CSV Table Logger

If you wish to record some configuration information and metrics locally in a CSV file during training (in a format consistent with the "Table View" on the SwanLab webpage), we highly recommend using the CSV Logger plugin.

Improving the Plugin

All SwanLab plugins are open-source. You can view the Github source code. We welcome your suggestions and PRs!

Plugin Usage

1. Initialize the CSV Logger:

from swanlab.plugin.writer import CSVWriter

csv_writer = CSVWriter(dir="logs")

The dir parameter specifies the save path for the CSV file. By default, it is saved in the current working directory.

2. Pass the Plugin:

swanlab.init(
    ...
    callbacks=[csv_writer]
)

After executing the code, a swanlab_run.csv file will be generated in the logs directory, and data recording will begin. For each subsequent training session, a new row will be added to this CSV file.

If you want to specify a different file name, you can pass the filename parameter:

csv_writer = CSVWriter(dir="logs", filename="my_csv_file.csv")

Example Code

import swanlab
from swanlab.plugin.writer import CSVWriter
import random

csv_writer = CSVWriter(dir="logs")

# Create a SwanLab project
swanlab.init(
    # Set the project name
    project="my-awesome-project",
    
    # Set hyperparameters
    config={
        "learning_rate": 0.02,
        "architecture": "CNN",
        "dataset": "CIFAR-100",
        "epochs": 10,
        "batch_size": 128
    },
    callbacks=[csv_writer]
)

# Simulate a training session
epochs = 10
offset = random.random() / 5
for epoch in range(2, epochs):
  acc = 1 - 2 ** -epoch - random.random() / epoch - offset
  loss = 2 ** -epoch + random.random() / epoch + offset

  # Log training metrics
  swanlab.log({"acc": acc, "loss2": loss})

# [Optional] Finish training, which is necessary in notebook environments
swanlab.finish()

Register plugins externally

If you are using the integration of SwanLab with other frameworks and thus find it difficult to locate swanlab.init, you can use the swanlab.register_callbacks method to pass in plugins externally:

import swanlab

# Equivalent to swanlab.init(callbacks=[...])
swanlab.register_callbacks([...])