使用TensorBoard可视化pytorch模型、数据以及模型训练过程

傷城~ 2023-01-16 03:24 265阅读 0赞

tensorboard常用函数列表如下：

<table style="margin-left:21pt;"> 
 <tbody> 
 <tr> 
 <td style="vertical-align:top;"> <code>torch.utils.</code><code>tensorboard.writer.SummaryWriter</code> </td> 
 <td style="vertical-align:top;width:488px;"> tensorboard的入口，创建event files </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;"> <code>add_scalar</code>(tag,&nbsp;scalar_value,&nbsp;global_step=None,&nbsp;walltime=None) </td> 
 <td style="vertical-align:top;width:488px;"> 添加标量到event file </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;"> <code>add_scalars</code>(main_tag,tag_scalar_dict,&nbsp;global_step=None,&nbsp;walltime=None) </td> 
 <td style="vertical-align:top;width:488px;"> 添加一组标量到event file </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;"> <code>add_histogram</code>(tag,&nbsp;values,&nbsp;global_step=None,&nbsp;bins='tensorflow',&nbsp;walltime=None,&nbsp;max_bins=None) </td> 
 <td style="vertical-align:top;width:488px;"> 添加柱状图 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;"> <code>add_image</code>(tag,&nbsp;img_tensor,&nbsp;global_step=None,&nbsp;walltime=None,&nbsp;dataformats='CHW') </td> 
 <td style="vertical-align:top;width:488px;"> 添加图片 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;"> <code>add_images</code>(tag,&nbsp;img_tensor,&nbsp;global_step=None,&nbsp;walltime=None,&nbsp;dataformats='NCHW') </td> 
 <td style="vertical-align:top;width:488px;"> 添加图片 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;"> <code>add_figure</code>(tag,figure,&nbsp;global_step=None,&nbsp;close=True,&nbsp;walltime=None) </td> 
 <td style="vertical-align:top;width:488px;"> 将图片经过matplotlib处理后添加到TensorBoard </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;"> <code>add_video</code>(tag,&nbsp;vid_tensor,&nbsp;global_step=None,&nbsp;fps=4,&nbsp;walltime=None) </td> 
 <td style="vertical-align:top;width:488px;"> 添加视频 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;"> <code>add_audio</code>(tag,&nbsp;snd_tensor,&nbsp;global_step=None,&nbsp;sample_rate=44100,&nbsp;walltime=None) </td> 
 <td style="vertical-align:top;width:488px;"> 添加音频 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;"> <code>add_text</code>(tag,&nbsp;text_string,&nbsp;global_step=None,&nbsp;walltime=None) </td> 
 <td style="vertical-align:top;width:488px;"> 添加文本 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;"> <code>add_graph</code>(model,&nbsp;input_to_model=None,&nbsp;verbose=False) </td> 
 <td style="vertical-align:top;width:488px;"> 添加网络结构图 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;"> <code>add_embedding</code>(mat,&nbsp;metadata=None,&nbsp;label_img=None,&nbsp;global_step=None,&nbsp;tag='default',&nbsp;metadata_header=None) </td> 
 <td style="vertical-align:top;width:488px;"> 将高维数据映射到低维进行表示 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;"> <code>add_pr_curve</code>(tag,&nbsp;labels,&nbsp;predictions,&nbsp;global_step=None,&nbsp;num_thresholds=127,&nbsp;weights=None,&nbsp;walltime=None) </td> 
 <td style="vertical-align:top;width:488px;"> 添加PR曲线 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;"> <code>add_custom_scalars</code>(layout) </td> 
 <td style="vertical-align:top;width:488px;"> &nbsp; </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;"> <code>add_mesh</code>(tag,&nbsp;vertices,&nbsp;colors=None,&nbsp;faces=None,&nbsp;config_dict=None,&nbsp;global_step=None,&nbsp;walltime=None) </td> 
 <td style="vertical-align:top;width:488px;"> &nbsp; </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;"> <code>add_hparams</code>(hparam_dict=None,&nbsp;metric_dict=None) </td> 
 <td style="vertical-align:top;width:488px;"> 超参数 </td> 
 </tr> 
 <tr> 
 <td style="vertical-align:top;"> <code>close</code>() </td> 
 <td style="vertical-align:top;width:488px;"> &nbsp; </td> 
 </tr> 
 </tbody> 
</table>

