AM1_AI/Benchmarking.py at main · tomasctg/AM1_AI · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
import time
from tqdm import tqdm
import numpy as np
import pandas as pd
from numpy.random import RandomState
from Datasets import split_transpose
import tracemalloc


RNG_SEED = 6553


class Benchmark:
    def __init__(self, X, y, n_runs=1000, warmup=100, mem_runs=100, test_sz=0.3, rng_seed=RNG_SEED, same_splits=True):
        self.X = X
        self.y = y
        self.n = n_runs
        self.warmup = warmup
        self.mem_runs = mem_runs
        self.test_sz = test_sz
        self.det = same_splits
        if self.det:
            self.rng_seed = rng_seed
        else:
            self.rng = RandomState(rng_seed)

        self.data = dict()

        print("Benching params:")
        print("Total runs:", self.warmup+self.mem_runs+self.n)
        print("Warmup runs:", self.warmup)
        print("Peak Memory usage runs:", self.mem_runs)
        print("Running time runs:", self.n)
        approx_test_sz = int(self.y.size * self.test_sz)
        print("Train size rows (approx):", self.y.size - approx_test_sz)
        print("Test size rows (approx):", approx_test_sz)
        print("Test size fraction:", self.test_sz)

    def bench(self, model_class, **kwargs):
        name = model_class.__name__
        # train_time, test_time, accuracy
        time_data = np.empty((self.n, 3), dtype=float)
        # train_peak_mem, test_peak_mem
        mem_data = np.empty((self.mem_runs, 2), dtype=float)
        rng = RandomState(self.rng_seed) if self.det else self.rng

        for i in range(self.warmup):
            # Instantiate model with error check for unsupported parameters
            model = model_class(**kwargs)

            # Generate current train-test split
            X_train, X_test, y_train, y_test = split_transpose(
                self.X, self.y,
                test_size=self.test_sz,
                random_state=rng
            )
            # Run training and prediction (timing or memory measurement not recorded)
            model.fit(X_train, y_train)
            model.predict(X_test)

        for i in tqdm(range(self.mem_runs), total=self.mem_runs, desc=f"{name} (MEM)"):

            model = model_class(**kwargs)

            X_train, X_test, y_train, y_test = split_transpose(
                self.X, self.y,
                test_size=self.test_sz,
                random_state=rng
            )

            tracemalloc.start()

            t1 = time.perf_counter()
            model.fit(X_train, y_train)
            t2 = time.perf_counter()

            _, train_peak = tracemalloc.get_traced_memory()
            tracemalloc.reset_peak()

            model.predict(X_test)
            t3 = time.perf_counter()
            _, test_peak = tracemalloc.get_traced_memory()
            tracemalloc.stop()

            mem_data[i,] = (
                train_peak / (1024 * 1024),
                test_peak / (1024 * 1024)
            )

        for i in tqdm(range(self.n), total=self.n, desc=f"{name} (TIME)"):
            model = model_class(**kwargs)

            X_train, X_test, y_train, y_test = split_transpose(
                self.X, self.y,
                test_size=self.test_sz,
                random_state=rng
            )

            t1 = time.perf_counter()
            model.fit(X_train, y_train)
            t2 = time.perf_counter()
            preds = model.predict(X_test)
            t3 = time.perf_counter()

            time_data[i,] = (
                (t2 - t1) * 1000,
                (t3 - t2) * 1000,
                (y_test.flatten() == preds.flatten()).mean()
            )

        self.data[name] = (time_data, mem_data)

    def summary(self, baseline=None):
        aux = []
        for name, (time_data, mem_data) in self.data.items():
            result = {
                'model': name,
                'train_median_ms': np.median(time_data[:, 0]),
                'train_std_ms': time_data[:, 0].std(),
                'test_median_ms': np.median(time_data[:, 1]),
                'test_std_ms': time_data[:, 1].std(),
                'mean_accuracy': time_data[:, 2].mean(),
                'train_mem_median_mb': np.median(mem_data[:, 0]),
                'train_mem_std_mb': mem_data[:, 0].std(),
                'test_mem_median_mb': np.median(mem_data[:, 1]),
                'test_mem_std_mb': mem_data[:, 1].std()
            }
            aux.append(result)
        df = pd.DataFrame(aux).set_index('model')

        if baseline is not None and baseline in self.data:
            df['train_speedup'] = df.loc[baseline,
                                         'train_median_ms'] / df['train_median_ms']
            df['test_speedup'] = df.loc[baseline,
                                        'test_median_ms'] / df['test_median_ms']
            df['train_mem_reduction'] = df.loc[baseline,
                                               'train_mem_median_mb'] / df['train_mem_median_mb']
            df['test_mem_reduction'] = df.loc[baseline,
                                              'test_mem_median_mb'] / df['test_mem_median_mb']
        return df