Save and Load State Using Checkpoints#

Leveraging detached mode, you can easily save the state at a particular point during training and restore it when needed. This is especially useful for resuming training after interruptions or failures.

For the full script, visit the GitHub repository.

Objectives#

These step-by-step instructions walk you through:

Initializing the core context with checkpoint storage
Loading the latest checkpoint
Resuming sending metrics to the trial
Saving checkpoints periodically

After completing this guide, you will be able to:

Understand how checkpoints operate in detached mode
Implement state-saving and restoration in your training routine
Use the Core API to handle checkpoints effectively

Prerequisites#

Required

A Determined cluster

Recommended

Perform Simple Metrics Reporting

Step 1: Initialize the Core Context with Checkpoint Storage#

To begin, you need to set up the core context, specifying the checkpoint storage path. If recovering from a failure, an external experiment and trial ID can be used to identify which artifact to log metrics to:

def main():
    core_v2.init(
        defaults=core_v2.DefaultConfig(
            name="unmanaged-2-checkpoints",
            checkpoint_storage="/path/to/checkpoint",
        ),
        unmanaged=core_v2.UnmanagedConfig(
            external_experiment_id="my-existing-experiment",
            external_trial_id="my-existing-trial",
        ),
    )

Step 2: Load the Latest Checkpoint#

Fetch the latest checkpoint and load it:

latest_checkpoint = core_v2.info.latest_checkpoint
initial_i = 0
if latest_checkpoint is not None:
    with core_v2.checkpoint.restore_path(latest_checkpoint) as path:
        with (path / "state").open() as fin:
            ckpt = fin.read()
            i_str, _ = ckpt.split(",")
            initial_i = int(i_str)

Step 3: Resume Sending Metrics to the Trial#

Continue logging metrics to the trial from where you left off:

for i in range(initial_i, initial_i + 100):
    loss = random.random()
    print(f"training loss is: {loss}")
    core_v2.train.report_training_metrics(steps_completed=i, metrics={"loss": loss})

Step 4: Save Checkpoints Periodically#

Store a new checkpoint every 10 steps:

if (i + 1) % 10 == 0:
    with core_v2.checkpoint.store_path({"steps_completed": i}) as (path, uuid):
        with (path / "state").open("w") as fout:
            fout.write(f"{i},{loss}")

End your training script and close the core context:

core_v2.close()

Navigate to <DET_MASTER_IP:PORT> in your web browser to see the experiment.

Next Steps#

Having walked through this guide, you now understand how to effectively use checkpoints in detached mode. Try more examples using detached mode or learn more about Determined by visiting the tutorials.

Save and Load State Using Checkpoints

Contents