forked from junjieliang672/L-DKGPR
-
Notifications
You must be signed in to change notification settings - Fork 0
/
L-DKGPR.py
436 lines (386 loc) · 16.3 KB
/
L-DKGPR.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
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.optim import Adam
from torchvision import datasets, transforms
import torch.distributions as tdist
import matplotlib.pyplot as plt
import numpy as np
from torch.utils.data import DataLoader, Dataset
from sklearn.manifold import TSNE
from sklearn.metrics import normalized_mutual_info_score as ami
from collections import Counter
from sklearn.metrics import r2_score
import seaborn as sns
from scipy import io
import time
from collections import defaultdict
import pandas as pd
from sklearn.model_selection import train_test_split
import argparse
from torch.utils.data import Dataset, DataLoader
class MyDataset(Dataset):
def __init__(self, trainX, trainy, trainId, trainOid):
self.trainX = trainX
self.trainy = trainy
self.trainId = trainId
self.trainOid = trainOid
def __getitem__(self, i):
return self.trainX[i], self.trainy[i], self.trainId[i], self.trainOid[
i] # the last index is the observation index
def __len__(self):
return len(self.trainy)
class RBFKernel(nn.Module):
def __init__(self, input_dim):
super(RBFKernel, self).__init__()
self.input_dim = input_dim
self.log_std = nn.Parameter(torch.zeros([1]))
self.log_ls = nn.Parameter(torch.zeros([self.input_dim]))
def _square_scaled_dist(self, X, Z=None):
ls = self.log_ls.exp()
scaled_X = X / ls[None, :]
scaled_Z = Z / ls[None, :]
X2 = scaled_X.pow(2).sum(1, keepdim=True)
Z2 = scaled_Z.pow(2).sum(1, keepdim=True)
XZ = scaled_X @ scaled_Z.t()
r2 = X2 - 2 * XZ + Z2.t()
return r2.clamp(min=0)
def forward(self, X, Z=None):
if Z is None:
Z = X
assert X.shape[1] == Z.shape[1]
base = -0.5 * self._square_scaled_dist(X, Z)
base += 2 * self.log_std
return base.clamp(min=-10, max=10).exp()
def kernelMix(k1, k2):
return k1 + k2
class Encoder(nn.Module):
def __init__(self, input_dim, z_dim, hidden_dim):
super(Encoder, self).__init__()
self.net = nn.Sequential(
nn.Linear(input_dim, hidden_dim),
nn.CELU(),
nn.Dropout(dropout_rate),
nn.Linear(hidden_dim, hidden_dim),
nn.CELU(),
nn.Dropout(dropout_rate),
nn.Linear(hidden_dim, z_dim)
)
def forward(self, x):
return self.net(x)
class DeepRBFKernel(nn.Module):
def __init__(self, input_dim, z_dim, hidden_dim, num_indv, n_induce, indv_dim):
super(DeepRBFKernel, self).__init__()
self.encoder = Encoder(input_dim, z_dim, hidden_dim)
self.rbf_indv = RBFKernel(indv_dim)
self.rbf_obsr = RBFKernel(z_dim)
self.indv_dim = indv_dim
self.num_indv = num_indv
self.indv_embedding = nn.Parameter(torch.rand([num_indv + n_induce, indv_dim]))
def forward(self, X, xid, Z=None, zid=None):
embed_X = self.encoder(X)
embed_Xid = self.indv_embedding[xid]
if Z is None:
k1 = self.rbf_indv(embed_Xid, embed_Xid)
k2 = self.rbf_obsr(embed_X, embed_X)
return kernelMix(k1, k2)
else:
embed_Zid = self.indv_embedding[zid + self.num_indv]
k1 = self.rbf_indv(embed_Xid, embed_Zid)
k2 = self.rbf_obsr(embed_X, Z)
return kernelMix(k1, k2)
def indv_kernel(self, xid):
embed_Xid = self.indv_embedding[xid]
return self.rbf_indv(embed_Xid, embed_Xid)
def rbf_direct(self, Z, zid):
embed_Zid = self.indv_embedding[zid + self.num_indv]
k1 = self.rbf_indv(embed_Zid, embed_Zid)
k2 = self.rbf_obsr(Z, Z)
return kernelMix(k1, k2)
def kernel_compute(X, xid, Z, zid, kernel):
kxz = kernel(X, xid, Z, zid)
kzz = kernel.rbf_direct(Z, zid)
kxx = kernel(X, xid)
kzz = kzz + torch.eye(len(kzz)).to(kzz.device) * factor
