save

Private Deployment Only

This API is only available in the self-hosted (private deployment) version of SwanLab. It is not supported on the public cloud version.

save(
    glob_str: Union[str, bytes, Path],
    base_path: Optional[Union[str, Path]] = None,
    policy: Literal["now", "end", "live"] = "live",
) -> List[str]

Parameter	Description
glob_str	(Union[str, bytes, Path]) Glob pattern for files to save, e.g. "checkpoints/*.pt".
base_path	(Optional[Union[str, Path]]) Base directory for resolving relative paths. Defaults to current working directory.
policy	(Literal["now", "end", "live"]) Save policy: "now" upload immediately; "end" defer until run finishes; "live" watch and re-upload on changes. Defaults to "live".

Introduction

swanlab.save allows you to save files matched by a glob pattern into the current run. This is useful for persisting model checkpoints, logs, or any other artifacts during training.

The function returns a list of relative file paths that were matched and saved.

Usage

import swanlab

swanlab.init(project="my-project", mode="local")

# Save all .pt files in the checkpoints directory
saved = swanlab.save("checkpoints/*.pt")
print(saved)  # ['checkpoints/epoch_1.pt', 'checkpoints/epoch_2.pt']

swanlab.finish()

Save Policies

Policy	Behavior
now	Upload matched files immediately.
end	Defer upload until the run finishes.
live	Watch for file changes and re-upload automatically. (default)

Example: Save at the End of Training

import swanlab

swanlab.init(project="my-project")

# Train your model...
# Save the final checkpoint when training ends
swanlab.save("output/model_final.pt", policy="end")

swanlab.finish()

Example: Live Watch for Checkpoints

import swanlab

swanlab.init(project="my-project")

# Save and watch for changes — useful during long training runs
swanlab.save("checkpoints/*.pt", policy="live")

# ... training loop ...
swanlab.finish()

Notes

1. The glob_str pattern is resolved relative to base_path (or current working directory if not specified).
2. Only regular files are saved — directories are automatically filtered out.
3. There is a limit on the number of files that can be saved in a single call (controlled by save_batch setting).