opu-kernel-experiments/evaluate_features.py at master · joneswack/opu-kernel-experiments · 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
import os
import argparse
import json
import time
import itertools
import collections

import numpy as np
import torch
from sklearn.model_selection import train_test_split

import util.data
from util.ridge_regression import RidgeRegression
from util.random_features import project_data
from util.kernels import polynomial_kernel


def run_experiment(data, proj_params, alpha, cg_config, device_config):
    """
    Runs a regression for a single hyperparameter combination.
    Returns validation/test scores and projection/regression timings.
    """

    train_data, test_data, train_labels, test_labels = data

    # depending on device_config, we receive either a GPU tensor or a np matrix
    projection, projection_time = project_data(torch.cat([train_data, test_data], dim=0),
                                    device_config, **proj_params)

    if not device_config['use_cpu_memory']:
        # we need to move the labels to GPU
        train_labels = train_labels.to('cuda:' + str(device_config['active_gpus'][0]))
        test_labels = test_labels.to('cuda:' + str(device_config['active_gpus'][0]))

    since = time.time()

    try:
        if proj_params['num_features'] > 10000:
            if proj_params['num_features'] > len(train_labels):
                # in this case we prefer to invert the n x n matrix instead of the D x D one.
                # linear kernel <=> polynomial with degree 1.
                kernel = lambda device_config, X, Y: polynomial_kernel(device_config, X, Y, var=1., bias=0, degree=1.)
            else:
                kernel = None
            clf = RidgeRegression(device_config, solver='cg_torch', kernel=kernel, **cg_config)
            clf.fit(projection[:len(train_data)], train_labels, alpha)
        else:
            clf = RidgeRegression(device_config, solver='cholesky_torch', kernel=None)
            clf.fit(projection[:len(train_data)], train_labels, alpha)
    except RuntimeError as e:
        print(e)
        return 0, 0, 0

    regression_time = time.time() - since

    test_score = clf.score(projection[len(train_data):], test_labels)

    return test_score, projection_time, regression_time


def parse_args():
    parser = argparse.ArgumentParser()
    parser.add_argument('--dataset_config', type=str, required=True,
                        help='Path to dataset configuration file')
    parser.add_argument('--hyperparameter_config', type=str, required=True,
                        help='Path to hyperparameter configuration file')
    parser.add_argument('--device_config', type=str, required=True,
                        help='Path to device configuration file')

    args = parser.parse_args()

    return args

if __name__ == '__main__':
    args = parse_args()

    print('Loading dataset: {}'.format(args.dataset_config))
    data = util.data.load_dataset(args.dataset_config, binarize_data=True)

    print('Loading hyperparameters: {}'.format(args.hyperparameter_config))
    hyperparameter_config = util.data.load_hyperparameters(args.hyperparameter_config)

    print('Loading device config: {}'.format(args.device_config))
    device_config = util.data.load_device_config(args.device_config)

    # iterate over all the kernel configs
    for kernel, params in hyperparameter_config.items():
        log_name = '_'.join([data[0], kernel])
        log_folder = 'feature_evaluation'
        csv_handler = util.data.DF_Handler(log_folder, log_name)
        log_handler = util.data.Log_Handler(log_folder, log_name)

        log_handler.append('Running experiments for kernel {}'.format(kernel))

        if 'precomputed_features' in params:
            # we can use precomputed OPU features
            precomputed = np.load(params['precomputed_features'])

        # iterate over all cv hyperparameters
        num_experiments = len(params['num_features'])
        for idx, feature_dim in enumerate(params['num_features']):
            print('Progress: {} / {} ({:.2f}%)'.format(idx, num_experiments, 100*float(idx) / num_experiments))

            log_handler.append('Feature Dimension: {}'.format(feature_dim))

            alpha = params['alpha']
            kernel_params = params['kernel_hyperparameters']
            cg_config = params['cg_config']
            proj_params = {**kernel_params, **{'projection': params['projection'], 'num_features': feature_dim}}

            if 'precomputed_features' in params:
                # we can use precomputed OPU features
                proj_params['precomputed'] = precomputed[:, :feature_dim]
                proj_params['raw_features'] = False

            # the lengthscale is sqrt(1 / 2*gamma)
            if 'gamma' in proj_params.keys():
                if proj_params['gamma'] == 'auto':
                    proj_params['lengthscale'] = proj_params.pop('gamma')
                else:
                    proj_params['lengthscale'] = np.sqrt(1. / (2*proj_params.pop('gamma')))

            test_score, proj_time, regr_time = run_experiment(
                data[1:], proj_params, alpha, cg_config, device_config)

            log_dictionary = {**proj_params, **{
                'alpha': alpha, 'test_score': test_score,
                'proj_time': proj_time, 'regr_time': regr_time
            }}

            if 'precomputed' in log_dictionary and log_dictionary['precomputed'] is not None:
                log_dictionary['precomputed'] = True

            csv_handler.append(log_dictionary)
            csv_handler.save()
            log_handler.append('Result: {}'.format(log_dictionary))

    log_handler.append('Experiments completed!')