02. 실행 비교와 메트릭 및 아티팩트 조회

$env:MLFLOW_TRACKING_URI = "sqlite:///mlflow.db"
$env:MLFLOW_REGISTRY_URI = $env:MLFLOW_TRACKING_URI
$env:MLFLOW_EXPERIMENT_NAME = "mlflow-wine-study"

import json
from pathlib import Path

import mlflow


experiment = mlflow.get_experiment_by_name("mlflow-wine-study")
if experiment is None:
    raise RuntimeError("Experiment not found in the current tracking backend")

runs = mlflow.search_runs(
    experiment_ids=[experiment.experiment_id],
    filter_string="tags.study = 'mlflow-aml'",
    max_results=1000,
)
if runs.empty:
    raise RuntimeError("No study runs found")

completed = runs.loc[runs["status"].str.upper().eq("FINISHED")].copy()
if completed.empty:
    raise RuntimeError("No completed runs found")

for key in ["params.dataset_sha256", "params.split_seed"]:
    if completed[key].nunique(dropna=False) != 1:
        raise RuntimeError(f"Runs have different comparison conditions: {key}")

ranked = completed.sort_values(
    ["metrics.validation_accuracy", "metrics.validation_log_loss"],
    ascending=[False, True],
)
columns = [
    "run_id", "params.C", "metrics.validation_accuracy",
    "metrics.validation_log_loss",
]
print(ranked[columns].to_string(index=False))

selected = {
    "tracking_uri": mlflow.get_tracking_uri(),
    "run_id": str(ranked.iloc[0]["run_id"]),
}
Path("selected-run.json").write_text(
    json.dumps(selected, indent=2), encoding="utf-8"
)
print("Selection written to selected-run.json")

.\.venv\Scripts\python.exe .\compare.py

import json
from pathlib import Path

import mlflow
from mlflow.tracking import MlflowClient


selection = json.loads(
    Path("selected-run.json").read_text(encoding="utf-8")
)
if selection["tracking_uri"] != mlflow.get_tracking_uri():
    raise RuntimeError("Tracking backend changed; run compare.py again")

run_id = selection["run_id"]
client = MlflowClient()
run = client.get_run(run_id)
print("parameters:", run.data.params)
print("latest metrics:", run.data.metrics)

for point in client.get_metric_history(run_id, "validation_accuracy"):
    print("metric history:", point.step, point.value)

for artifact in client.list_artifacts(run_id):
    print("artifact:", artifact.path, "directory:", artifact.is_dir)

model_path = mlflow.artifacts.download_artifacts(
    run_id=run_id,
    artifact_path="model",
)
print("downloaded model folder:", model_path)

.\.venv\Scripts\python.exe .\inspect_run.py