下面通过具体代码实例演示pytorch模型中使用tensorboard模块进行模型、数据、训练指标的可视化：

1、加载数据集

%matplotlib inline
    import numpy as np
    import torch
    import torch.nn as nn
    import torchvision.transforms as transforms
    import torch.optim as optim
    import torch.nn.functional as F
    from matplotlib import pyplot as plt
    import torchvision
    import os
    import sys
    
    # 设置下载的pretrained模型保存位置
    os.environ["TORCH_HOME"] = "./models"
    
    # transforms，预处理
    transform = transforms.Compose(
        [
            transforms.ToTensor(),
            transforms.Normalize((0.5, ), (0.5, ))
        ]
    )
    
    # datasets
    trainset = torchvision.datasets.FashionMNIST("./data", download=True, train=True, transform=transform)
    testset = torchvision.datasets.FashionMNIST("./data", download=True, train=False, transform=transform)
    
    # dataloaders
    trainloader = torch.utils.data.DataLoader(trainset, batch_size=4, shuffle=True, num_workers=2, drop_last=True)
    testloader = torch.utils.data.DataLoader(testset, batch_size=4, shuffle=False, num_workers=2, drop_last=False)
    
    # constant for classes
    classes = ('T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat', 'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle Boot')
    
    # helper function to show an image
    # (used in the `plot_classes_preds` function below)
    def matplotlib_imshow(img, one_channel=False):
        if one_channel:
            img = img.mean(dim=0)
        img = img / 2 + 0.5     # unnormalize
        npimg = img.numpy()
        if one_channel:
            plt.imshow(npimg, cmap="Greys")
        else:
            plt.imshow(np.transpose(npimg, (1, 2, 0)))
        plt.show()
    
    for batch_images, batch_labels in trainloader:
        print(batch_images.shape)
        print(batch_labels.shape)
        for img in batch_images:
            print(img.shape)
            matplotlib_imshow(img, one_channel=True)
        break

2、定义网络模型

class Net(nn.Module):
        def __init__(self):
            super(Net, self).__init__()
            self.conv1 = nn.Conv2d(1, 6, 5)
            self.pool = nn.MaxPool2d(2, 2)
            self.conv2 = nn.Conv2d(6, 16, 5)
            self.fc1 = nn.Linear(16 * 4 * 4, 120)
            self.fc2 = nn.Linear(120, 84)
            self.fc3 = nn.Linear(84, 10)
    
        def forward(self, x):
            x = self.pool(F.relu(self.conv1(x)))
            x = self.pool(F.relu(self.conv2(x)))
            x = x.view(-1, 16 * 4 * 4)
            x = F.relu(self.fc1(x))
            x = F.relu(self.fc2(x))
            x = self.fc3(x)
            return x
    
    
    net = Net()
    net

3、定义损失函数和优化方法

criterion = nn.CrossEntropyLoss()
    optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)

4、tensorboard初始化

启动tensorboard命令: tensorboard --logdir=logs

访问tensorboard页面：[http://localhost:6006][http_localhost_6006]

from torch.utils.tensorboard import SummaryWriter
    
    # 创建TensorBoard写入对象
    writer = SummaryWriter("./runs/fashion_mnist_experiment_1")
    
    # 将训练数据图像写入TensorBoard
    dataiter = iter(trainloader)
    images, labels = dataiter.next()
    
    # 创建图像网格区域
    img_grid = torchvision.utils.make_grid(images)
    print(type(img_grid))
    
    matplotlib_imshow(img_grid, one_channel=True)

5、add\_image添加图片到tensorboard

writer.add_image("fasion mnist images", img_grid)

IMAGES标签下可以看到如下图像：

![20210423201443936.png][]

6、add\_graph可视化网络模型结构

writer.add_graph(net, images)

GRAPHS标签下可以看到如下网络结构：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2N4eDY1NA_size_16_color_FFFFFF_t_70][]

7、add\_scalar添加标量，标量值随着global\_step变化，就构成了变化趋势曲线

for i in range(10):
        writer.add_scalar("y = x^2", i ** 2, i)

