Add simple benchmark setup

kdtree/benchmarks/bench.py

+import argparse
+import glob
+import importlib
+import json
+import os
+import subprocess
+import time
+import kdtree
+import numpy as np
+import pandas as pd
+import seaborn as sns
+import threadpoolctl
+import yaml
+
+from pprint import pprint
+from matplotlib import pyplot as plt
+from memory_profiler import memory_usage
+from sklearn import set_config
+from sklearn.neighbors import KDTree
+
+
+# Be gentle with eyes
+plt.rcParams["figure.dpi"] = 200
+
+
+def benchmark(config, results_folder, bench_name):
+    datasets = config["datasets"]
+    estimators = config["estimators"]
+    leaf_sizes = config["leaf_sizes"]
+
+    n_trials = config.get("n_trials", 3)
+    return_distance = config.get("return_distance", False)
+    one_GiB = 1e9
+    benchmarks = pd.DataFrame()
+
+    env_specs_file = f"{results_folder}/{bench_name}.json"
+
+    # TODO: This is ugly, but I haven't found something better.
+    commit = (
+        str(subprocess.check_output(["git", "rev-parse", "--short", "HEAD"]))
+        .replace("b'", "")
+        .replace("\\n'", "")
+    )
+
+    env_specs = dict(
+        threadpool_info=threadpoolctl.threadpool_info(),
+        commit=commit,
+        config=config,
+    )
+
+    set_config(assume_finite=True)
+
+    with open(env_specs_file, "w") as outfile:
+        json.dump(env_specs, outfile)
+
+    for dataset in datasets:
+        for leaf_size in leaf_sizes:
+            for trial in range(n_trials):
+                dataset = {k: int(float(v)) for k, v in dataset.items()}
+                ns_train, n_features = dataset.values()
+                X_train = np.random.rand(ns_train, n_features)
+                bytes_processed_data = X_train.nbytes
+
+                t0_ = time.perf_counter()
+                tree = KDTree(X_train, leaf_size=256)
+                t1_ = time.perf_counter()
+                time_elapsed = round(t1_ - t0_, 5)
+
+                row = dict(
+                    trial=trial,
+                    implementation="sklearn",
+                    leaf_size=leaf_size,
+                    n_samples_train=ns_train,
+                    n_features=n_features,
+                    time_elapsed=time_elapsed,
+                    throughput=bytes_processed_data / time_elapsed / one_GiB,
+                )
+
+                benchmarks = benchmarks.append(row, ignore_index=True)
+                pprint(row)
+                print("---")
+
+                t0_ = time.perf_counter()
+                tree = kdtree.KDTree(X_train, leaf_size=256)
+                t1_ = time.perf_counter()
+                time_elapsed = round(t1_ - t0_, 5)
+
+                row = dict(
+                    trial=trial,
+                    implementation="kdtree",
+                    leaf_size=leaf_size,
+                    n_samples_train=ns_train,
+                    n_features=n_features,
+                    time_elapsed=time_elapsed,
+                    throughput=bytes_processed_data / time_elapsed / one_GiB,
+                )
+
+                benchmarks = benchmarks.append(row, ignore_index=True)
+                pprint(row)
+                print("---")                
+
+                benchmarks.to_csv(
+                    f"{results_folder}/{bench_name}.csv",
+                    mode="w+",
+                    index=False,
+                )
+
+    # Overriding again now that all the dyn. lib. have been loaded
+    env_specs["threadpool_info"] = threadpoolctl.threadpool_info()
+
+    with open(env_specs_file, "w") as outfile:
+        json.dump(env_specs, outfile)
+
+
+def report(results_folder, bench_name):
+    df = pd.read_csv(glob.glob(f"{results_folder}/*.csv")[0])
+    with open(glob.glob(f"{results_folder}/*.json")[0], "r") as json_file:
+        env_specs = json.load(json_file)
+
+    cols = [
+        "n_samples_train",
+        "n_features",
+    ]
+
+    df[cols] = df[cols].astype(np.uint32)
+    df['d'] = df.n_features.apply(str)
+    df['leaf'] = df.leaf_size.apply(str)
+
+    df_grouped = df.groupby(cols)
+
+    for i, (vals, df) in enumerate(df_grouped):
+        # 16:9 ratio
+        fig = plt.figure(figsize=(24, 13.5))
+        ax = plt.gca()
+        splot = sns.barplot(
+            y="leaf", x="throughput", hue="implementation", data=df, ax=ax
+        )
+        _ = ax.set_xlabel("Throughput (in GB/s)")
+        _ = ax.set_ylabel("Leaf Size")
+        _ = ax.tick_params(labelrotation=45)
+
+        # Adding the numerical values of "x" to bar
+        for p in splot.patches:
+            _ = splot.annotate(
+                f"{p.get_width():.4e}",
+                (p.get_width(), p.get_y() + p.get_height() / 2),
+                ha="center",
+                va="center",
+                size=10,
+                xytext=(0, -12),
+                textcoords="offset points",
+            )
+
+        title = (
+            f"KDTree@{env_specs['commit']} - "
+            f"Euclidean Distance, dtype=np.float64, {df.trial.max() + 1} trials - Bench. Name: {bench_name}\n"
+        )
+        title += (
+            "n_samples_train=%s - n_features=%s" % vals
+        )
+        _ = fig.suptitle(title, fontsize=16)
+        plt.savefig(f"{results_folder}/{bench_name}_{i}.pdf", bbox_inches="tight")
+
+    # Unifying pdf files into one
+    pdf_files = sorted(glob.glob(f"{results_folder}/{bench_name}*.pdf"))
+    subprocess.check_output(
+        ["pdfunite", *pdf_files, f"{results_folder}/{bench_name}.pdf"]
+    )
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser("benchmark")
+
+    parser.add_argument("bench_name")
+
+    args = parser.parse_args()
+
+    bench_name = args.bench_name
+    with open("benchmarks/config.yml", "r") as f:
+        config = yaml.full_load(f)
+
+    results_folder = f"benchmarks/results/{bench_name}"
+    os.makedirs(results_folder, exist_ok=True)
+
+    print(f"Benchmarking {bench_name}")
+    benchmark(config, results_folder, bench_name)
+    print(f"Benchmark results wrote in {results_folder}")
+
+    print(f"Reporting results for {bench_name}")
+    report(results_folder, bench_name)
+    print(f"Reporting results wrote in {results_folder}")
\ No newline at end of file

kdtree/benchmarks/config.yml

+estimators:
+  - name: sklearn
+    estimator: sklearn.neighbors.KDTree
+  - name: cython+
+    estimator: kdtree.KDTree
+
+n_trials: 1
+
+datasets:
+  - n_samples_train: 1e3
+    n_features: 10
+  - n_samples_train: 1e3
+    n_features: 50
+  - n_samples_train: 1e3
+    n_features: 100
+  - n_samples_train: 1e3
+    n_features: 1000
+  
+  - n_samples_train: 1e6
+    n_features: 10
+  - n_samples_train: 1e6
+    n_features: 50
+  - n_samples_train: 1e6
+    n_features: 100
+
+
+leaf_sizes:
+  - 64
+  - 128
+  - 256
+  - 512
+  - 1024
+  - 2048
\ No newline at end of file