Log metrics horizontally in W&B to simplify comparison to future runs

src/art/utils/benchmarking/log_constant_metrics_wandb.py

@@ -0,0 +1,49 @@
+"""Utilities for logging constant baseline metrics to Weights & Biases."""
+
+import wandb
+
+import art
+
+
+async def log_constant_metrics_wandb(
+    model: art.Model,
+    num_steps: int,
+    split_metrics: dict[str, dict[str, float]],
+) -> None:
+    """
+    Log constant metrics to W&B as horizontal lines across all training steps.
+
+    Creates a W&B run and logs the same values at every step from 0 to
+    `num_steps`, producing horizontal reference lines on charts. Useful for
+    comparing training curves against static baselines.
+
+    Parameters
+    ----------
+    model : art.Model
+        The model whose `project` and `name` are used for the W&B run.
+    num_steps : int
+        Total training steps. Metrics are logged at steps 0 through `num_steps`.
+    split_metrics : dict[str, dict[str, float]]
+        Nested dict mapping split names (e.g., "train", "val") to metric dicts.
+        Each metric is logged as "{split}/{metric_name}".
+
+        Example: `{"train": {"loss": 0.5}, "val": {"loss": 0.4, "accuracy": 0.8}}`
+    """
+    run = wandb.init(
+        project=model.project,
+        name=model.name,
+        reinit="create_new",
+    )
+
+    # Prefix metrics with their split names
+    prefixed_metrics = {
+        f"{split}/{key}": value
+        for split, metrics in split_metrics.items()
+        for key, value in metrics.items()
+    }
+
+    # Log at every step to create horizontal lines
+    for step in range(num_steps + 1):
+        run.log(prefixed_metrics, step=step)
+
+    run.finish()