Evaluate projections¶

The ZADU class is the recommended interface. It validates specifications, plans shared exact work, reuses original-space resources, and returns results in specification order.

Build a runner¶

from zadu import ExecutionConfig, ZADU

specs = [
    {"id": "tnc", "params": {"k": 20}},
    {"id": "snc", "params": {"k": 30, "random_state": 0}},
]

runner = ZADU(
    specs,
    original,
    return_local=False,
    execution=ExecutionConfig(memory_budget="4GiB"),
)
scores = runner.measure(projection)

The constructor accepts:

Argument	Meaning
`spec_list`	Ordered measure specifications
`orig`	Original high-dimensional samples
`return_local`	Return pointwise values where a measure supports them
`verbose`	Retained compatibility flag
`geodesic`	Treat original two-column coordinates as longitude/latitude in radians
`max_memory_bytes`	Legacy byte-count memory limit
`execution`	Preferred `ExecutionConfig` interface

Do not set both max_memory_bytes and execution.memory_budget to conflicting values.

Read results and diagnostics separately¶

measure() returns only scientific results. Execution metadata lives in last_run_info:

scores = runner.measure(projection)

print(scores)
print(runner.last_run_info["backend"])
print(runner.last_run_info["planned_peak_bytes"])
print(runner.last_run_info["resources"])

The diagnostic record includes the selected provider, resource fallbacks, memory estimates, build and metric timings, resource consumers, dtype, and reuse. Do not mix these fields into metric-score output or serialize them as if they were scientific results.

Use geodesic original coordinates¶

For spherical positions, pass longitude and latitude in radians as the first two original columns:

runner = ZADU(specs, spherical_coordinates, geodesic=True)
scores = runner.measure(projection)

Geodesic distance applies only to the registered original space. Projection coordinates remain Euclidean. Unsupported accelerator resources fall back to the exact NumPy path and report the reason in diagnostics.

Direct measure calls¶

Standalone measure functions remain useful for one-off or research workflows, but they do not share resources across metrics. See Direct measure functions.