Skip to content

Container Metrics

Managing multiple metrics individually can lead to boilerplate code, especially when manually synchronizing, computing, and resetting states across distributed workers. The d9d framework provides container metrics to bundle and manage collections of metrics efficiently.

Compose Metric

The ComposeMetric wraps a mapping string keys to Metric instances into a single unifying interface.

By design, you cannot call .update() directly on a ComposeMetric since different metrics may require different underlying arguments. Instead, you access and update the children directly. However, collective lifecycle methods like .sync(), .compute(), .reset(), .state_dict(), and .to() automatically cascade to all underlying metrics.

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
import torch
from d9d.metric.impl.aggregation import SumMetric, WeightedMeanMetric
from d9d.metric.impl.container import ComposeMetric

# 1. Group multiple metrics together
metrics = ComposeMetric({
    "loss": WeightedMeanMetric(),
    "total_samples": SumMetric(),
})

# 2. Update children individually based on their specific signatures
metrics["loss"].update(torch.tensor(0.5), torch.tensor(32.0))
metrics["total_samples"].update(torch.tensor(32.0))

# 3. Lifecycle operations naturally propagate to all children via the container
metrics.to("cuda")
metrics.sync(dist_context)

# 4. Compute returns a dictionary mapping metric names to their final aggregated results
results = metrics.compute()

# 5. Reset all metrics for the next epoch/evaluation step
metrics.reset()

d9d.metric.impl.container.ComposeMetric

Bases: Metric[dict[str, Any]]