
        sum_num += torch.eq(pred, labels.to(device)).sum()

    # 计算预测正确的比例
    # item() 方法用于将张量中的单个元素提取为 Python 标量值,
    # 一个epoch中所有批次的所有图片的预测正确总数/验证集中图片总数   求得预测的正确率,进行返回
    acc = sum_num.item() / num_samples

    return acc






训练过程:

3670 images were found in the dataset.
2939 images for training.
731 images for validation.
2939it [00:00, 4558.29it/s]
731it [00:00, 4928.77it/s]
Using 8 dataloader workers every process
not using pretrain-weights.
[epoch 0] mean loss 1.416: 100%|██████████| 184/184 [00:42<00:00,  4.30it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.22it/s]
[epoch 0] accuracy: 0.473
[epoch 1] mean loss 1.27: 100%|██████████| 184/184 [00:42<00:00,  4.32it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.22it/s]
[epoch 1] accuracy: 0.59
[epoch 2] mean loss 1.155: 100%|██████████| 184/184 [00:42<00:00,  4.36it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.24it/s]
[epoch 2] accuracy: 0.64
[epoch 3] mean loss 1.088: 100%|██████████| 184/184 [00:42<00:00,  4.36it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.24it/s]
[epoch 3] accuracy: 0.662
[epoch 4] mean loss 1.043: 100%|██████████| 184/184 [00:42<00:00,  4.35it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.20it/s]
[epoch 4] accuracy: 0.683
[epoch 5] mean loss 1.019: 100%|██████████| 184/184 [00:42<00:00,  4.34it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.22it/s]
[epoch 5] accuracy: 0.644
[epoch 6] mean loss 0.979: 100%|██████████| 184/184 [00:42<00:00,  4.35it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.20it/s]
[epoch 6] accuracy: 0.705
[epoch 7] mean loss 0.959: 100%|██████████| 184/184 [00:42<00:00,  4.35it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.22it/s]
[epoch 7] accuracy: 0.713
[epoch 8] mean loss 0.913: 100%|██████████| 184/184 [00:42<00:00,  4.35it/s]
validation...: 100%|██████████| 46/46 [00:21<00:00,  2.18it/s]
[epoch 8] accuracy: 0.715
[epoch 9] mean loss 0.881: 100%|██████████| 184/184 [00:43<00:00,  4.27it/s]
validation...: 100%|██████████| 46/46 [00:21<00:00,  2.18it/s]
[epoch 9] accuracy: 0.741
[epoch 10] mean loss 0.843: 100%|██████████| 184/184 [00:42<00:00,  4.32it/s]
validation...: 100%|██████████| 46/46 [00:21<00:00,  2.17it/s]
[epoch 10] accuracy: 0.741
[epoch 11] mean loss 0.846: 100%|██████████| 184/184 [00:42<00:00,  4.35it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.22it/s]
[epoch 11] accuracy: 0.74
[epoch 12] mean loss 0.839: 100%|██████████| 184/184 [00:42<00:00,  4.34it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.22it/s]
[epoch 12] accuracy: 0.759
[epoch 13] mean loss 0.808: 100%|██████████| 184/184 [00:42<00:00,  4.33it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.22it/s]
[epoch 13] accuracy: 0.713
[epoch 14] mean loss 0.769: 100%|██████████| 184/184 [00:42<00:00,  4.35it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.22it/s]
[epoch 14] accuracy: 0.767
[epoch 15] mean loss 0.759: 100%|██████████| 184/184 [00:42<00:00,  4.35it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.24it/s]
[epoch 15] accuracy: 0.717
[epoch 16] mean loss 0.759: 100%|██████████| 184/184 [00:42<00:00,  4.35it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.22it/s]
[epoch 16] accuracy: 0.756
[epoch 17] mean loss 0.711: 100%|██████████| 184/184 [00:42<00:00,  4.32it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.22it/s]
[epoch 17] accuracy: 0.784
[epoch 18] mean loss 0.73: 100%|██████████| 184/184 [00:42<00:00,  4.35it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.24it/s]
[epoch 18] accuracy: 0.766
[epoch 19] mean loss 0.681: 100%|██████████| 184/184 [00:42<00:00,  4.35it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.22it/s]
[epoch 19] accuracy: 0.787
[epoch 20] mean loss 0.66: 100%|██████████| 184/184 [00:42<00:00,  4.35it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.22it/s]
[epoch 20] accuracy: 0.776
[epoch 21] mean loss 0.658: 100%|██████████| 184/184 [00:42<00:00,  4.33it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.22it/s]
[epoch 21] accuracy: 0.778
[epoch 22] mean loss 0.663: 100%|██████████| 184/184 [00:42<00:00,  4.35it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.22it/s]
[epoch 22] accuracy: 0.782
[epoch 23] mean loss 0.623: 100%|██████████| 184/184 [00:42<00:00,  4.35it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.22it/s]
[epoch 23] accuracy: 0.796
[epoch 24] mean loss 0.589: 100%|██████████| 184/184 [00:42<00:00,  4.35it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.21it/s]
[epoch 24] accuracy: 0.8
[epoch 25] mean loss 0.616: 100%|██████████| 184/184 [00:42<00:00,  4.32it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.20it/s]
[epoch 25] accuracy: 0.802
[epoch 26] mean loss 0.586: 100%|██████████| 184/184 [00:42<00:00,  4.35it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.21it/s]
[epoch 26] accuracy: 0.811
[epoch 27] mean loss 0.569: 100%|██████████| 184/184 [00:42<00:00,  4.35it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.22it/s]
[epoch 27] accuracy: 0.814
[epoch 28] mean loss 0.559: 100%|██████████| 184/184 [00:42<00:00,  4.35it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.22it/s]
[epoch 28] accuracy: 0.817
[epoch 29] mean loss 0.567: 100%|██████████| 184/184 [00:42<00:00,  4.35it/s]
validation...: 100%|██████████| 46/46 [00:20<00:00,  2.22it/s]
[epoch 29] accuracy: 0.814

打开tensorboard面板:

由于是在每个epoch训练和验证完后进行一次对预测图片的的预测,所以可以查看每个epoch,在tensorboard界面中是step表示,的预测情况,具体打印的信息在代码中进行控制。

?GRAPH可以查看详细数据在模型里的流向情况。

在代码里定义的要查看的每一层的权重参数的分布情况。

DISTRIBUTED是tensorboard提供的另一种查看方式。

文章来源:https://blog.csdn.net/m0_56294205/article/details/135115249
本文来自互联网用户投稿,该文观点仅代表作者本人,不代表本站立场。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。