forked from ermongroup/BCD-Nets
-
Notifications
You must be signed in to change notification settings - Fork 0
/
dag_utils.py
390 lines (341 loc) · 13 KB
/
dag_utils.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
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
import logging
import numpy as np
import networkx as nx
from typing import Dict
import jax.random as rnd
import numpy as np
from typing import Any
from cdt.data import load_dataset
# from dag_data import is_dag
import csv
debug_list = lambda x: list(x)
class SyntheticDataset(object):
_logger = logging.getLogger(__name__)
def __init__(
self,
n,
d,
graph_type,
degree,
sem_type,
noise_scale=1.0,
dataset_type="linear",
quadratic_scale=None,
):
self.n = n
self.d = d
self.graph_type = graph_type
self.degree = degree
self.sem_type = sem_type
self.noise_scale = noise_scale
self.dataset_type = dataset_type
self.w_range = (0.5, 2.0)
self.quadratic_scale = quadratic_scale
self._setup()
self._logger.debug("Finished setting up dataset class")
def _setup(self):
self.W, self.W_2, self.P = SyntheticDataset.simulate_random_dag(
self.d,
self.degree,
self.graph_type,
self.w_range,
(self.dataset_type != "linear"),
)
if self.dataset_type != "linear":
assert self.W_2 is not None
self.W_2 = self.W_2 * self.quadratic_scale
self.X = SyntheticDataset.simulate_sem(
self.W,
self.n,
self.sem_type,
self.w_range,
self.noise_scale,
self.dataset_type,
self.W_2,
)
@staticmethod
def simulate_random_dag(d, degree, graph_type, w_range, return_w_2=False):
"""Simulate random DAG with some expected degree.
Args:
d: number of nodes
degree: expected node degree, in + out
graph_type: {erdos-renyi, barabasi-albert, full}
w_range: weight range +/- (low, high)
return_w_2: boolean, whether to return an additional
weight matrix used for quadratic terms
Returns:
W: weighted DAG
[Optional] W: weighted DAG with same occupancy but different weights
"""
if graph_type == "erdos-renyi":
prob = float(degree) / (d - 1)
B = np.tril((np.random.rand(d, d) < prob).astype(float), k=-1)
elif graph_type == "barabasi-albert":
m = int(round(degree / 2))
B = np.zeros([d, d])
bag = [0]
for ii in range(1, d):
dest = np.random.choice(bag, size=m)
for jj in dest:
B[ii, jj] = 1
bag.append(ii)
bag.extend(dest)
elif graph_type == "full": # ignore degree, only for experimental use
B = np.tril(np.ones([d, d]), k=-1)
else:
raise ValueError("Unknown graph type")
# random permutation
P = np.random.permutation(np.eye(d, d)) # permutes first axis only
B_perm = P.T.dot(B).dot(P)
U = np.random.uniform(low=w_range[0], high=w_range[1], size=[d, d])
U[np.random.rand(d, d) < 0.5] *= -1
W = (B_perm != 0).astype(float) * U
U_2 = np.random.uniform(low=w_range[0], high=w_range[1], size=[d, d])
U_2[np.random.rand(d, d) < 0.5] *= -1
W_2 = (B_perm != 0).astype(float) * U_2
# At the moment the generative process is P.T @ lower @ P, we want
# it to be P' @ upper @ P'.T.
# We can return W.T, so we are saying W.T = P'.T @ lower @ P.
# We can then return P.T, as we have
# (P.T).T @ lower @ P.T = W.T
if return_w_2:
return W.T, W_2.T, P.T
else:
return W.T, None, P.T
@staticmethod
def simulate_gaussian_dag(d, degree, graph_type, w_std):
"""Simulate dense DAG adjacency matrix
Args:
d: number of nodes
degree: expected node degree, in + out
graph_type: {erdos-renyi, barabasi-albert, full}
w_range: weight range +/- (low, high)
return_w_2: boolean, whether to return an additional
weight matrix used for quadratic terms
Returns:
W: weighted DAG
[Optional] W: weighted DAG with same occupancy but different weights
"""
lower_entries = np.random.normal(loc=0.0, scale=w_std, size=(d * (d - 1) // 2))
L = np.zeros((d, d))
# We want the ground-truth W.T to be generated from PLP^\top
# This is since we encode W.T as PLP^\top in the approach.
L[np.tril_indices(d, -1)] = lower_entries
P = np.random.permutation(np.eye(d, d)) # permutes first axis only
W = (P @ L @ P.T).T
return W, None, P, L
@staticmethod
def simulate_sem(
W,
n,
sem_type,
w_range,
noise_scale=1.0,
dataset_type="nonlinear_1",
W_2=None,
sigmas=None,
) -> np.ndarray:
"""Simulate samples from SEM with specified type of noise.