SCALARS标签下可以看到如下变化曲线：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2N4eDY1NA_size_16_color_FFFFFF_t_70 1][]

8、add\_embedding将高维数据映射到低维进行表示

def select_n_random(data, labels, n=100):
        assert len(data) == len(labels)
        permutation_index = torch.randperm(len(data))
        return data[permutation_index][:n], labels[permutation_index][:n]
    
    images, labels = select_n_random(trainset.data, trainset.targets)
    class_labels = [classes[l] for l in labels]
    
    print(images.shape, labels.shape)
    print(images.unsqueeze(1).shape)
    print(class_labels[:10])
    
    features = images.view(-1, 28 * 28)
    print(features.shape)
    
    # features的形状是 (N, D）
    # metadata对应类别的字符串
    # label_img的形状是(N, C, H, W),也就是一个batch的图像数据
    # D = C * H * W
    writer.add_embedding(features, metadata=class_labels, label_img=images.unsqueeze(1))

PROJECTORS标签下可以看到如下三维图像：

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2N4eDY1NA_size_16_color_FFFFFF_t_70 2][]

9、add\_figure、add\_scalar、add\_pr\_curve记录模型训练准确率、损失函数、PR曲线、ROC曲线等

from sklearn.metrics import roc_curve
    from sklearn.metrics import auc
    
    # helper functions
    
    # 计算输入图像的预测类别和对应的预测概率
    def images_to_probs(net, images):
        '''
        Generates predictions and corresponding probabilities from a trained
        network and a list of images
        '''
        output = net(images)
        # convert output probabilities to predicted class
        _, preds_tensor = torch.max(output, 1)
        preds = np.squeeze(preds_tensor.numpy())
        return preds, [F.softmax(el, dim=0)[i].item() for i, el in zip(preds, output)]
    
    # 使用matplotlib的subplots将一组images进行加载显示
    def plot_classes_preds(net, images, labels):
        '''
        Generates matplotlib Figure using a trained network, along with images
        and labels from a batch, that shows the network's top prediction along
        with its probability, alongside the actual label, coloring this
        information based on whether the prediction was correct or not.
        Uses the "images_to_probs" function.
        '''
        preds, probs = images_to_probs(net, images)
        # plot the images in the batch, along with predicted and true labels
        fig = plt.figure(figsize=(12, 4))
        for idx in np.arange(4):
            ax = fig.add_subplot(1, 4, idx+1, xticks=[], yticks=[])
            matplotlib_imshow(images[idx], one_channel=True)
            ax.set_title("{0}, {1:.1f}%\n(label: {2})".format(
                classes[preds[idx]],
                probs[idx] * 100.0,
                classes[labels[idx]]),
                        color=("green" if preds[idx]==labels[idx].item() else "red"))
        return fig
    
    # 计算准确率
    def accuracy(net, images, labels):
        preds, probs = images_to_probs(net, images)
    #     print(preds, labels)
    #     print(preds == labels.numpy(), (preds == labels.numpy()).sum())
        
        acc = (preds == labels.numpy()).sum() / len(labels)
        return acc
    
    # 计算在数据集上的loss、accuracy以及预测概率
    def evaluate(net, dataloader):
        eval_loss = 0.0
        eval_acc = 0.0
        eval_probs = []
        eval_preds = []
        
        for i, data in enumerate(dataloader, 0):
            inputs, labels = data
            outputs = net(inputs)
            loss = criterion(outputs, labels).item()
            acc = accuracy(net, inputs, labels).item()
            eval_loss += loss
            eval_acc += acc
            preds, probs = images_to_probs(net, inputs)
            eval_probs.append(probs)
            eval_preds.append(preds)
    #         print(i, loss, acc)
            
        return eval_loss / len(dataloader), eval_acc / len(dataloader), np.concatenate(eval_probs, axis=0), np.concatenate(eval_preds, axis=0)
    
    # 绘制各个类别的ROC曲线
    def plot_roc_curve(probs, labels, classes, figsize=(40, 4)):
        fig, axes = plt.subplots(1, len(classes), figsize=figsize)
        
        labels = labels.numpy()
        for cls_index, cls_name in enumerate(classes, 0):
            # 如果分类结果中只有一类是不能绘制ROC曲线和计算AUC的
            cls_labels = labels.copy()
            # 将训练标签中当前类型的label置为1（正类），其余类型的label置为0（负类）
            cls_labels[labels != cls_index] = 0
            cls_labels[labels == cls_index] = 1
    #         print(cls_labels[:50])
            