Lz = torch.cholesky(kzz)
inv_Lz = torch.inverse(Lz)
kzz_inv = inv_Lz.t() @ inv_Lz
first = kxz @ kzz_inv
return kxz, kzz, kxx, Lz, kzz_inv, first
def conditional_dist(kxz, kxx, first, u):
mean = first @ u # nxc
var = kxx - first @ kxz.t() # nxn
return mean, var
def solve_posterior(sgp):
zid = torch.arange(sgp.n_induce).type(torch.long)
beta = 1 / sgp.log_sigma.exp().pow(2)
x = trainX
y = trainy
xid = trainId
if args.cuda:
x = trainX.cuda()
y = trainy.cuda()
xid = trainId.cuda()
zid = zid.to('cuda')
kxz, kzz, kxx, Lz, kzz_inv, first = kernel_compute(x, xid, sgp.embed_Z, zid, sgp.deepkernel)
B = kzz + beta * kxz.T @ kxz
B_chol = torch.cholesky(B)
B_chol_inv = torch.inverse(B_chol)
B_inv = B_chol_inv.T @ B_chol_inv
A = kzz @ B_inv
u = beta * A @ kxz.T @ y
Sigma = A @ kzz
U = torch.cholesky(Sigma + torch.eye(len(Sigma)).to(Sigma.device) * factor)
P = (torch.ones([sgp.n_induce, sgp.n_induce]) + torch.eye(sgp.n_induce)).to(x.device)
P = torch.cholesky(P)
L = (U.diag() / P.diag()).diag()
return u, L
class SparseGPRegression(nn.Module):
def __init__(self, input_dim, z_dim, hidden_dim, n_induce, num_indv, indv_dim):
super(SparseGPRegression, self).__init__()
self.n_induce = n_induce
self.input_dim = input_dim
self.deepkernel = DeepRBFKernel(input_dim, z_dim, hidden_dim, num_indv, n_induce,indv_dim=indv_dim)
self.embed_Z = nn.Parameter(torch.rand([n_induce, z_dim]))
self.log_sigma = nn.Parameter(torch.zeros(1))
def forward(self, x, y, xid, u, L):
sigma2 = self.log_sigma.exp().pow(2)
zid = torch.arange(self.n_induce).type(torch.long).to(x.device)
kxz, kzz, kxx, Lz, kzz_inv, first = kernel_compute(x, xid, self.embed_Z, zid, self.deepkernel)
m = first @ u
v = first @ L
Sigma_q = L @ L.T
l1 = 1 / sigma2 * (y.T @ y - 2 * y.T @ m + m.T @ m + (v.T @ v).sum())
l2 = (Lz.diag().log().sum() - L.diag().log().sum()) * 2 + \
(kzz_inv @ Sigma_q).trace() + (u.T @ kzz_inv @ u)
return l1 / 2, l2 / 2
def predict_logit(self, x, xid, u):
zid = torch.arange(self.n_induce).type(torch.long).to(x.device)
kxz, kzz, kxx, Lz, kzz_inv, first = kernel_compute(x, xid, self.embed_Z, zid, self.deepkernel)
m, _ = conditional_dist(kxz, kxx, first, u.view(-1, 1))
return m.view(-1)
def train(args):
valid_r2_past = -1e10
count = 0
for epoch in range(args.epochs):
if epoch == 0:
sgp.eval()
with torch.no_grad():
u, L = solve_posterior(sgp)
sgp.train()
epoch_loss = 0.
preds = []
ys = []
for i, (x, y, xid, oid) in enumerate(train_loader):
ys.extend(y.numpy())
x = x.type(torch.float)
y = y.type(torch.float)
xid = xid.type(torch.long)
if args.cuda:
x = x.cuda()
y = y.cuda()
xid = xid.cuda()
optimizer.zero_grad()
l1, l2 = sgp(x, y, xid, u, L)
loss = l1 + l2
pred = sgp.predict_logit(x, xid, u)
preds.extend(pred.detach().cpu().numpy())
epoch_loss += loss.item()
loss.backward()
optimizer.step()
normalizer_train = len(train_loader.dataset)
total_epoch_loss_train = epoch_loss / normalizer_train
print("[epoch %03d] average training loss: %.4f" % (epoch, total_epoch_loss_train))
ys, preds = np.array(ys), np.array(preds)
print(f'epoch {epoch}: r2 -> {r2_score(ys, preds)}, rmse -> {np.sqrt(np.mean((ys - preds) ** 2))}')
u,L,valid_r2 = test(args,True,u,L)
if valid_r2 < valid_r2_past:
count += 1
else:
count = 0
valid_r2_past = valid_r2
if epoch % args.test_frequency == 0:
if epoch > 0:
u,L,final_test = test(args,False,None,None)
else:
u,L,final_test = test(args,False, u, L)
if count >= args.valid_dec_count:
break
return final_test
def test(args,isValid,u=None,L=None):
sgp.eval()
if u is None:
with torch.no_grad():
u, L = solve_posterior(sgp)
test_loss = 0.
