-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathanomaly_scores.py
220 lines (168 loc) · 8.48 KB
/
anomaly_scores.py
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
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
import torch
import numpy as np
from tqdm import tqdm
from GDSS.utils.data_loader import dataloader
from GDSS.utils.graph_utils import adjs_to_graphs, count_nodes, get_laplacian
from GDSS.utils.plot import plot_graphs_list
from GDSS.reconstruction import Reconstructor
from GDSS.likelihood import LikelihoodEstimator
def calculate_scores(config, loader, data_len, exp_name, num_sample=1, plot_graphs=True):
reconstructor = Reconstructor(config)
x_scores = torch.zeros(data_len)
adj_scores = torch.zeros(data_len)
gen_graph_list = []
orig_graph_list = []
batch_start_pos = 0
for i, batch in tqdm(enumerate(loader)):
x = batch[0]
adj = batch[1]
# Normalization terms (number of nodes in each graph and number of features)
num_nodes = count_nodes(adj)
num_feat = x.shape[2]
x_err = torch.zeros(x.shape[0])
adj_err = torch.zeros(adj.shape[0])
for _ in range(num_sample):
with torch.no_grad():
x_reconstructed, adj_reconstructed = reconstructor(batch)
x_reconstructed = x_reconstructed.to('cpu')
adj_reconstructed = adj_reconstructed.to('cpu')
x_err = x_err + torch.linalg.norm(x - x_reconstructed, dim=[1, 2]) / (num_nodes * num_feat)
adj_err = adj_err + torch.linalg.norm(adj - adj_reconstructed, dim=[1, 2]) / (num_nodes**2)
x_err /= num_sample
adj_err /= num_sample
bs = x.shape[0]
batch_end_pos = batch_start_pos + bs
x_scores[batch_start_pos:batch_end_pos] = x_err
adj_scores[batch_start_pos:batch_end_pos] = adj_err
batch_start_pos = batch_end_pos
# Convert the first batch to networkx for plotting
if i == 0 and plot_graphs:
eps = 1e-9
rel_x_err = torch.linalg.norm(x - x_reconstructed, dim=[2])
rel_x_err /= (torch.linalg.norm(x, dim=[2]) + eps)
nx_graphs, empty_nodes = adjs_to_graphs(adj.numpy(), False, return_empty=True)
orig_graph_list.extend(nx_graphs)
gen_graph_list.extend(adjs_to_graphs(adj_reconstructed.numpy(), False, empty_nodes=empty_nodes))
if plot_graphs:
pos_list = plot_graphs_list(graphs=orig_graph_list, title=f'orig_{exp_name}', rows=4, cols=2, save_dir='./')
_ = plot_graphs_list(graphs=gen_graph_list, title=f'reconstruction_{exp_name}', rows=4, cols=2, save_dir='./',
pos_list=pos_list, rel_x_err=rel_x_err)
return x_scores, adj_scores
def calculate_energy(x, adj, sym):
L = get_laplacian(adj, sym=sym)
E = torch.bmm(x.transpose(-1, -2), L)
E = torch.bmm(E, x)
E = torch.diagonal(E, offset=0, dim1=-2, dim2=-1).sum(-1)
return E
def calculate_energy_scores(config, loader, data_len, exp_name, num_sample=1, plot_graphs=True):
reconstructor = Reconstructor(config)
E_orig = torch.zeros(data_len)
E_rec = torch.zeros((data_len, num_sample))
X_norm_orig = torch.zeros(data_len)
X_norm_rec = torch.zeros((data_len, num_sample))
gen_graph_list = []
orig_graph_list = []
batch_start_pos = 0
for i, batch in tqdm(enumerate(loader)):
x = batch[0]
adj = batch[1]
bs = x.shape[0]
batch_end_pos = batch_start_pos + bs
X_norm_orig[batch_start_pos:batch_end_pos] = torch.linalg.norm(x, dim=(1,2))
batch_E_orig = calculate_energy(x, adj, sym=config.model.sym)
batch_E_rec = torch.zeros((bs, num_sample))
for sample_idx in range(num_sample):
with torch.no_grad():
x_reconstructed, adj_reconstructed = reconstructor(batch)
x_reconstructed = x_reconstructed.to('cpu')
adj_reconstructed = adj_reconstructed.to('cpu')
batch_E_rec[:, sample_idx] = calculate_energy(x_reconstructed, adj_reconstructed, sym=config.model.sym)
X_norm_rec[batch_start_pos:batch_end_pos, sample_idx] = torch.linalg.norm(x_reconstructed, dim=(1,2))
