TreeGrad/plot_inaccuracy.py at main · watml/TreeGrad · 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
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
import os
NUM_THREAD = 1
os.environ["OMP_NUM_THREADS"] = f"{NUM_THREAD}"
os.environ["OPENBLAS_NUM_THREADS"] = f"{NUM_THREAD}"
os.environ["MKL_NUM_THREADS"] = f"{NUM_THREAD}"
os.environ["VECLIB_MAXIMUM_THREADS"] = f"{NUM_THREAD}"
os.environ["NUMEXPR_NUM_THREADS"] = f"{NUM_THREAD}"
import numpy as np
from createTreeModel import createTreeModel
from utilFuncs import treeUtility
from linearTreeShap import linear_treeshap
from TreeGrad import treegrad_shap
from TreeProb import treeprob, treeprob_worsetime
import argparse
import matplotlib.pyplot as plt
import seaborn as sns
from matplotlib.ticker import MaxNLocator


n_samples = 5
depths = range(35, 66)


if not os.path.exists('data_inaccuracy.npz'):
    parser = argparse.ArgumentParser()
    parser.add_argument("-p", type=int, default=1, help="number of processes")
    n_processes = parser.parse_args().p
    print('number of processes:', n_processes)

    lts = np.empty((len(depths), n_samples), dtype=np.float64)
    lts_wc = np.empty_like(lts)

    tp_min = np.empty_like(lts)
    tp_nomin = np.empty_like(lts)
    tp_worsetime_min = np.empty_like(lts)
    tp_worsetime_nomin = np.empty_like(lts)

    tg_shap_min = np.empty_like(lts)
    tg_shap_nomin = np.empty_like(lts)

    for i, depth in enumerate(depths):
        print(depth)
        model, X_test, y_test = createTreeModel(1219, 0, 2025, depth)
        for sample_id in range(n_samples):
            x = X_test[sample_id]
            util = treeUtility(model, x, 0)
            gt = util.groundtruth_bruteforce((1, 1), n_processes=n_processes)

            r = linear_treeshap(model, x, 0)
            lts[i, sample_id] = np.linalg.norm(r - gt)

            r = linear_treeshap(model, x, 0, True)
            lts_wc[i, sample_id] = np.linalg.norm(r - gt)

            r = treeprob(model, x, (1, 1), 0)
            tp_min[i, sample_id] = np.linalg.norm(r - gt)

            r = treeprob(model, x, (1, 1), 0, test=True)
            tp_nomin[i, sample_id] = np.linalg.norm(r - gt)

            r = treeprob_worsetime(model, x, (1, 1), 0)
            tp_worsetime_min[i, sample_id] = np.linalg.norm(r - gt)

            r = treeprob_worsetime(model, x, (1, 1), 0, test=True)
            tp_worsetime_nomin[i, sample_id] = np.linalg.norm(r - gt)

            r = treegrad_shap(model, x, (1, 1), 0)
            tg_shap_min[i, sample_id] = np.linalg.norm(r - gt)

            r = treegrad_shap(model, x, (1, 1), 0, test=True)
            tg_shap_nomin[i, sample_id] = np.linalg.norm(r - gt)


    np.savez_compressed('data_inaccuracy.npz',
                        lts=lts,
                        lts_wc=lts_wc,
                        tp_min=tp_min,
                        tp_nomin=tp_nomin,
                        tp_worsetime_min=tp_worsetime_min,
                        tp_worsetime_nomin=tp_worsetime_nomin,
                        tg_shap_min=tg_shap_min,
                        tg_shap_nomin=tg_shap_nomin)

else:
    data = np.load('data_inaccuracy.npz')
    lts = data['lts']
    lts_wc = data['lts_wc']
    tp_min = data['tp_min']
    tp_nomin = data['tp_nomin']
    tp_worsetime_min = data['tp_worsetime_min']
    tp_worsetime_nomin = data['tp_worsetime_nomin']
    tg_shap_min = data['tg_shap_min']
    tg_shap_nomin = data['tg_shap_nomin']


p1 = sns.color_palette('hls', 8)
p2 = sns.color_palette('husl', 8)
p3 = sns.color_palette("Set2")
colors = [p1[-2], p2[1], p2[3]]
fig, ax = plt.subplots(figsize=(32, 24))
plt.grid()
ax.tick_params(axis='x', labelsize=80)
ax.tick_params(axis='y', labelsize=80)
plt.yscale('log')
ax.xaxis.set_major_locator(MaxNLocator(integer=True)) # force integer xtick labels
labels = ['Linear TreeShap', 'Linear TreeShap (well-conditioned)', 'TreeGrad-Shap']
index = 0
for label, curve in zip(labels, [lts, lts_wc, tg_shap_min]):
    curve_mean = curve.mean(axis=1)
    curve_std = curve.std(axis=1)

    ax.plot(depths, curve_mean, linewidth=10, label=label, color=colors[index])
    ax.fill_between(depths, curve_mean - curve_std, curve_mean + curve_std, alpha=0.2, color=colors[index])
    index += 1

plt.xlabel(r'depth $D$', fontsize=100)
plt.ylabel(r'$\|\hat{\phi} - \phi\|_{2}$', fontsize=100)
plt.legend(fontsize=60, framealpha=0.5)
plt.savefig('inaccuracy_main.pdf', bbox_inches='tight')
plt.close(fig)


colors = [p1[-2], p2[1], p2[3], p1[-4], p3[2], p1[2], p1[-1], p1[-3]]
fig, ax = plt.subplots(figsize=(32, 24))
plt.grid()
ax.tick_params(axis='x', labelsize=80)
ax.tick_params(axis='y', labelsize=80)
plt.yscale('log')


labels = ['Linear TreeShap', 'Linear TreeShap (well-conditioned)', 'TreeGrad-Shap w/ min',
          'TreeGrad-Shap w/o min',
          'TreeProb w/ min', 'TreeProb w/o min', 'TreeProb (modified) w/ min',
          'TreeProb (modified) w/o min']
index = 0
for label, curve in zip(labels, [lts, lts_wc, tg_shap_min, tg_shap_nomin, tp_min, tp_nomin, tp_worsetime_min,
                                 tp_worsetime_nomin]):
    curve_mean = curve.mean(axis=1)
    curve_std = curve.std(axis=1)

    ax.plot(depths, curve_mean, linewidth=10, label=label, color=colors[index])
    ax.fill_between(depths, curve_mean - curve_std, curve_mean + curve_std, alpha=0.2, color=colors[index])
    index += 1

plt.xlabel(r'depth $D$', fontsize=100)
plt.ylabel(r'$\|\hat{\phi} - \phi\|_{2}$', fontsize=100)
plt.legend(fontsize=60, framealpha=0.5)
plt.savefig('inaccuracy_app.pdf', bbox_inches='tight')
plt.close(fig)