            # 在预测概率中，保留当前类别的预测概率（正类），负类的预测概率为1 - 当前类别的概率
            cls_probs = probs.copy()
            cls_probs[labels != cls_index] = 1 - cls_probs[labels != cls_index]
            cls_probs = cls_probs.copy()
    #         print(cls_probs[:50])
            
            fpr, tpr, thr = roc_curve(cls_labels, cls_probs)
    #         print(fpr[:50])
    #         print(tpr[:50])
            axes[cls_index].plot(fpr, tpr, label="AUC: {:.2f}".format(auc(fpr, tpr)))
            axes[cls_index].set_xlabel("FPR")
            axes[cls_index].set_ylabel("TPR")
            axes[cls_index].set_title(classes[cls_index])
            axes[cls_index].legend(loc="lower right")
    #     plt.show()
        return fig
    
    # 绘制各个类别的PR曲线
    def plot_pr_curve(writer, probs, labels, classes, step):
        labels = labels.numpy()
        for cls_index, cls_name in enumerate(classes, 0):
            # 如果分类结果中只有一类是不能绘制ROC曲线和计算AUC的
            cls_labels = labels.copy()
            # 将训练标签中当前类型的label置为1（正类），其余类型的label置为0（负类）
            cls_labels[labels != cls_index] = 0
            cls_labels[labels == cls_index] = 1
    #         print(cls_labels[:50])
            
            # 在预测概率中，保留当前类别的预测概率（正类），负类的预测概率为1 - 当前类别的概率
            cls_probs = probs.copy()
            cls_probs[labels != cls_index] = 1 - cls_probs[labels != cls_index]
            cls_probs = cls_probs.copy()
    #         print(cls_probs[:50])
            writer.add_pr_curve("PR/{}".format(cls_name), cls_labels, cls_probs, global_step=step)
        return writer
    
    # 计算各个类别的准确率
    def plot_accuracy(writer, prebs, labels, classes):
        labels = labels.numpy()
        acc_dict = {}
        for cls_index, cls_name in enumerate(classes, 0):
            cls_labels = labels[labels == cls_index]
            cls_prebs = prebs[labels == cls_index]
    
            acc = (cls_labels == cls_prebs).sum() / cls_labels.shape[0]
            acc_dict[cls_name] = acc
            
        for cls_name, cls_acc in acc_dict.items():
            writer.add_scalar("Accuracy/{}".format(cls_name), cls_acc, epoch * len(trainloader) + i)
        return writer

定义模型训练过程：

# 记录训练误差和训练准确率
    running_loss = 0.0
    running_acc = 0.0
    for epoch in range(3):  # loop over the dataset multiple times
    
        for i, data in enumerate(trainloader, 0):
            # get the inputs; data is a list of [inputs, labels]
            inputs, labels = data
    
            # zero the parameter gradients
            optimizer.zero_grad()
    
            # forward + backward + optimize
            outputs = net(inputs)
            loss = criterion(outputs, labels)
            loss.backward()
            optimizer.step()
    
            running_loss += loss.item()
            running_acc += accuracy(net, inputs, labels).item()
            if i % 1000 == 999:    # every 1000 mini-batches...
                net.eval()
                # 记录训练集的整体损失和准确率
                # ...log the running loss
                writer.add_scalar('Loss/training loss', running_loss / 1000, epoch * len(trainloader) + i)
                writer.add_scalar("Accuracy/training accuracy", running_acc / 1000, epoch * len(trainloader) + i)
                
                # 记录测试集的整体损失和准确率
                eval_loss, eval_acc, eval_probs, eval_prebs = evaluate(net, testloader)
                writer.add_scalar("Loss/testing loss", eval_loss, epoch * len(trainloader) + i)
                writer.add_scalar("Accuracy/testing accuracy", eval_acc, epoch * len(trainloader) + i)
                
