Rewind a run

Rewind a run to modify its history. When you rewind a run, W&B resets the state of the run to the specified step while maintaining the same run ID. Use rewind to correct logged data or change the trajectory of an experiment from a specific point without starting a new run or losing the run’s identity.

The ability to rewind a run is in active development. It is in preview for Multi-tenant Cloud and Dedicated Cloud, and not yet available in Self-Managed.Because of known performance limitations with rewind, W&B typically recommends forking as an alternative.

Prerequisites

Before you rewind a run, ensure you meet the following prerequisites:

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

Limitations

Rewind doesn’t support the following:

Log rewind: W&B resets logs in the new run segment.
System metrics rewind: W&B logs only new system metrics after the rewind point.
Artifact association: W&B associates artifacts with the source run that produced them.

W&B recomputes the summary metrics for the run you rewind based on the newly logged history. This produces the following behavior:

History truncation: W&B truncates the history to the rewind point, which allows new data logging.
Summary metrics: W&B recomputes summary metrics based on the newly logged history.
Configuration preservation: W&B preserves the original configurations, and you can merge new configurations.

Rewind and forking compatibility

Forking complements a rewind. When you fork from a run, W&B creates a new branch off a run at a specific point so you can try different parameters or models. When you rewind a run, you can correct or modify the run history itself.

To rewind a run from a specific step and log new data from that point, pass the run ID and step to the resume_from parameter in wandb.init(). The resume_from parameter accepts a string in the format [RUN-ID]?_step=[STEP], where [RUN-ID] is the ID of the run you want to rewind and [STEP] is the step to rewind from. The following example demonstrates the rewind workflow. Suppose you log a linear line for 300 steps:

import wandb

# Initialize the first run and log some metrics
with wandb.init(project="[PROJECT]", entity="wandb") as run:
    for i in range(300):
        # Plot a linear line
        run.log({"metric": i, "step": i})

In your project’s workspace, you see a line plot from step 0 to step 300:

Later, you want to rewind the run from step 200 and log a metric called additional_metric that logs i*1.1 from step 200 to step 300. From step 250, you want to log a subtle wavy pattern (i**2 + 2*sin(i/3)) instead of a linear line:

import math

run_ID = "[RUN-ID]" # Replace with the run ID of the run you want to rewind

# Rewind from the first run at a specific step and log the metric starting from step 200
with wandb.init(project="[PROJECT]", entity="wandb", resume_from=f"{run_ID}?_step=200") as run:

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

The following image shows the updated project’s workspace. The plot shows these changes after the rewind:

The line plot shows the original linear line from step 0 to step 200, and the subtle wavy pattern starts from step 250.
W&B creates a plot labeled additional_metric that starts from step 200.

Original linear line plot alongside the additional_metric plot starting from step 200

The run now contains the corrected history from step 200 onward while preserving the original run ID. W&B also archives the original run so you can revisit it later.

View an archived run

After you rewind a run, W&B preserves the original run as an archived resumption that you can revisit at any time. Viewing the archived run lets you compare pre-rewind history with the current state and trace the sequence of rewinds applied to a run. To view an archived run in the W&B App:

Navigate to the Overview tab on the run’s page. This tab provides a view of the run’s details and history.
In 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, so you can trace back to the original run and review the entire rewind history.

Use the Forked From field to navigate the tree of archived resumptions and gain insights into the sequence and origin of each rewind.

Fork from a run that you rewind

After you rewind a run, you can branch off the corrected history to explore alternate trajectories without modifying the rewound run itself. This is useful when you want to test alternate parameters or models from a known-good point in the corrected history. To fork from a rewound run, use the fork_from argument in wandb.init() and specify the source run ID and the step to fork from:

import wandb

# Fork the run from a specific step
forked_run = wandb.init(
    project="[PROJECT]",
    entity="[ENTITY]",
    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()

Guides

Integrations

Reference

Rewind a run

Prerequisites

Limitations

Rewind and forking compatibility

Rewind a run

View an archived run

Fork from a run that you rewind

​Prerequisites

​Limitations

​Rewind and forking compatibility

​Rewind a run

​View an archived run

​Fork from a run that you rewind

Prerequisites

Limitations

Rewind and forking compatibility

Rewind a run

View an archived run

Fork from a run that you rewind