Args:
W: weigthed DAG
n: number of samples
sem_type: {linear-gauss,linear-exp,linear-gumbel}
noise_scale: scale parameter of noise distribution in linear SEM
Returns:
X: [n,d] sample matrix
"""
G = nx.DiGraph(W)
d = W.shape[0]
X = np.zeros([n, d], dtype=np.float64)
if sigmas is None:
sigmas = np.ones((d,)) * noise_scale
ordered_vertices = list(nx.topological_sort(G))
assert len(ordered_vertices) == d
for j in ordered_vertices:
parents = list(G.predecessors(j))
if dataset_type == "linear":
eta = X[:, parents].dot(W[parents, j])
elif dataset_type == "quadratic":
eta = X[:, parents].dot(W[parents, j]) + (X[:, parents] ** 2).dot(
W_2[parents, j]
)
else:
raise ValueError("Unknown dataset type")
if sem_type == "linear-gauss":
X[:, j] = eta + np.random.normal(scale=sigmas[j], size=n)
elif sem_type == "linear-exp":
X[:, j] = eta + np.random.exponential(scale=sigmas[j], size=n)
elif sem_type == "linear-gumbel":
X[:, j] = eta + np.random.gumbel(scale=sigmas[j], size=n)
else:
raise ValueError("Unknown sem type")
return X
@staticmethod
def intervene_sem(
W, n, sem_type, sigmas=None, idx_to_fix=None, value_to_fix=None,
):
"""Simulate samples from SEM with specified type of noise.
Args:
W: weigthed DAG
n: number of samples
sem_type: {linear-gauss,linear-exp,linear-gumbel}
noise_scale: scale parameter of noise distribution in linear SEM
Returns:
X: [n,d] sample matrix
"""
G = nx.DiGraph(W)
d = W.shape[0]
X = np.zeros([n, d])
if len(sigmas) == 1:
sigmas = np.ones(d) * sigmas
ordered_vertices = list(nx.topological_sort(G))
assert len(ordered_vertices) == d
for j in ordered_vertices:
parents = list(G.predecessors(j))
if j == idx_to_fix:
X[:, j] = value_to_fix
else:
eta = X[:, parents].dot(W[parents, j])
if sem_type == "linear-gauss":
X[:, j] = eta + np.random.normal(scale=sigmas[j], size=n)
elif sem_type == "linear-exp":
X[:, j] = eta + np.random.exponential(scale=sigmas[j], size=n)
elif sem_type == "linear-gumbel":
X[:, j] = eta + np.random.gumbel(scale=sigmas[j], size=n)
else:
raise ValueError("Unknown sem type")
return X
def sample(
self,
W,
n_obs,
sem_type,
w_range,
n_inters=0,
noise_scale=1.0,
dataset_type="nonlinear_1",
W_2=None,
sigmas=None,
):
X_obs = self.simulate_sem(W, n_obs, sem_type, w_range, noise_scale, dataset_type, W_2, sigmas)
d = W.shape[0]
interv_targets = np.zeros((n_obs + n_inters, d), dtype=bool)
X = X_obs
if n_inters:
X_inters = []
intervention_nodes = np.random.choice(list(range(d)), size=n_inters)
for i, node in enumerate(intervention_nodes):
interv_targets[n_obs + i][node] = True
X_inters.append(self.intervene_sem(W, 1, sem_type, sigmas, idx_to_fix=node, value_to_fix=0))
X = np.vstack((X_obs, np.array(X_inters).squeeze()))
return X, interv_targets
def count_accuracy(W_true, W_est, W_und=None) -> Dict["str", float]:
"""
Compute FDR, TPR, and FPR for B, or optionally for CPDAG B + B_und.
