Merge pull request #136 from lucasimi/develop

Added benchmark for custom metrics

Author: lucasimi

tests/test_bench_metrics.py

@@ -0,0 +1,103 @@
+import unittest
+import timeit
+
+import pandas as pd
+import numpy as np
+
+import tdamapper.utils.metrics as metrics
+
+
+def euclidean_numpy(a, b):
+    return np.sqrt(np.sum((a - b) ** 2))
+
+
+def euclidean_numpy_linalg(a, b):
+    return np.linalg.norm(a - b)
+
+
+def manhattan_numpy(a, b):
+    return np.sum(np.abs(a - b))
+
+
+def manhattan_numpy_linalg(a, b):
+    return np.linalg.norm(a - b, ord=1)
+
+
+def chebyshev_numpy(a, b):
+    return np.max(np.abs(a - b))
+
+
+def chebyshev_numpy_linalg(a, b):
+    return np.linalg.norm(a - b, ord=np.inf)
+
+
+def eval_dist(X, d):
+    for i in range(X.shape[0] - 1):
+        d(X[i], X[i+1])
+
+
+def run_dist_bench(X, d):
+    return timeit.timeit(lambda: eval_dist(X, d), number=30)
+
+
+def run_euclidean_bench(X):
+    t_numpy = run_dist_bench(X, euclidean_numpy)
+    t_numpy_linalg = run_dist_bench(X, euclidean_numpy_linalg)
+    t_tdamapper = run_dist_bench(X, metrics.euclidean())
+    return {
+        'metric': 'euclidean',
+        'numpy': t_numpy,
+        'numpy_linalg': t_numpy_linalg,
+        'tdamapper': t_tdamapper,
+    }
+
+
+def run_chebyshev_bench(X):
+    t_numpy = run_dist_bench(X, chebyshev_numpy)
+    t_numpy_linalg = run_dist_bench(X, chebyshev_numpy_linalg)
+    t_tdamapper = run_dist_bench(X, metrics.chebyshev())
+    return {
+        'metric': 'chebyshev',
+        'numpy': t_numpy,
+        'numpy_linalg': t_numpy_linalg,
+        'tdamapper': t_tdamapper,
+    }
+
+
+def run_manhattan_bench(X):
+    t_numpy = run_dist_bench(X, manhattan_numpy)
+    t_numpy_linalg = run_dist_bench(X, manhattan_numpy_linalg)
+    t_tdamapper = run_dist_bench(X, metrics.manhattan())
+    return {
+        'metric': 'manhattan',
+        'numpy': t_numpy,
+        'numpy_linalg': t_numpy_linalg,
+        'tdamapper': t_tdamapper,
+    }
+
+
+def merge(d, d_part):
+    for k, v in d_part.items():
+        if k not in d:
+            d[k] = []
+        d[k].append(v)
+    return d
+
+
+def run_bench(X):
+    d = {}
+    d_part = run_euclidean_bench(X)
+    merge(d, d_part)
+    d_part = run_chebyshev_bench(X)
+    merge(d, d_part)
+    d_part = run_manhattan_bench(X)
+    merge(d, d_part)
+    return pd.DataFrame(d)
+
+
+class TestBenchMetrics(unittest.TestCase):
+
+    def test_bench(self):
+        X = np.random.rand(1000, 1000)
+        df_bench = run_bench(X)
+        print(df_bench)