Set Experiment Configuration

Use swanlab.config to save your training configurations, such as:

Hyperparameters
Input settings, such as dataset name or model type
Any other variables for your experiment

swanlab.config allows you to easily analyze your experiments and reproduce your work in the future. You can also compare configurations of different experiments in the SwanLab application and see how different training configurations affect model outputs.

Set Experiment Configuration

config is typically defined at the beginning of the training script. Of course, different AI workflows may vary, so config also supports defining at different locations in the script to meet flexible needs.

The following sections outline different scenarios for defining experiment configurations.

Set in init

The following code snippet demonstrates how to define config using a Python dictionary and how to pass this dictionary as a parameter when initializing a SwanLab experiment:

python

import swanlab

# Define a config dictionary
config = {
  "hidden_layer_sizes": [64, 128],
  "activation": "ELU",
  "dropout": 0.5,
  "num_classes": 10,
  "optimizer": "Adam",
  "batch_normalization": True,
  "seq_length": 100,
}

# Pass the config dictionary when you initialize SwanLab
run = swanlab.init(project="config_example", config=config)

Accessing values in config is similar to accessing values in other dictionaries in Python:

Access values using the key name as an index

python

hidden_layer_sizes = swanlab.config["hidden_layer_sizes"]

Access values using the get() method

python

activation = swanlab.config.get["activation"]

Access values using dot notation
python
```
dropout = swanlab.config.dropout
```

Set with argparse

You can set config using an argparse object. argparse is a very powerful module in the Python standard library (Python >= 3.2) for parsing program arguments from the command-line interface (CLI). This module allows developers to easily write user-friendly command-line interfaces.

You can directly pass the argparse object to set config:

python

import argparse
import swanlab

# Initialize Argparse
parser = argparse.ArgumentParser()
parser.add_argument('--epochs', default=20)
parser.add_argument('--learning-rate', default=0.001)
args = parser.parse_args()

swanlab.init(config=args)

This is equivalent to swanlab.init(config={"epochs": 20, "learning-rate": 0.001})

Set at Different Locations in the Script

You can add more parameters to the config object at different locations throughout your script.

The following code snippet shows how to add new key-value pairs to the config object:

python

import swanlab

# Define a config dictionary
config = {
  "hidden_layer_sizes": [64, 128],
  "activation": "ELU",
  "dropout": 0.5,
  # ... other configuration items
}

# Pass the config dictionary when you initialize SwanLab
run = swanlab.init(project="config_example", config=config)

# Update config after initializing SwanLab
swanlab.config["dropout"] = 0.8
swanlab.config.epochs = 20
swanlab.config.set["batch_size", 32]

Set with Configuration Files

You can initialize config using json and yaml configuration files. For details, please refer to Create Experiment with Configuration File.

Set Experiment Configuration ​