Pin and compare runs

Learn how to use pinned and baseline runs to keep track of important runs and efficiently evaluate model experiments.

In a workspace with many runs, it can be difficult to keep track of your best performers, production models, failed experiments, or important reference points. The W&B App provides features to help organize and compare runs:

• Pinned runs: Pin a maximum of 5 runs to keep them visible in the workspace and at the top of the runs list.
• Baseline run: Specify a baseline run as your reference point for comparisons. The baseline run is always visible in the workspace and at the top of the runs list. In line plots, the baseline appears with visually distinct styling to help with comparison.

These features are particularly useful for:

• Comparing new experiments against your production model.
• Tracking multiple candidate models during experimentation.
• Evaluating whether new runs improve on your best results.

Pinned and baseline runs are available for W&B Multi-tenant Cloud only.

See Limitations.

Pin runs

Pin up to 5 runs to keep them easily accessible at the top of your workspace. Pinned runs remain visible regardless of sorting or filtering applied to other runs. Pinned runs appear at the top of the run selector with a circular pin icon, separated from other runs by a visual divider.

To pin a run:

1. Navigate to your workspace.
2. In the run selector or runs table, find the run you want to pin.
3. Click the action ... menu, then select Pin run.

To unpin a run, click the pin icon, or follow the procedure to pin the run, but select Unpin run instead.

Manage the baseline run

You can designate one run as the baseline for the workspace to use it as a reference point for evaluating other runs in your workspace.

In the runs selector and runs table, the baseline run appears at the top alongside pinned runs, and has a bookmark icon instead of a pin.

In line plots, lines for the baseline run appear bolder than other lines. When hovering over the plot or legend, the baseline run’s line is dashed.

Set a baseline run

To set a baseline run:

1. Navigate to your workspace.
2. In the run selector or runs table, find the run you want to use as your baseline.
3. Click the action ... menu, then select Set as baseline.

The baseline run appears at the top of the run selector, separated from other runs by a visual divider. The baseline run has a bookmark icon instead of a circle.

Change the baseline run

Only one run can be the baseline at a time. To change which run is your baseline:

1. Navigate to your workspace.
2. In the run selector or runs table, find the run you want to use as your new baseline.
3. Click the action ... menu, then select Replace baseline.

The new run becomes the baseline. Depending on its configuration, the previous baseline run appears alongside other pinned or unpinned runs.

Remove the baseline designation

To remove the baseline designation:

1. Navigate to your workspace.
2. In the run selector or runs table, find the run you want to use as your new baseline.
3. Click the action ... menu, then select Remove baseline.

Depending on its configuration, the previous baseline run appears alongside other pinned or unpinned runs.

Compare runs to the baseline

The baseline run is always visible in line plots for metrics the run has logged. In line plots, lines for the baseline run appear bolder than other lines.

• Hover over a part of the plot to display a tooltip with values for all visible runs, including the baseline run and pinned runs.
  
• Hover over the baseline run’s legend label to display the line prominently. It appears as a heavy dashed line. Lines for other visible runs appear with reduced saturation.
  
• Hover over another run’s legend label to display that run’s line prominently and compare it with the baseline, which appears as a heavy dashed line. Lines for other visible runs appear with reduced saturation.
  

Use cases

This section describes some scenarios where pinned and baseline runs can help guide your experiments.

• Track production models: Ensure that new models meet your quality bar before deployment.
  1. Set your production model as the baseline.
  2. Compare all experiments against your deployed model to identify candidates that outperform production.
• Compare hyperparameter experiments: Evaluate hyperparameter sweeps or manual experiments against your best-known configuration.
  1. Set your best known configuration as the baseline.
  2. Pin promising candidates as you discover them.
  3. Use the line plots to visually compare runs against the baseline.
  4. Continue to update the baseline as you find better configurations.

Example workflow

This section illustrates how pinned and baseline runs can help you to compare runs.

1. Run this example code, which simulates a hyperparameter-tuning scenario with a series of runs. Replace placeholders surrounded with angle brackets (<>) with your own values.

   import wandb
   import random
   import math
   
   def train_model(learning_rate, batch_size, run_name, tags=None):
       """Simulate training a model with given hyperparameters."""
       config = {
           "learning_rate": learning_rate,
           "batch_size": batch_size,
           "optimizer": "adam",
           "architecture": "resnet50"
       }
       
       with wandb.init(
         # Replace with your team and project name
           project="hyperparameter-tuning",
           entity="<team>",
           name=run_name,
           config=config,
           tags=tags or []
       ) as run:
           # Simulate training loop
           for epoch in range(50):
               # Simulated metrics
               accuracy = 0.6 + 0.3 * (1 - math.exp(-learning_rate * epoch / 10))
               loss = 1.0 * math.exp(-learning_rate * epoch / 10)
               
               run.log({
                   "epoch": epoch,
                   "accuracy": accuracy,
                   "loss": loss
               })
   
   # Create baseline run with standard configuration
   train_model(
       learning_rate=0.001,
       batch_size=64,
       run_name="baseline-config",
       tags=["baseline", "production"]
   )
   
   # Experiment with different learning rates
   train_model(
       learning_rate=0.003,
       batch_size=64,
       run_name="lr-experiment-0.003",
       tags=["experiment"]
   )
   
   train_model(
       learning_rate=0.0001,
       batch_size=64,
       run_name="lr-experiment-0.0001",
       tags=["experiment"]
   )
   

   After running this code, your workspace has three runs.

2. Set baseline-config as your baseline run.

3. Pin baseline-config to keep it visible.

4. Compare the experiment runs against the baseline using the line plots in the workspace.

5. Pin promising experiments for further investigation. In this example, after 50 epochs, lr-experiment-0.003 has the highest accuracy (~0.64) and the lowest loss (~0.86).

Limitations

Pinned and baseline runs are available for W&B Multi-tenant Cloud only.

The following features are not yet supported for pinned and baseline runs:

• Grouping: When viewing runs in the run selector or runs table, if runs are grouped by a column, pinned and baseline runs are not visually distinct from other runs.
• Reports: In a run set in a W&B Report, pinned and baseline runs are not visually distinct from other runs.
• Workspace view only: The baseline does not appear when viewing a single run’s workspace.
• Line plots only: Baseline comparison is available only for line plots, and is not yet available for other panels such as bar plots or media panels.