libact/examples/batch_query_plot.py at master · ntucllab/libact · 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
#!/usr/bin/env python3
"""
Batch-mode active learning example.

Sequential active learning retrains the model after every single query
(make_query -> label -> update), which is impractical when training is
expensive. Batch querying selects batch_size points per round
(make_query_batch -> label -> update_batch), cutting the number of
training rounds by a factor of batch_size.

This script compares, on the diabetes dataset:

- sequential uncertainty sampling (one query per training round),
- plain top-k batch uncertainty sampling (one round per batch, but the
  batch may contain redundant near-duplicate points), and
- DiversityWeightedMeta over the same base strategy (one round per
  batch, with diversity-aware batches).
"""

import copy
import os

import numpy as np
import matplotlib.pyplot as plt
try:
    from sklearn.model_selection import train_test_split
except ImportError:
    from sklearn.cross_validation import train_test_split

# libact classes
from libact.base.dataset import Dataset, import_libsvm_sparse
from libact.models import LogisticRegression
from libact.query_strategies import DiversityWeightedMeta, UncertaintySampling
from libact.labelers import IdealLabeler


def run_sequential(trn_ds, tst_ds, lbr, model, qs, quota):
    """One query per round: quota training rounds in total."""
    n_labeled_axis, E_out = [], []

    for _ in range(quota):
        ask_id = qs.make_query()
        lb = lbr.label(trn_ds.data[ask_id][0])
        trn_ds.update(ask_id, lb)

        model.train(trn_ds)
        n_labeled_axis.append(trn_ds.len_labeled())
        E_out.append(1 - model.score(tst_ds))

    return n_labeled_axis, E_out


def run_batch(trn_ds, tst_ds, lbr, model, qs, quota, batch_size):
    """batch_size queries per round: quota/batch_size training rounds."""
    n_labeled_axis, E_out = [], []

    for _ in range(quota // batch_size):
        ask_ids = qs.make_query_batch(batch_size)
        labels = [lbr.label(trn_ds.data[ask_id][0]) for ask_id in ask_ids]
        trn_ds.update_batch(ask_ids, labels)

        model.train(trn_ds)
        n_labeled_axis.append(trn_ds.len_labeled())
        E_out.append(1 - model.score(tst_ds))

    return n_labeled_axis, E_out


def split_train_test(dataset_filepath, test_size, n_labeled):
    X, y = import_libsvm_sparse(dataset_filepath).format_sklearn()

    X_train, X_test, y_train, y_test = \
        train_test_split(X, y, test_size=test_size)
    trn_ds = Dataset(X_train, np.concatenate(
        [y_train[:n_labeled], [None] * (len(y_train) - n_labeled)]))
    tst_ds = Dataset(X_test, y_test)
    fully_labeled_trn_ds = Dataset(X_train, y_train)

    return trn_ds, tst_ds, fully_labeled_trn_ds


def main():
    dataset_filepath = os.path.join(
        os.path.dirname(os.path.realpath(__file__)), 'diabetes.txt')
    test_size = 0.33    # fraction of samples assigned to the test set
    n_labeled = 10      # number of samples that are initially labeled
    quota = 120         # number of samples to query in total
    batch_size = 10     # number of samples per batch query

    trn_ds, tst_ds, fully_labeled_trn_ds = \
        split_train_test(dataset_filepath, test_size, n_labeled)
    trn_ds2 = copy.deepcopy(trn_ds)
    trn_ds3 = copy.deepcopy(trn_ds)
    lbr = IdealLabeler(fully_labeled_trn_ds)

    # 1) Sequential uncertainty sampling: one training round per label.
    qs1 = UncertaintySampling(trn_ds, method='lc', model=LogisticRegression())
    n1, E1 = run_sequential(
        trn_ds, tst_ds, lbr, LogisticRegression(), qs1, quota)
    print('sequential US      : %3d training rounds for %d labels'
          % (quota, quota))

    # 2) Plain top-k batch: batch_size labels per training round. The
    #    batch may contain redundant near-duplicates.
    qs2 = UncertaintySampling(trn_ds2, method='lc', model=LogisticRegression())
    n2, E2 = run_batch(
        trn_ds2, tst_ds, lbr, LogisticRegression(), qs2, quota, batch_size)
    print('top-k batch US     : %3d training rounds for %d labels'
          % (quota // batch_size, quota))

    # 3) Diversity-aware batches over the same base strategy.
    qs3 = DiversityWeightedMeta(
        trn_ds3,
        base_query_strategy=UncertaintySampling(
            trn_ds3, method='lc', model=LogisticRegression()),
        lmbda=0.5,
        random_state=1126,
    )
    n3, E3 = run_batch(
        trn_ds3, tst_ds, lbr, LogisticRegression(), qs3, quota, batch_size)
    print('diversity batch US : %3d training rounds for %d labels'
          % (quota // batch_size, quota))

    plt.plot(n1, E1, 'g', label='sequential US')
    plt.plot(n2, E2, 'b--o', label='top-k batch US')
    plt.plot(n3, E3, 'r--s', label='DiversityWeightedMeta batch')
    plt.xlabel('Number of labeled samples')
    plt.ylabel('Test error')
    plt.title('Sequential vs batch-mode active learning')
    plt.legend(loc='upper center', bbox_to_anchor=(0.5, -0.05),
               fancybox=True, shadow=True, ncol=3)
    plt.show()


if __name__ == '__main__':
    main()