                # 记录测试集各个类别的准确率
    #             for cls_name, cls_acc in plot_accuracy(eval_prebs, testset.targets, classes).items():
    #                 writer.add_scalar("Accuracy/{}".format(cls_name), cls_acc, epoch * len(trainloader) + i)
                writer = plot_accuracy(writer, eval_prebs, testset.targets, classes)
                
                # 记录测试集各个类别的P-R曲线
                writer = plot_pr_curve(writer, eval_probs, testset.targets, classes, 
                                       step=epoch * len(trainloader) + i)
                
                # 记录测试集各个类别的ROC曲线
                writer.add_figure("testing roc curve", 
                                  plot_roc_curve(eval_probs, testset.targets, classes), 
                                  global_step=epoch * len(trainloader) + i)
    
                # 输出样本的真实标签和预测标签、预测概率
                # ...log a Matplotlib Figure showing the model's predictions on a
                # random mini-batch
                writer.add_figure('predictions vs. actuals',
                                plot_classes_preds(net, inputs, labels),
                                global_step=epoch * len(trainloader) + i)
                
                print("epoch: {}, iter: {}, train loss: {}, train acc: {}, eval loss: {}, eval acc: {}".format(epoch, i, 
                                                                                                               running_loss / 1000, 
                                                                                                               running_acc / 1000, 
                                                                                                               eval_loss, 
                                                                                                               eval_acc))
                running_loss = 0.0
                running_acc = 0.0
                net.train()
    writer.close()
    print('Finished Training')

tensorboard页面可以看到训练的损失函数、准确率变化曲线，模型预测的PR曲线和ROC曲线如下：

1）、损失函数曲线

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2N4eDY1NA_size_16_color_FFFFFF_t_70 3][]

2）、准确率曲线

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2N4eDY1NA_size_16_color_FFFFFF_t_70 4][]

3）、各个类别的准确率曲线

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2N4eDY1NA_size_16_color_FFFFFF_t_70 5][]

4）、各个类别的PR曲线

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2N4eDY1NA_size_16_color_FFFFFF_t_70 6][]

5）、各个类别的ROC曲线

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2N4eDY1NA_size_16_color_FFFFFF_t_70 7][]

参考链接

[VISUALIZING MODELS, DATA, AND TRAINING WITH TENSORBOARD][VISUALIZING MODELS_ DATA_ AND TRAINING WITH TENSORBOARD]

[TORCH.UTILS.TENSORBOARD][]

[使用scikit-learn计算分类器的ROC曲线及AUC值][scikit-learn_ROC_AUC]

[http_localhost_6006]: http://localhost:6006
[20210423201443936.png]: /images/20221022/c23da91ea49646968fa18ac65d94369c.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2N4eDY1NA_size_16_color_FFFFFF_t_70]: /images/20221022/caecb071715041478d2028a52a560255.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2N4eDY1NA_size_16_color_FFFFFF_t_70 1]: /images/20221022/90cfd968299e4ea9869c5b5871c38a1d.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2N4eDY1NA_size_16_color_FFFFFF_t_70 2]: /images/20221022/0279f7dbe3504c3c92a0d1a954fc1fe1.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2N4eDY1NA_size_16_color_FFFFFF_t_70 3]: /images/20221022/32389267e3364a8f8edeed354d2dba5a.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2N4eDY1NA_size_16_color_FFFFFF_t_70 4]: /images/20221022/d6c4d3033ac143cc89977b52d3828208.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2N4eDY1NA_size_16_color_FFFFFF_t_70 5]: /images/20221022/a8e1a394508944b7b9898c3d6badcded.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2N4eDY1NA_size_16_color_FFFFFF_t_70 6]: /images/20221022/4ee76b6fe4e74cde84c618eb5ee60d27.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L2N4eDY1NA_size_16_color_FFFFFF_t_70 7]: /images/20221022/a0f49474e0c84cd2acb5461c1e1f3b84.png
[VISUALIZING MODELS_ DATA_ AND TRAINING WITH TENSORBOARD]: https://pytorch.org/tutorials/intermediate/tensorboard_tutorial.html
[TORCH.UTILS.TENSORBOARD]: https://pytorch.org/docs/1.6.0/tensorboard.html
[scikit-learn_ROC_AUC]: https://blog.csdn.net/cxx654/article/details/106737317