PyTorch之—可视化（Visdom/TensorboardX）

一、Visdom

Visdom是Facebook在2017年发布的一款针对PyTorch的可视化工具。visdom由于其功能简单，一般会被定义为服务器端的matplot，也就是说我们可以直接使用python的控制台模式进行开发并在服务器上执行，将一些可视化的数据传送到Visdom服务上，通过Visdom服务进行可视化。
官网：https://github.com/facebookresearch/visdom

安装与使用

执行命令：pip install visdom安装即可。
使用命令：python -m visdom.server 在本地启动服务器，启动后会提示It’s Alive! You can navigate to http://localhost:8097 这就说明服务已经可用。
输入地址：http://localhost:8097 即可看到页面。

此时终端：

(AI_PyTorch_GPU) E:\pytorch_AI>python -m visdom.server
Checking for scripts.
It's Alive!
INFO:root:Application Started
You can navigate to http://localhost:8097
INFO:tornado.access:200 GET / (::1) 14.96ms
INFO:tornado.access:200 GET /static/css/react-resizable-styles.css?v=9f91a8dbf4d8f7ef1399e625660405f4 (::1) 1.00ms
INFO:tornado.access:200 GET /static/css/react-grid-layout-styles.css?v=7dc8934d2f9ac5303b8f0bb1148152a0 (::1) 2.00ms
INFO:tornado.access:200 GET /static/js/bootstrap.min.js?v=5869c96cc8f19086aee625d670d741f9 (::1) 1.00ms
INFO:tornado.access:200 GET /static/js/react-react.min.js?v=bca103da5b5404d93783ccf73e0e9d1e (::1) 0.00ms
INFO:tornado.access:200 GET /static/fonts/layout_bin_packer?v=6c46683ed70fbb1443caf3531243836d (::1) 1.00ms
INFO:tornado.access:200 GET /static/css/bootstrap.min.css?v=ec3bb52a00e176a7181d454dffaea219 (::1) 188.49ms
INFO:tornado.access:200 GET /static/js/react-dom.min.js?v=950495cc51ccb90612cf0fe0bb44f8f3 (::1) 11.97ms
INFO:tornado.access:200 GET /static/js/jquery.min.js?v=e071abda8fe61194711cfc2ab99fe104 (::1) 5.98ms
INFO:tornado.access:200 GET /static/css/style.css?v=f1b2f4d2442b8263f30058d8085c1ab3 (::1) 1.02ms
INFO:tornado.access:200 GET /static/js/mathjax-MathJax.js?v=49565b9ce89c64da075a5a39969b366e (::1) 1.00ms
INFO:tornado.access:200 GET /static/js/main.js?v=4316a77ded15b92b6e884b60b72f6513 (::1) 21.94ms
INFO:tornado.access:200 GET /static/js/plotly-plotly.min.js?v=00dcca8ba45aab5f63dc7fc55a544d3d (::1) 81.78ms
INFO:tornado.access:200 GET /static/fonts/glyphicons-halflings-regular.woff2 (::1) 1.00ms
INFO:tornado.access:101 GET /socket (::1) 1.00ms
INFO:root:Opened new socket from ip: ::1
INFO:tornado.access:200 POST /env/main (::1) 0.99ms
INFO:tornado.access:200 POST /env/main (::1) 0.00ms
INFO:tornado.access:200 GET /favicon.png (::1) 2.99ms
INFO:tornado.access:200 GET /extensions/MathMenu.js?V=2.7.1 (::1) 2.03ms
INFO:tornado.access:200 GET /extensions/MathZoom.js?V=2.7.1 (::1) 1.96ms

端口8097是默认的端口可以在启动命令后加 -port参数指定端口，常用的参数还有 --hostname，-base_url等

Visdom的服务在启动时会自动下载一些静态文件，可能会有报错问题。这里直接复制到Lib\site-packages\visdom中即可。
C:\Anaconda3\envs\<"虚拟环境名">\Lib\site-packages\visdom
关于文件详情以及安装 详情请点击

