Rewind a run

caution

The ability to rewind a run is in private preview. Contact W&B Support at support@wandb.com to request access to this feature.

W&B currently does not support:

• Log rewind: Logs are reset in the new run segment.
• System metrics rewind: Only new system metrics after the rewind point are logged.
• Artifact association: Artifacts are associated with the source run that produced them.

info

• To rewind a run, you must have W&B Python SDK version >= 0.17.1.
• You must use monotonically increasing steps. You can not use non-monotonic steps defined with define_metric() because it will disrupt the required chronological order of run history and system metrics.

Rewind a run to correct or modify the history of a run without losing the original data. In addition, when you rewind a run, you can log new data from that point in time. The summary metrics for the run you rewind are recomputed based on the newly logged history. This means the following behavior:

• History truncation: The history is truncated to the rewind point, allowing new data logging.
• Summary metrics: Recomputed based on the newly logged history.
• Configuration preservation: Original configurations are preserved and can be merged with new configurations.

When you rewind a run, W&B resets the state of the run to the specified step, preserving the original data and maintaining a consistent run ID. This means that:

• Run archiving: Original runs are archived and accessible from the Run Overview tab.
• Artifact association: Artifacts are associated with the run that produced them.
• Immutable run IDs: Introduced for consistent forking from a precise state.
• Copy immutable run ID: A button to copy the immutable run ID for improved run management.

Rewind and forking compatibility

Forking compliments a rewind by providing more flexibility in managing and experimenting with your runs.

When you fork from a run, W&B creates a new branch off a run at a specific point to try different parameters or models.

When you rewind a run, W&B let's you correct or modify the run history itself.

Rewind a run

Use resume_from with wandb.init() to "rewind" a run’s history to a specific step. Specify the name of the run and the step you want to rewind from:

import wandb
import math

# Initialize the first run and log some metrics
# Replace with your_project_name and your_entity_name!
run1 = wandb.init(project="your_project_name", entity="your_entity_name")
for i in range(300):
    run1.log({"metric": i})
run1.finish()

# Rewind from the first run at a specific step and log the metric starting from step 200
run2 = wandb.init(project="your_project_name", entity="your_entity_name", resume_from=f"{run1.id}?_step=200")

# Continue logging in the new run
# For the first few steps, log the metric as is from run1
# After step 250, start logging the spikey pattern
for i in range(200, 300):
    if i < 250:
        run2.log({"metric": i, "step": i})  # Continue logging from run1 without spikes
    else:
        # Introduce the spikey behavior starting from step 250
        subtle_spike = i + (2 * math.sin(i / 3.0))  # Apply a subtle spikey pattern
        run2.log({"metric": subtle_spike, "step": i})
    # Additionally log the new metric at all steps
    run2.log({"additional_metric": i * 1.1, "step": i})
run2.finish()

View an archived run

After you rewind a run, you can explore archived run with the W&B App UI. Follow these steps to view archived runs:

1. Access the Overview Tab: Navigate to the Overview tab on the run's page. This tab provides a comprehensive view of the run's details and history.
2. Locate the Forked From field: Within the Overview tab, find the Forked From field. This field captures the history of the resumptions. The Forked From field includes a link to the source run, allowing you to trace back to the original run and understand the entire rewind history.

By using the Forked From field, you can effortlessly navigate the tree of archived resumptions and gain insights into the sequence and origin of each rewind.

Fork from a run that you rewind

To fork from a rewound run, use the fork_from argument in wandb.init() and specify the source run ID and the step from the source run to fork from:

import wandb

# Fork the run from a specific step
forked_run = wandb.init(
    project="your_project_name",
    entity="your_entity_name",
    fork_from=f"{rewind_run.id}?_step=500",
)

# Continue logging in the new run
for i in range(500, 1000):
    forked_run.log({"metric": i*3})
forked_run.finish()