《TensorBoard简单使用指南》

简介

• TensorBoard，可视化工具

• 原本是tensorflow的可视化工具，pytorch从1.2.0开始支持tensorboard。之前的版本也可以使用tensorboardX代替

• 可视化工具，展示网络图、张量的指标变化、张量的分布

• TensorBoard运行时开启本地服务器，监听6006端口。在浏览器发出请求时，获取数据并绘制可视化图像

• 主要功能：

  （1）Scalars:展示训练过程中的准确率、损失值、权重/偏置的变化情况。

  （2）Images:展示训练过程中记录的图像。

  （3）Audio:展示训练过程中记录的音频。

  （4）Graphs:展示模型的数据流图，以及训练在各个设备上消耗的内存和时间。

  （5）Distributions:展示训练过程中记录的数据的分部图。

  （6）Histograms:展示训练过程中记录的数据的柱状图。

  （7）Embeddings:展示词向量后的投影分部。

使用

准备

• 安装

pip install tensorboard

• 导入

from torch.utils.tensorboard import SummaryWriter

• 命令行使用(环境变量)

tensorboard --logdir=logs

• 端口默认6006，冲突，指定端口

tensorboard --logdir=logs --port 8888

• 指定主机

tensorboard --logdir=logs --host 127.0.0.1

使用步骤

• 使用步骤如下：

  • 读取数据并转换

  • 设置TensorBoard,设置保存路径

    writer = SummaryWriter('logs')
    

  • 写入TensorBoard

  • 使用tensorboard检查模型架构

    • 跟踪训练模型时表现
    • 训练后评估模型的表现

  • 使用tensorboard创建可视化的交互式版本

    tensorboard --logdir=logs
    

标量可视化

• add_scalar(self, tag, scalar_value, global_step=None, walltime=None, new_style=False)
  • tag：标签
  • scalar_value：监测变量的值
  • global_step：步长索引

writer = SummaryWriter("logs/log_scalar")
for i in range(1, 100):
    writer.add_scalar('y=2x', i * 2, i)
writer.close()

图像可视化

• add_image(self, tag, img_tensor, global_step=None, walltime=None, dataformats='CHW')
  • 支持的图像格式 orch.Tensor, numpy.array
  • 默认shape(3, H, W)，其余需要指定格式 ，格式是RGB格式
• opencv读取图片

img = cv2.imdecode(np.fromfile(path, dtype='uint8'), -1)
writer.add_image('img_read', img, 0, dataformats='HWC')

• 栅格图像

img_grid = torchvision.utils.make_grid(images)
writer.add_image('four_fashion_mnist_images', img_grid)

模型结构可视化

• add_graph：可视化复杂模型结构

writer.add_graph(net, images)

模型训练监测

• test_accuracy

  

• test_loss

  

• train_loss

  

import torchvision, torch
from torch import nn
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter

class Model(nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.model = nn.Sequential(
            nn.Conv2d(3, 32, 5, 1, 2),
            nn.MaxPool2d(2),
            nn.Conv2d(32, 32, 5, 1, 2),
            nn.MaxPool2d(2),
            nn.Conv2d(32, 64, 5, 1, 2),
            nn.MaxPool2d(2),
            nn.Flatten(),
            nn.Linear(64*4*4, 64),
            nn.Linear(64, 10)
        )
    def forward(self, x):
        x = self.model(x)
        return x

train_dataset = torchvision.datasets.CIFAR10(
    root=r"E:\Data\CIFAR10",
    train=True,
    download=False,
    transform=torchvision.transforms.ToTensor()
)

test_dataset = torchvision.datasets.CIFAR10(
    root=r"E:\Data\CIFAR10",
    train=False,
    transform=torchvision.transforms.ToTensor()
)

train_dataset_size = len(train_dataset)
test_dataset_size = len(test_dataset)
print("训练数据集的长度为:{}".format(train_dataset_size))
print("测试数据集的长度为:{}".format(test_dataset_size))

train_dataloader = DataLoader(train_dataset, batch_size=64)
test_dataloader = DataLoader(test_dataset, batch_size=64)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# 创建神经网络
net = Model()
net = net.to(device)

# 损失函数
loss_fn = nn.CrossEntropyLoss()
loss_fn = loss_fn.to(device)

# 优化器
learning_rate = 1e-2
optimizer = torch.optim.SGD(net.parameters(), lr = learning_rate)

# 设置训练网络的一些参数
total_train_step = 0 # 记录训练次数
total_test_step = 0 # 记录测试次数
epoch = 10 # 记录训练轮数

# 添加tensorboard
writer = SummaryWriter("logs")

for i in range(epoch):
    print("~~~~~~~~~~~~~~第{}轮训练开始~~~~~~~~~~~~~~~".format(i+1))
    # 训练开始
    net.train()

    for data in train_dataloader:
        imgs, targets = data
        imgs = imgs.to(device)
        targets = targets.to(device)

        outputs = net(imgs)
        loss = loss_fn(outputs, targets)
        # 优化器优化模型
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        total_train_step += 1
        if total_train_step % 100 == 0: 
            print("训练次数：{}, loss:{:.3f}".format(total_train_step, loss.item()))
            writer.add_scalar("train_loss", loss.item(), total_train_step)

