fix(dbscan): GPU clustering metrics, and eps quantile in the estimator's metric

configs/experiments/dbscan_metrics.json

@@ -0,0 +1,79 @@
+{
+    "INCLUDE": ["../common/sklearn.json"],
+    "PARAMETERS_SETS": {
+        "dbscan metrics": {
+            "algorithm": {
+                "estimator": "DBSCAN",
+                "estimator_params": [
+                    { "metric": "euclidean" },
+                    { "metric": "manhattan" },
+                    { "metric": "chebyshev" },
+                    { "metric": "cosine" },
+                    { "metric": "minkowski", "p": 3 }
+                ]
+            }
+        },
+        "dbscan eps and jobs": {
+            "algorithm": {
+                "estimator_params": {
+                    "eps": "[SPECIAL_VALUE]distances_quantile:0.01",
+                    "min_samples": 5,
+                    "n_jobs": "[SPECIAL_VALUE]physical_cpus"
+                }
+            },
+            "bench": { "n_runs": 5, "time_limit": 900 }
+        },
+        "sklearn brute and tree": [
+            { "algorithm": { "library": "sklearn", "device": "cpu", "estimator_params": { "algorithm": "brute" } } },
+            { "algorithm": { "library": "sklearn", "device": "cpu", "estimator_params": { "algorithm": "auto" } } }
+        ],
+        "sklearnex cpu": {
+            "algorithm": { "library": "sklearnex", "device": "cpu", "estimator_params": { "algorithm": "brute" } }
+        },
+        "sklearnex gpu": {
+            "algorithm": {
+                "library": "sklearnex",
+                "device": "gpu",
+                "estimator_params": { "algorithm": "brute" },
+                "sklearn_context": { "array_api_dispatch": true }
+            },
+            "data": { "format": "dpnp", "order": "C" }
+        },
+        "dbscan datasets": {
+            "data": [
+                {
+                    "source": "make_blobs",
+                    "generation_kwargs": [
+                        { "centers": 20, "n_samples": 50000, "n_features": 8, "cluster_std": 1.0 },
+                        { "centers": 20, "n_samples": 200000, "n_features": 8, "cluster_std": 1.0 },
+                        { "centers": 20, "n_samples": 100000, "n_features": 32, "cluster_std": 1.0 },
+                        { "centers": 20, "n_samples": 50000, "n_features": 256, "cluster_std": 1.0 },
+                        { "centers": 100, "n_samples": 500000, "n_features": 2, "cluster_std": 1.0 }
+                    ],
+                    "split_kwargs": { "ignore": true }
+                },
+                {
+                    "dataset": "cifar",
+                    "split_kwargs": { "ignore": true },
+                    "preprocessing_kwargs": { "normalize": "mean" }
+                },
+                {
+                    "dataset": "covtype",
+                    "split_kwargs": { "train_size": 100000, "test_size": null },
+                    "preprocessing_kwargs": { "normalize": "standard" }
+                }
+            ]
+        }
+    },
+    "TEMPLATES": {
+        "sklearn dbscan": {
+            "SETS": ["sklearn brute and tree", "dbscan metrics", "dbscan eps and jobs", "dbscan datasets"]
+        },
+        "sklearnex cpu dbscan": {
+            "SETS": ["sklearnex cpu", "dbscan metrics", "dbscan eps and jobs", "dbscan datasets"]
+        },
+        "sklearnex gpu dbscan": {
+            "SETS": ["sklearnex gpu", "dbscan metrics", "dbscan eps and jobs", "dbscan datasets"]
+        }
+    }
+}

sklbench/benchmarks/sklearn_estimator.py

@@ -225,7 +225,7 @@ def get_subset_metrics_of_estimator(
                         )
                     }
                 )
-            if len(np.unique(y)) < 128:
+            if len(np.unique(y_compat)) < 128:
                 metrics.update(
                     {
                         "homogeneity": (

sklbench/utils/special_params.py

@@ -21,7 +21,7 @@
 import numpy as np
 import pandas as pd
 from psutil import cpu_count
-from sklearn.metrics import euclidean_distances
+from sklearn.metrics import pairwise_distances
 
 from ..datasets import dataset_loading_functions
 from .bench_case import get_bench_case_value, set_bench_case_value
@@ -288,7 +288,18 @@ def assign_case_special_values_on_run(
         # conversion to lower precision is required
         # to produce same distances quantile for different dtypes of x
         x_sample = x_sample.astype("float32")
-        dist = np.tril(euclidean_distances(x_sample, x_sample)).reshape(-1)
+        # the quantile is taken in the estimator's own metric, so eps has the right scale
+        metric = get_bench_case_value(
+            bench_case, "algorithm:estimator_params:metric", "euclidean"
+        )
+        metric_kwargs = {}
+        if metric == "minkowski":
+            metric_kwargs["p"] = get_bench_case_value(
+                bench_case, "algorithm:estimator_params:p", 2
+            )
+        dist = np.tril(
+            pairwise_distances(x_sample, x_sample, metric=metric, **metric_kwargs)
+        ).reshape(-1)
         dist = dist[dist != 0]
         quantile = float(np.quantile(dist, quantile))
         set_bench_case_value(bench_case, "algorithm:estimator_params:eps", quantile)