1. Configuration environment

Use the pip install statement to install Tensorboard in the console (tensorflow needs to be installed at the same time)


pip install tensorflow
pip install tensorboard

Add a reference and set a path for it

from torch.utils.tensorboard import SummaryWriter

writer = SummaryWriter(log_dir = '日志路径')

2. About SummaryWiter

SummaryWriter is the core of Tensorboard. After the declaration, Tensorboard will start the service, and then you can access the data visually in the browser; you also need to use the close statement to close the service after use. Similar to flask's app.run

writer.close()

3. Add data

Generally speaking, commonly used functions are: add_scalar (add scalar), add_scalars, add_image (add image), and its usage is as follows:

1.add_scalar

Used to add scalars, such as loss functions, that can be used to draw graphs.

for i in range(100):
    writer.add_scalar(tag="accuracy", # 标签，即数据属于哪一类，不同的标签对应不同的图
                      scalar_value=i * random.uniform(0.8, 1),  # 纵坐标的值
                      global_step=i  # 迭代次数，即图的横坐标
                      )
    #延迟代码，与数据无关
    time.sleep(2 * random.uniform(0.5, 1.5))

Tips.Tag can be split using / (such as: record/avg_loss; record/total_loss), which will put multiple images under the same Tag

2.add_scalars

Add multiple graph lines to the same graph

for epoch in range(100):
    writer.add_scalars('scalar/scalars_test', {'xsinx': epoch * np.sin(epoch), 'xcosx': epoch * np.cos(epoch)}, epoch)

3.add_graph

Import the model to realize model visualization

model = Net1()    #实例化模型
with SummaryWriter(comment='Net1') as w:
    w.add_graph(model, (dummy_input,))

The second parameter is the input vector , which can also be initialized by the following method

init_img = torch.zeros((1, 3, 224, 224), device=device)

4.add_image

To add an image file, you need to specify the channel format (dataformats)

single write

writer.add_image(tag = "test", img_array,1, dataformats='HWC')

multiple write

The pictures need to be packaged into a batch first , and then passed in uniformly

for i in range(16):
    img_batch[i, 0] = np.arange(0, 10000).reshape(100, 100) / 10000 / 16 * i
    img_batch[i, 1] = (1 - np.arange(0, 10000).reshape(100, 100) / 10000) / 16 * i

writer.add_images('my_image_batch', img_batch, 0)

[PyTorch] Using Tensorboard in the PyTorch environment