随笔- 8 文章- 0 评论- 5 阅读- 7661

关于tensorboard笔记

------------恢复内容开始------------

关于tensorboard:

它的使用是在训练模型之后的分析模型结果阶段。作用是将cnn度量可视化，使我们可以更清楚地了解这个训练过程中发生了什么事情。具体而言：

Tracking and visualizing metrics such as loss and accuracy
Visualizing the model graph (ops and layers)
Viewing histograms of weights, biases, or other tensors as they change over time
Projecting embeddings to a lower dimensional space
Displaying images, text, and audio data
Profiling TensorFlow programs
And much more

关于使用：

pip install tensorboard（这个版本要在1.15及以上）
Pytorch的1.1.0版本就加入了tensorboard的实用程序包：
from torch.utils.tensorboard import SummaryWriter(SummaryWriter是一个类）

To use TensorBoard our task is to get the data we want displayed saved to a file that TensorBoard can read.

创建一个SummaryWriter实例、PyTorch网络实例并对一组的images和labels进行分析。之后图像和网络图都被加进了Tensorboard.

tb  = SummaryWriter()
 
network  = Network()
 
images, labels  = next ( iter (train_loader))
 
grid  = torchvision.utils.make_grid(images)
 
tb.add_image( 'images' , grid)
 
tb.add_graph(network, images)
 
tb.close()

Running TensorBoard:（运行.py文件，会在.py文件所在的目录下生成一个runs文件夹）
默认SummaryWriter实体会把数据写入磁盘中的./runs目录下，运行命令时通过传参告诉数据位置：

tensorboard --logdir=runs

访问TensorBoard接口：

http://localhost:6006

（这时可以看到的式网络图和图片数据）

在终端cd 到events.out文件所在的上上级目录，然后tensorboard --logdir=上级目录

然后通过谷歌点开访问接口

http://localhost:6006/

TensorBoard Histograms And Scalars:

我们可以添加到TensorBoard的下一个导入类型是数值数据。我们可以添加将随时间或epoch显示的标量值。我们还可以在直方图中添加值，以查看值的频率分布。

例子：
tb.add_scalar('Loss', total_loss, epoch)

tb.add_scalar('Number Correct', total_correct, epoch)

tb.add_scalar('Accuracy', total_correct / len(train_set), epoch)

tb.add_histogram('conv1.bias', network.conv1.bias, epoch)

tb.add_histogram('conv1.weight', network.conv1.weight, epoch)

tb.add_histogram('conv1.weight.grad', network.conv1.weight.grad, epoch)

在训练循环体中放置上面的calls:

 1 network = Network()
 2 
 3 train_loader = torch.utils.data.DataLoader(train_set, batch_size=100)
 4 
 5 optimizer = optim.Adam(network.parameters(), lr=0.01)
 6 
 7  
 8 
 9 images, labels = next(iter(train_loader))
10 
11 grid = torchvision.utils.make_grid(images)
12 
13  
14 
15 tb = SummaryWriter()
16 
17 tb.add_image('images', grid)
18 
19 tb.add_graph(network, images)
20 
21 for epoch in range(1):
22 
23  
24 
25     total_loss = 0
26 
27     total_correct = 0
28 
29  
30 
31     for batch in train_loader: # Get Batch
32 
33  
34 
35         # Pass Batch
36 
37         # Calculate Loss
38 
39         # Calculate Gradient
40 
41         # Update Weights
42 
43  
44 
45     tb.add_scalar('Loss', total_loss, epoch)
46 
47     tb.add_scalar('Number Correct', total_correct, epoch)
48 
49     tb.add_scalar('Accuracy', total_correct / len(train_set), epoch)
50 
51  
52 
53     tb.add_histogram('conv1.bias', network.conv1.bias, epoch)
54 
55     tb.add_histogram('conv1.weight', network.conv1.weight, epoch)
56 
57     tb.add_histogram(
58 
59         'conv1.weight.grad'
60 
61         ,network.conv1.weight.grad
62 
63         ,epoch
64 
65     )
66 
67  
68 
69     print(
70 
71         "epoch", epoch,
72 
73         "total_correct:", total_correct,
74 
75         "loss:", total_loss
76 
77     )
78 
79  
80 
81 tb.close()

加入到Tensorboard中的这些值还会随着玩过训练实时更新。

TensorBoard的真正强大之处在于它能够对多次运行进行比较。这允许我们通过改变超参数值和比较运行来快速试验，以查看哪些参数工作得最好。

另一个例子：(来自参考https://blog.csdn.net/wuzhihuaw/article/details/121357355）

import numpy as np
from torch.utils.tensorboard  import SummaryWriter
 
writer  = SummaryWriter(comment = 'tensorboard_test' )
 
for x  in range ( 50 ):
    writer.add_scalar( 'y=2x' , x  * 2 , x)
    writer.add_scalar( 'y=pow(2, x)' ,  2 * * x, x)
 
    writer.add_scalars( 'data/scalar_group' , { "xsinx" : x  * np.sin(x),
                                             "xcosx" : x  * np.cos(x),
                                             "arctanx" : np.arctan(x)}, x)
writer.close()