Visdom是由Plotly 提供的可视化支持，所以提供一下可视化的接口

可视化接口:

vis.scatter : 2D 或 3D 散点图
vis.line : 线图
vis.stem : 茎叶图
vis.heatmap : 热力图
vis.bar : 条形图
vis.histogram: 直方图
vis.boxplot : 箱型图
vis.surf : 表面图
vis.contour : 轮廓图
vis.quiver : 绘出二维矢量场
vis.image : 图片
vis.text : 文本
vis.mesh : 网格图
vis.save : 序列化状态

小案例

import math
import numpy as np
from visdom import Visdom
import time


env = Visdom()
assert env.check_connection() #测试一下链接，链接错误的话会报错
Y = np.linspace(0, 2 * math.pi, 70)
X = np.column_stack((np.sin(Y), np.cos(Y)))
env.stem(
        X=X,
        Y=Y,
        opts=dict(legend=['Sine', 'Cosine'])
    )

# 通过env参数指定Environments，如果名称包含了下划线,那么visdom会跟根据下划线分割并自动分组
envtest = Visdom(env='test_mesh')
assert envtest.check_connection()


x = [0, 0, 1, 1, 0, 0, 1, 1]
y = [0, 1, 1, 0, 0, 1, 1, 0]
z = [0, 0, 0, 0, 1, 1, 1, 1]
X = np.c_[x, y, z]
i = [7, 0, 0, 0, 4, 4, 6, 6, 4, 0, 3, 2]
j = [3, 4, 1, 2, 5, 6, 5, 2, 0, 1, 6, 3]
k = [0, 7, 2, 3, 6, 7, 1, 1, 5, 5, 7, 6]
Y = np.c_[i, j, k]
envtest.mesh(X=X, Y=Y, opts=dict(opacity=0.5))

# 更新损失函数
x,y=0,0
env2 = Visdom()
pane1= env2.line(
    X=np.array([x]),
    Y=np.array([y]),
    opts=dict(title='dynamic data'))

for i in range(10):
    time.sleep(1)        # 每隔一秒钟打印一次数据
    x+=i
    y=(y+i)*1.5
    print(x,y)
    env2.line(
        X=np.array([x]),
        Y=np.array([y]),
        win=pane1,       # win参数确认使用哪一个pane
        update='append') # 我们做的动作是追加，除了追加意外还有其他方式

二、TensorBoardX

安装与使用 (依次执行如下命令)

pip install tensorboard    #安装tensorboard
pip install tensorboardx   #安装tensorboardx

tensorboard --logdir runs
http://localhost:6006/    #在chrome浏览器中打开

案例一

项目地址：https://github.com/lanpa/tensorboardX

python examples/demo.py
tensorboard --logdir runs
http://localhost:6006/    #在chrome浏览器中打开

界面说明

• SCALAR
  对标量数据进行汇总和记录，通常用来可视化训练过程中随着迭代次数准确率(val acc)、损失值(train/test loss)、学习率(learning rate)、每一层的权重和偏置的统计量(mean、std、max/min)等的变化曲线

• IMAGES
  可视化当前轮训练使用的训练/测试图片或者 feature maps

• GRAPHS
  可视化计算图的结构及计算图上的信息，通常用来展示网络的结构

• HISTOGRAMS
  可视化张量的取值分布，记录变量的直方图(统计张量随着迭代轮数的变化情况）

• PROJECTOR
  全称Embedding Projector 高维向量进行可视化

• 使用
  在使用前请先去确认执行tensorboard --logdir logs 并保证 http://localhost:6006/ 页面能够正常打开。tensorboard 使用详情请点击

案例二

import torch
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
from PIL import Image
from torch.utils.data import DataLoader
from torchvision import transforms
from torchvision import models,datasets
from tensorboardX import SummaryWriter


cat_img = Image.open('./Felis_silvestris_catus_lying_on_rice_straw.jpg')

transform_224 = transforms.Compose([
        transforms.Resize(224),
        transforms.CenterCrop(224),
        transforms.ToTensor(),
    ])
cat_img_224=transform_224(cat_img)

# 将图片展示在tebsorboard中：
writer = SummaryWriter(log_dir='./logs', comment='cat image') # 这里的logs要与--logdir的参数一样
writer.add_image("cat",cat_img_224)
writer.close()# 执行close立即刷新，否则将每120秒自动刷新