preds = []
ys = []
if isValid:
loader = valid_loader
else:
loader = test_loader
for i, (x, y, xid, oid) in enumerate(loader):
ys.extend(y.numpy())
x = x.type(torch.float)
y = y.type(torch.float)
xid = xid.type(torch.long)
if args.cuda:
x = x.cuda()
y = y.cuda()
xid = xid.cuda()
with torch.no_grad():
l1, l2 = sgp(x, y, xid, u, L)
loss = l1 + l2
test_loss += loss.item()
pred = sgp.predict_logit(x, xid, u)
preds.extend(pred.cpu().numpy())
# report test diagnostics
normalizer_test = len(loader.dataset)
total_epoch_loss_test = test_loss / normalizer_test
print(f"\t average {'valid' if isValid else 'test'} loss: %.4f" % (total_epoch_loss_test))
r2 = r2_score(ys, preds)
print(f"\t {'valid' if isValid else 'test'} r2: {r2}")
return u,L,r2
def arge_parser():
fn = 'tadpole_small'
parser = argparse.ArgumentParser(description='L-DKGPR')
# parser.add_argument('--file', type=str, help='location of the input file, should be a .mat file',default='..//cluster3_1.mat')
parser.add_argument('--file', type=str, help='location of the input file, should be a .mat file',default=f'data/{fn}.mat')
parser.add_argument('--batch_size', type=int, help='minibatch size for training, default 1024',
default=1024)
parser.add_argument('--dropout', type=float, help='dropout rate, default 0.2', default=0.2)
parser.add_argument('--D', type=int, help='latent dimensions, default 10',default=10)
parser.add_argument('--hidden_dim',type=int, help='number of hidden units, default 16', default=16)
parser.add_argument('--factor',type=float,help='the value added to diagonal of correlation matrix to avoid singularity, default 0.0001; If singular problem exists, please choose a larger value.', default=1e-2)
parser.add_argument('--M', type=int, help='the number of inducing points, default 10', default=10)
parser.add_argument('--lr', type=float,
help='learning rate for Theta - phi, default 0.001. Decrease lr if bad performance is achieved', default=1e-3)
parser.add_argument('--lr_indv_embedding', type=float,
help='learning rate for phi, default 0.01. Decrease lr if bad performance is achieved', default=1e-1)
parser.add_argument('--epochs', type=int,
help='learning epochs for training, default 300', default=300)
parser.add_argument('--cuda', dest='cuda',
help='use gpu', action='store_true')
parser.add_argument('--cpu', dest='cuda', help='use cpu only', action='store_false')
parser.add_argument('--test_frequency', type=int,
help='output the test loss in every {test_frequency} epochs, default 5', default=5)
parser.add_argument('--save_path', type=str,
help="the path to the saved model after training.", default=f'saveModels/{fn}')
parser.add_argument('--load_path', type=str,
help="load a saved model If specified, then no training is performed. default 'None'",
# default='saveModels/lastRun')
default = 'None')
parser.add_argument('--seed', type=int, help='random seed, default 0', default=0)
parser.add_argument('--valid_dec_count', type=int, help='tolerance for time of decreasing r2 on validation set', default=10)
parser.add_argument('--number_cluster',type=int,help='number of cluster in correlation plot. Only valid in real-life data. default 2.',default=3)
args = parser.parse_args()
return args
def visualizeCorr(sgp,args):
sgp.cpu()
if args.file.split('/')[-2] == 'simulation':
final_corr = data['corr']
allX = torch.tensor(data['data']).type(torch.float)
allIid = data['iid'].reshape(-1)
plt.figure()
sns.heatmap(
final_corr,
cmap="YlGnBu",
square=True,
robust=True,
xticklabels=False,
