Tensorboard quick start (pytoch uses tensorboard)

TensorBoard brief introduction

TensorBoard yes Google Developed a machine learning visualization tool . It is mainly used to record the machine learning process , for example ：

• Record Loss change 、 Accuracy changes etc.
• Record The picture changes 、 phonological change 、 Text changes, etc , For example, doing GAN when , You can record a generated picture after a period of time
• Draw a model

Please refer to Official documents

TensorBoard Interface is introduced

TensorBoard install

Use it directly pip Can be installed ：

pip install tensorboard

After installation , Enter at the command line ：

tensorboard --help

If it can output normally , The installation is successful .

TensorBoard function

Just run the start command ：

tensorboard --logdir my_log

my_log yes TensorBoard Of log Directory of files .Tensorboard The data displayed in the panel comes from log file , Generally, a complete run generates one log file .

for example , stay Pytorch in , We will call new One SummaryWriter object , A log file , Then we will call it add_something Method , Go to log Write a log inside , After the TensorBoard You can see the data in the panel . Finally, after the training , call close Method end .

If you see the following output , Indicating successful startup ：

TensorBoard 2.8.0 at http://localhost:6006/ (Press CTRL+C to quit)

At this time, you only need to enter http://localhost:6006/ You can enter TensorBoard Interface .

Pytorch Use TensorBoard

Pytorch Use Tensorboard There are three main uses API：

• SummaryWriter： This is used to create a log file ,TensorBoard When viewing the panel , You also need to select the one to view log file .
• add_something： towards log Add data to the file . For example, through add_scalar Add line chart data ,add_image You can add pictures .
• close： When the training is over , We can go through close Method end log write in .

Next , Let's simulate the change of accuracy during training .

The first thing you need to new One SummaryWriter object ：

from torch.utils.tensorboard import SummaryWriter

writer = SummaryWriter(log_dir='runs/mock_loss')

After running this line of code , You can see that a runs/mock_accuracy Folder , And it has event journal

Now it's ready to start tensorboard Here we are :

tensorboard --logdir runs

Now enter tensorboard After the page , I can't see anything , Because we haven't told log Write any data inside ：

Next use add_scalar Draw a line chart of accuracy ：

for i in range(100):
    writer.add_scalar(tag="accuracy", #  It can be temporarily understood as the name of the image 
                      scalar_value=i * random.uniform(0.8, 1),  #  The value of the ordinate 
                      global_step=i  #  What is the current iteration , It can be understood as the value of abscissa 
                      )
    time.sleep(2 * random.uniform(0.5, 1.5))

Simulation here 1~3 Complete an iteration in seconds and calculate the accuracy , Then add the accuracy to accuracy Below this figure .

After a while , Let's refresh the page , You can see our accuracy curve ：

Because the data is still being written , So the curve is still changing .

Tensor Broken line diagram （Scalars）

• TODO： Introduction of other graphics