# 更新损失函数
# 更新损失函数和训练批次我们与visdom一样使用模拟展示，这里用到的是tensorboard的SCALAR页面
x = torch.FloatTensor([100])
y = torch.FloatTensor([500])

for epoch in range(100):
    x /= 1.5
    y /= 1.5
    loss = y - x
    with SummaryWriter(log_dir='./logs', comment='train') as writer: #with语法:自动调用close方法
        writer.add_histogram('his/x', x, epoch)
        writer.add_histogram('his/y', y, epoch)
        writer.add_scalar('data/x', x, epoch)
        writer.add_scalar('data/y', y, epoch)
        writer.add_scalar('data/loss', loss, epoch)
        writer.add_scalars('data/data_group', {'x': x,
                                               'y': y,
                                               'loss': loss}, epoch)

查看图像：

• 方法一：
  激活环境 --> 进入logs目录 -->执行命令 --> chrome下打开http://localhost:8080/

  (AI_PyTorch_GPU) E:\pytorch_AI\logs>tensorboard --logdir=E:\\pytorch_AI\\logs --port=8080
  TensorBoard 1.12.2 at http://DESKTOP-JQI7IBN:8080 (Press CTRL+C to quit)
  

  如图：
  

• 方法二：
  激活环境 --> 进入logs父目录 -->执行命令 --> chrome下打开http://localhost:6006/

  (AI_PyTorch_GPU) E:\pytorch_AI>tensorboard --logdir logs
  TensorBoard 1.12.2 at http://DESKTOP-JQI7IBN:6006 (Press CTRL+C to quit)
  

  如图：
  

使用PROJECTOR对高维向量可视化

PROJECTOR的的原理是通过PCA，T-SNE等方法将高维向量投影到三维坐标系（降维度）。Embedding Projector从模型运行过程中保存的checkpoint文件中读取数据，默认使用主成分分析法（PCA）将高维数据投影到3D空间中，也可以通过设置设置选择T-SNE投影方法，这里做一个简单的展示。

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader
from torchvision import transforms
from torchvision import datasets
from tensorboardX import SummaryWriter




""" ==================使用PROJECTOR对高维向量可视化====================

    PROJECTOR的的原理是通过PCA，T-SNE等方法将高维向量投影到三维坐标系（降维度）。
    Embedding Projector从模型运行过程中保存的checkpoint文件中读取数据，
    默认使用主成分分析法（PCA）将高维数据投影到3D空间中，也可以通过设置设置选择T-SNE投影方法，
    这里做一个简单的展示。
"""

BATCH_SIZE=256
EPOCHS=20
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

train_loader = DataLoader(datasets.MNIST('./mnist', train=True, download=False,
                                         transform=transforms.Compose([
                                             transforms.ToTensor(),
                                             transforms.Normalize((0.1307,), (0.3081,))
                                         ])),
                          batch_size=BATCH_SIZE, shuffle=True)


test_loader = torch.utils.data.DataLoader(
        datasets.MNIST('./mnist', train=False, transform=transforms.Compose([
                           transforms.ToTensor(),
                           transforms.Normalize((0.1307,), (0.3081,))
                       ])),
        batch_size=BATCH_SIZE, shuffle=True)

class ConvNet(nn.Module):
    def __init__(self):
        super().__init__()
        # 1,28x28
        self.conv1=nn.Conv2d(1,10,5)  # 10, 24x24
        self.conv2=nn.Conv2d(10,20,3) # 128, 10x10
        self.fc1 = nn.Linear(20*10*10,500)
        self.fc2 = nn.Linear(500,10)
    def forward(self,x):
        in_size = x.size(0)
        out = self.conv1(x)            #24
        out = F.relu(out)
        out = F.max_pool2d(out, 2, 2)  #12
        out = self.conv2(out)          #10
        out = F.relu(out)
        out = out.view(in_size,-1)
        out = self.fc1(out)
        out = F.relu(out)
        out = self.fc2(out)
        out = F.log_softmax(out,dim=1)
        return out
model = ConvNet().to(DEVICE)
optimizer = torch.optim.Adam(model.parameters())