yticklabels=False,
)
corr = sgp.deepkernel(allX, allIid).detach().cpu().numpy()
plt.figure()
sns.heatmap(
corr,
cmap='YlGnBu',
square=True,
robust=True,
xticklabels=False,
yticklabels=False,
)
plt.show()
else:
from sklearn.cluster import SpectralCoclustering
indv_corr = sgp.deepkernel.indv_kernel(torch.arange(len(idMap))).detach().cpu().numpy()
num_c = args.number_cluster
model = SpectralCoclustering(n_clusters=num_c, random_state=0)
model.fit(indv_corr)
fit_data = indv_corr[np.argsort(model.row_labels_)]
fit_data = fit_data[:, np.argsort(model.row_labels_)]
rows = np.random.permutation(np.arange(len(fit_data)))
rows = rows[:3300]
rows = np.sort(rows)
clusterRes = model.row_labels_
cl = np.argsort(clusterRes)
ax = sns.heatmap(
indv_corr[cl][:, cl],
cmap='YlGnBu',
square=True,
robust=True,
xticklabels=False,
yticklabels=False,
)
plt.show()
if __name__ == '__main__':
args = arge_parser()
file = args.file
factor = args.factor
dropout_rate = args.dropout
# the mat file should contains the following fields:
# trainId, trainOid, trainX, trainY; testId, testOid, testX, testY;
data = io.loadmat(file)
trainX = torch.from_numpy(data['trainX']).type(torch.float)
trainy = torch.from_numpy(data['trainY'].reshape(-1)).type(torch.float)
testX = torch.from_numpy(data['testX']).type(torch.float)
testy = torch.from_numpy(data['testY'].reshape(-1)).type(torch.float)
trainId_Ori = data['trainId'].reshape(-1)
testId_Ori = data['testId'].reshape(-1)
trainOid_Ori = data['trainOid'].reshape(-1)
testOid_Ori = data['testOid'].reshape(-1)
allX = torch.cat([trainX, testX], dim=0)
ally = torch.cat([trainy, testy], dim=0)
ids = set(list(np.concatenate([trainId_Ori, testId_Ori])))
oids = set(list(np.concatenate([trainOid_Ori, testOid_Ori])))
idMap = {}
i = 0
for x in ids:
idMap[x] = i
i += 1
trainId = torch.FloatTensor([idMap[x] for x in trainId_Ori]).type(torch.long)
testId = torch.FloatTensor([idMap[x] for x in testId_Ori]).type(torch.long)
minOid = np.min(list(oids))
trainOid = torch.FloatTensor(trainOid_Ori - minOid)
testOid = torch.FloatTensor(testOid_Ori - minOid)
allIid = torch.cat([trainId, testId])
allmean = ally.mean()
trainy -= allmean
testy -= allmean
# split train into train and valid
tr_idx = np.arange(len(trainX))
train_idx, valid_idx = train_test_split(tr_idx,test_size=int(2/7 * len(trainX)), random_state=args.seed)
train_ds = MyDataset(trainX[train_idx],trainy[train_idx],trainId[train_idx],trainOid[train_idx])
valid_ds = MyDataset(trainX[valid_idx],trainy[valid_idx],trainId[valid_idx],trainOid[valid_idx])
test_ds = MyDataset(testX, testy, testId, testOid)
train_loader = DataLoader(train_ds, batch_size=args.batch_size, shuffle=True)
valid_loader = DataLoader(valid_ds, batch_size=args.batch_size, shuffle=False)
test_loader = DataLoader(test_ds, batch_size=args.batch_size, shuffle=False)
sgp = SparseGPRegression(trainX.shape[1], args.D, args.hidden_dim, args.M, len(ids),args.D)
params = []
for n, p in sgp.named_parameters():
if n.startswith('embed_Z') or n.startswith('deepkernel.indv_embedding'):
params.append({'name': n, 'params': p, 'lr': args.lr_indv_embedding})
else:
params.append({'name': n, 'params': p, 'lr': args.lr})
if args.cuda:
sgp = sgp.to('cuda')
optimizer = Adam(params, lr=args.lr)
if args.load_path is 'None':
r2 = train(args)
if args.save_path is not 'None':
torch.save(sgp.state_dict(), args.save_path)
print(f'the final test r2 {r2}')
else:
sgp.load_state_dict(torch.load(args.load_path))
visualizeCorr(sgp, args)