Args:
W_true: ground truth graph
W_est: predicted graph
W_und: predicted undirected edges in CPDAG, asymmetric
Returns in dict:
fdr: (reverse + false positive) / prediction positive
tpr: (true positive) / condition positive
fpr: (reverse + false positive) / condition negative
shd: undirected extra + undirected missing + reverse
nnz: prediction positive
"""
B_true = W_true != 0
B = W_est != 0
B_und = None if W_und is None else W_und
d = B.shape[0]
# linear index of nonzeros
pred = np.flatnonzero(B)
cond = np.flatnonzero(B_true)
cond_reversed = np.flatnonzero(B_true.T)
cond_skeleton = np.concatenate([cond, cond_reversed])
# true pos
true_pos = np.intersect1d(pred, cond, assume_unique=True)
if B_und is not None:
# treat undirected edge favorably
pred_und = np.flatnonzero(B_und)
true_pos_und = np.intersect1d(pred_und, cond_skeleton, assume_unique=True)
true_pos = np.concatenate([true_pos, true_pos_und])
# false pos
false_pos = np.setdiff1d(pred, cond_skeleton, assume_unique=True)
if B_und is not None:
false_pos_und = np.setdiff1d(pred_und, cond_skeleton, assume_unique=True) # type: ignore
false_pos = np.concatenate([false_pos, false_pos_und])
# reverse
extra = np.setdiff1d(pred, cond, assume_unique=True)
reverse = np.intersect1d(extra, cond_reversed, assume_unique=True)
# compute ratio
pred_size = len(pred)
if B_und is not None:
pred_size += len(pred_und) # type: ignore
cond_neg_size = 0.5 * d * (d - 1) - len(cond)
fdr = float(len(reverse) + len(false_pos)) / max(pred_size, 1)
tpr = float(len(true_pos)) / max(len(cond), 1)
fpr = float(len(reverse) + len(false_pos)) / max(cond_neg_size, 1)
# structural hamming distance
B_lower = np.tril(B + B.T)
if B_und is not None:
B_lower += np.tril(B_und + B_und.T)
pred_lower = np.flatnonzero(B_lower)
cond_lower = np.flatnonzero(np.tril(B_true + B_true.T))
extra_lower = np.setdiff1d(pred_lower, cond_lower, assume_unique=True)
missing_lower = np.setdiff1d(cond_lower, pred_lower, assume_unique=True)
shd = len(extra_lower) + len(missing_lower) + len(reverse)
return {"fdr": fdr, "tpr": tpr, "fpr": fpr, "shd": shd, "pred_size": pred_size}
def dagify(W):
"""Successively removes edges with smallest absolute weights
until the graph with weight matrix W is a DAG"""
import numpy as onp
# def is_dag(W):
# return onp.abs(np.trace(jax.scipy.linalg.expm(W * W)) - dim) < 0.001
def is_dag(W):
G = nx.DiGraph(np.array(np.abs(W).T > 0))
return nx.is_directed_acyclic_graph(G)
dim = W.shape[0]
while not is_dag(W):
tmp = onp.array(W.copy())
tmp[tmp == 0.0] = onp.nan
min_idx = onp.nanargmin(onp.abs(tmp))
W = onp.array(W.flatten())
W[min_idx] = 0.0
W = W.reshape((dim, dim))
return W
def process_sachs(
center: bool = True,
print_labels: bool = False,
normalize=False,
n_data=None,
rng_key=None,
):
data = []
with open("./data/sachs_observational.csv") as csvfile:
filereader = csv.reader(csvfile, delimiter=",")
for i, row in enumerate(filereader):
if i == 0:
if print_labels:
print(row)
continue
data.append(np.array([float(x) for x in row]).reshape((1, -1)))
if n_data is None:
data_out = np.concatenate(data, axis=0)
else:
if rng_key is None:
data_out = np.concatenate(data, axis=0)[:n_data]
else:
data_out = np.concatenate(data, axis=0)
idxs = rnd.choice(rng_key, len(data_out), shape=(n_data,), replace=False)
data_out = data_out[idxs]
if center:
if normalize:
data_out = (data_out - np.mean(data_out, axis=0)) / np.std(data_out, axis=0)
else:
data_out = data_out - np.mean(data_out, axis=0)
return data_out
def get_sachs_ground_truth():
"""Labels are ['praf', 'pmek', 'plcg', 'PIP2', 'PIP3', 'p44/42', 'pakts473',
'PKA', 'PKC', 'P38', 'pjnk']."""
W = np.load("./data/sachs_ground_truth.npy")
return W