def train(model, DEVICE,train_loader, optimizer, epoch):
    n_iter=0
    model.train()
    for batch_idx, (data, target) in enumerate(train_loader):
        data,target = data.to(DEVICE),target.to(DEVICE)
        optimizer.zero_grad()
        output = model(data)
        loss = F.nll_loss(output, target)
        loss.backward()
        optimizer.step()
        if(batch_idx+1)%30 == 0:
            n_iter=n_iter+1
            print('Train Epoch: {} [{}/{} ({:.0f}%)]\t Loss: {:.6f}'.format(
                epoch, batch_idx * len(data), len(train_loader.dataset),
                100. * batch_idx / len(train_loader), loss.item()))

            #主要增加了一下内容
            out = torch.cat((output.data.cpu(), torch.ones(len(output), 1)), 1) # 因为是投影到3D的空间，所以我们只需要3个维度
            with SummaryWriter(log_dir='./pytorch_tensorboardX_03_logs', comment='mnist') as writer:
                #使用add_embedding方法进行可视化展示
                writer.add_embedding(
                    out,
                    metadata=target.data,
                    label_img=data.data,
                    global_step=n_iter)


def test(model, device, test_loader):
    model.eval()
    test_loss = 0
    correct = 0
    with torch.no_grad():
        for data, target in test_loader:
            data, target = data.to(device), target.to(device)
            output = model(data)
            test_loss += F.nll_loss(output, target, reduction='sum').item() # 损失相加
            pred = output.max(1, keepdim=True)[1]                           # 找到概率最大的下标
            correct += pred.eq(target.view_as(pred)).sum().item()

    test_loss /= len(test_loader.dataset)
    print('\n Test set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'
          .format(test_loss, correct, len(test_loader.dataset),
                  100. * correct / len(test_loader.dataset)))

for epoch in range(1, EPOCHS + 1):
    train(model, DEVICE, train_loader, optimizer, epoch)
    test(model, DEVICE, test_loader)

# 保存模型
torch.save(model.state_dict(), './pytorch_tensorboardX_03.pth')
# Test set: Average loss: 0.0400, Accuracy: 9913/10000 (99%)

如上代码直接点击运行。
可视化展示：

(AI_PyTorch_GPU) E:\pytorch_AI>tensorboard --logdir pytorch_tensorboardX_03_logs
TensorBoard 1.12.2 at http://DESKTOP-JQI7IBN:6006 (Press CTRL+C to quit)

复制http://localhost:6006/到chrome打开：

绘制网络结构

import torch
import numpy as np
from torchvision import models,transforms
from PIL import Image
from tensorboardX import SummaryWriter



vgg16 = models.vgg16() # 这里下载预训练好的模型
print(vgg16)           # 打印一下这个模型


transform_2 = transforms.Compose([
    transforms.Resize(224),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    # convert RGB to BGR
    # from <https://github.com/mrzhu-cool/pix2pix-pytorch/blob/master/util.py>
    transforms.Lambda(lambda x: torch.index_select(x, 0, torch.LongTensor([2, 1, 0]))),
    transforms.Lambda(lambda x: x*255),
    transforms.Normalize(mean = [103.939, 116.779, 123.68],
                          std = [ 1, 1, 1 ]),
])



cat_img = Image.open('./Felis_silvestris_catus_lying_on_rice_straw.jpg')
vgg16_input=transform_2(cat_img)[np.newaxis]# 因为pytorch的是分批次进行的，所以我们这里建立一个批次为1的数据集
print(vgg16_input.shape)


# 开始前向传播，打印输出值
raw_score = vgg16(vgg16_input)
raw_score_numpy = raw_score.data.numpy()
print(raw_score_numpy.shape, np.argmax(raw_score_numpy.ravel()))


# 将结构图在tensorboard进行展示
with SummaryWriter(log_dir='./vgg16_logs', comment='vgg16') as writer:
    writer.add_graph(vgg16, (vgg16_input,))

(AI_PyTorch_GPU) E:\pytorch_AI>tensorboard --logdir vgg16_logs
TensorBoard 1.12.2 at http://DESKTOP-JQI7IBN:6006 (Press CTRL+C to quit)