    # 每次训练完一轮，测试
    net.eval()
    total_test_loss = 0
    total_accuracy = 0
    with torch.no_grad():
        for data in test_dataloader:
            imgs, targets = data
            imgs = imgs.to(device)
            targets = targets.to(device)

            outputs = net(imgs)
            loss = loss_fn(outputs, targets)
            total_test_loss += loss
            accuracy = (outputs.argmax(1) == targets).sum()
            total_accuracy += accuracy
    print("测试集上的loss：{}".format(total_test_loss.item()))
    writer.add_scalar("test_loss", total_test_loss.item(), total_test_step)
    print("整体测试的正确率：{}".format(total_accuracy / test_dataset_size))
    writer.add_scalar("test_accuracy", total_accuracy / test_dataset_size, total_test_step)
    total_test_step += 1
writer.close()

client获取TensorBoard数据

• 按照Web应用程序，写了个简单client，获取scalar和image数据
  • add_scalar
  • add_image

class TensorBoardClient:
    def __init__(self, host="127.0.0.1", port=6006):
        self.host = host
        self.port = port
        self.baseURL = "http://%s:%d" % (host, port)
        self.headers = {"Connection": "keep-alive",
                   "User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/94.0.4606.81 Safari/537.36",
                   "Content-Type": "application/json",
                   "Accept": "*/*"
                   }
        self.session = requests.session()

    def plugins_listing(self):
        # 列出所有的插件
        url = "%s/data/plugins_listing" % self.baseURL
        print(url)
        response = self.session.get(url, headers=self.headers)
        print(response.status_code)
        print('plugins_listing ', response.text)

    def environment(self):
        # 运行环境
        url = "%s/data/environment" % self.baseURL
        response = self.session.get(url, headers=self.headers)
        print(response.status_code)
        print('environment ', response.text)

    def experiments(self):
        # experiments
        url = "%s/data/experiments" % self.baseURL
        response = self.session.get(url, headers=self.headers)
        print(response.status_code)
        print('experiments ', response.text)
        return response.json()

    def runs(self):
        # runs
        url = "%s/data/runs" % self.baseURL
        response = self.session.get(url, headers=self.headers)
        print(response.status_code)
        print('runs ', response.text)
        runs = response.json()[0]
        return runs

    def getScalar(self):
        # tags
        runs = self.runs()
        print('runs', runs)
        experiment = self.experiments()
        print('experiment', experiment)
        # 获取scalar的所有的tag
        url = "%s/data/plugin/scalars/tags" % self.baseURL
        response = self.session.get(url, headers=self.headers)
        print(response.status_code)
        print('tags ', response.text)
        tags = response.json()[runs]         # 似乎是runs，返回值
        for k, v in tags.items():
            # 针对每个tag单独显示内容
            print(k, v)
            # scalar?tag
            url = "%s/data/plugin/scalars/scalars?tag=%s&run=%s&experiment=" % (self.baseURL, k, runs)
            response = self.session.get(url, headers=self.headers)
            print(response.status_code)
            print(response.text)

    def _saveImage(self, filePath, response):
        size = 0
        with open(filePath, "wb") as f:  # 文件保存
            for chunk in response.iter_content(chunk_size=1024):
                size += f.write(chunk)

    def _byte2Image(self, byteData):
        img = cv2.imdecode(np.frombuffer(byteData, np.uint8), cv2.IMREAD_COLOR)
        return img

    def showImage(self, wnd_name, img):
        cv2.namedWindow(wnd_name, cv2.WINDOW_NORMAL)
        cv2.imshow(wnd_name, img)
        cv2.waitKey(0)

    def getImage(self):
        runs = self.runs()
        url = "%s/data/plugin/images/tags" % self.baseURL
        response = self.session.get(url, headers=self.headers)
        print(response.status_code)
        print(response.text)
        items = response.json()[runs]
        for k, v in items.items():
            # print(k, v)
            url = "%s/data/plugin/images/images?run=%s&sample=%d&tag=%s" % (self.baseURL, runs, 0, k)
            response = self.session.get(url, headers=self.headers) # 查询图片
            l_img = response.json()
            for img in l_img:
                query = img['query']
                url = "%s/data/plugin/images/individualImage?ts=%s&%s" % (self.baseURL, 1 ,query)
                response = self.session.get(url, headers=self.headers) # 查询图片
                # filePath = r'%s.jpg' % k
                # self._saveImage(filePath, response)
                # print(type(response.content))
                img = self._byte2Image(response.content)
                self.showImage(k, img)

QA

1. localhost 无法显示

tensorboard --logdir=‘log’ --host=127.0.0.1

2. 默认端口6006，端口被占用

tensorboard --logdir=‘log’ --port=6602

3. 路径不能有中文
4. 多个文件，读取最近一次
5. 有界面，无法显示数据，返回`return self._accumulators[run],KeyError: None

更换浏览器为Chrome浏览器，360浏览器，搜狗都不行

参考

使用tensorboard可视化模型