E_orig[batch_start_pos:batch_end_pos] = batch_E_orig
E_rec[batch_start_pos:batch_end_pos] = batch_E_rec
batch_start_pos = batch_end_pos
# Convert the first batch to networkx for plotting
if i == 0 and plot_graphs:
eps = 1e-9
rel_x_err = torch.linalg.norm(x - x_reconstructed, dim=[2])
rel_x_err /= (torch.linalg.norm(x, dim=[2]) + eps)
nx_graphs, empty_nodes = adjs_to_graphs(adj.numpy(), False, return_empty=True)
orig_graph_list.extend(nx_graphs)
gen_graph_list.extend(adjs_to_graphs(adj_reconstructed.numpy(), False, empty_nodes=empty_nodes))
if plot_graphs:
pos_list = plot_graphs_list(graphs=orig_graph_list, title=f'orig_{exp_name}', max_num=16, save_dir='./')
_ = plot_graphs_list(graphs=gen_graph_list, title=f'reconstruction_{exp_name}', max_num=16, save_dir='./',
pos_list=pos_list, rel_x_err=rel_x_err)
return E_orig, E_rec, X_norm_orig, X_norm_rec
def save_final_scores(config, dataset, exp_name, trajectory_sample, num_sample=1, num_steps=100, is_energy=False):
loader = dataloader(config,
dataset,
shuffle=False,
drop_last=False)
data_len = len(dataset)
endtime = config.sde.adj.endtime
T_lst = np.linspace(0, endtime, trajectory_sample + 2, endpoint=True)[1:-1]
if is_energy:
E_rec_final = torch.zeros((data_len, num_sample, trajectory_sample))
X_norm_rec_final = torch.zeros((data_len, num_sample, trajectory_sample))
else:
x_scores_final = torch.zeros((data_len, trajectory_sample))
adj_scores_final = torch.zeros((data_len, trajectory_sample))
for i, T in enumerate(T_lst):
config.sde.x.endtime = T
config.sde.adj.endtime = T
new_num_scales = int(T * num_steps)
config.sde.x.num_scales = new_num_scales
config.sde.adj.num_scales = new_num_scales
new_exp_name = f'{exp_name}_scales_{new_num_scales}'
if is_energy:
E_orig_final, E_rec, X_norm_orig_final, X_norm_rec = calculate_energy_scores(config, loader, data_len, new_exp_name,
num_sample=num_sample, plot_graphs=False)
E_rec_final[:, :, i] = E_rec
X_norm_rec_final[:, :, i] = X_norm_rec
else:
x_scores, adj_scores = calculate_scores(config, loader, data_len, new_exp_name,
num_sample=num_sample, plot_graphs=False)
x_scores_final[:, i] = x_scores
adj_scores_final[:, i] = adj_scores
with open(f'{exp_name}_final_scores.npy', 'wb') as f:
if is_energy:
np.save(f, E_orig_final.numpy())
np.save(f, E_rec_final.numpy())
np.save(f, X_norm_orig_final.numpy())
np.save(f, X_norm_rec_final.numpy())
else:
np.save(f, x_scores_final.numpy())
np.save(f, adj_scores_final.numpy())
def save_likelihood_scores(config, dataset, exp_name, num_sample):
likelihood_estimator = LikelihoodEstimator(config, num_sample)
loader = dataloader(config,
dataset,
shuffle=False,
drop_last=False,
dequantize=True)
data_len = len(dataset)
prior_constant_x = torch.zeros(data_len)
prior_constant_adj = torch.zeros(data_len)
prior_logp_x = torch.zeros(data_len)
prior_logp_adj = torch.zeros(data_len)
delta_logp = torch.zeros(data_len)
batch_start_pos = 0
for i, batch in tqdm(enumerate(loader)):
x = batch[0]
adj = batch[1]
bs = x.shape[0]
batch_end_pos = batch_start_pos + bs
likelihood_components = likelihood_estimator(batch)
prior_constant_x[batch_start_pos:batch_end_pos] = likelihood_components[0]
prior_constant_adj[batch_start_pos:batch_end_pos] = likelihood_components[1]
prior_logp_x[batch_start_pos:batch_end_pos] = likelihood_components[2]
prior_logp_adj[batch_start_pos:batch_end_pos] = likelihood_components[3]
delta_logp[batch_start_pos:batch_end_pos] = likelihood_components[4]
batch_start_pos = batch_end_pos
with open(f'{exp_name}_final_scores.npy', 'wb') as f:
np.save(f, prior_constant_x.numpy())
np.save(f, prior_constant_adj.numpy())
np.save(f, prior_logp_x.numpy())
np.save(f, prior_logp_adj.numpy())
np.save(f, delta_logp.numpy())