Lightning Fabric

DVCLive allows you to add experiment tracking capabilities to your Lightning Fabric projects.

Usage

You need to create a DVCLiveLogger and then log the parameters, metrics, and other info you want to log.

An example code snippet is shown below:

from dvclive.fabric import DVCLiveLogger
from lightning.fabric import Fabric
from lightning.fabric.utilities.rank_zero import rank_zero_only

...

fabric = Fabric()

# Create the DVCLiveLogger
logger = DVCLiveLogger()

# Log dict of hyperparameters
logger.log_hyperparams({"batch_size": 64, "epochs": 5, "lr": 1.0, ...})

for epoch in range(epochs):

    ...

    # Log dict of metrics
    logger.log_metrics({"loss": loss})

fabric.save("mnist_cnn.pt", model.state_dict())
# Check that `rank_zero_only.rank == 0` to avoid logging in other processes.
if rank_zero_only.rank == 0:
    # `logger.experiment` provides access to DVCLive's `Live` instance.
    logger.experiment.log_artifact("mnist_cnn.pt")

# Call finalize to save final results as a DVC experiment
logger.finalize("success")

Initialize

from dvclive.fabric import DVCLiveLogger

logger = DVCLiveLogger()

To customize tracking, include keyword arguments to be passed to the Live instance like:

logger = DVCLiveLogger(dir="my_directory")

Log parameters

To log hyperparameters, pass a dictionary to log_hyperparams():

logger.log_hyperparams({"batch_size": 64, "epochs": 5, "lr": 1.0, ...})

Log metrics

To log metrics, pass a dictionary to log_metrics():

logger.log_metrics({"train_loss": loss})
...
logger.log_metrics({"test_loss": test_loss, "test_acc": test_acc})

Optionally pass the step (if not passed, it will be auto-incremented):

logger.log_metrics({"train_loss": loss}, step=step)

Additional logging

To log models, other artifacts, or images, retrieve the Live instance from the logger:

fabric.save("mnist_cnn.pt", model.state_dict())
if rank_zero_only.rank == 0:
    logger.experiment.log_artifact("mnist_cnn.pt")

rank_zero_only.rank ensures it is only called on the rank zero process. This is handled automatically by the DVCLiveLogger in the other methods above, but in this example we are directly calling Live instead.

logger.experiment calls the Live instance so that you can call any Live methods.

End

Finally, end the experiment to trigger Live.end():

logger.finalize("success")

"success" is passed as the status and is a required argument.