-
Notifications
You must be signed in to change notification settings - Fork 0
/
seesfuture.py
432 lines (364 loc) · 16.5 KB
/
seesfuture.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
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import random
import time
from scipy.io import loadmat
import os
from encoder import *
from decoder import *
from utils import *
def epoch_time(start_time, end_time):
elapsed_time = end_time - start_time
elapsed_mins = int(elapsed_time / 60)
elapsed_secs = int(elapsed_time - (elapsed_mins * 60))
return elapsed_mins, elapsed_secs
class Sees_model(nn.Module):
def __init__(self, device='cpu', model_path='seesfuture.model'):
super(Sees_model, self).__init__()
print('Using device: {}'.format(device))
self.model_path = model_path
self.device = device
self.n_node = -1
self.support = None
self.train_ = [None, None]
self.val_ = [None, None]
self.test_ = [None, None]
def INIT(self, data_path):
self.read_data(data_path)
self.encoder = seesEncoderWithGraphAttention(INPUT_DIM,
HID_DIM,
N_LAYERS,
N_HEADS,
PF_DIM,
SEQ_LEN,
WALK_LEN,
self.n_node,
self.get_device())
self.decoder = seesDecoderWithMultiOut(HID_DIM,
OUT_DIM,
SEQ_LEN,
self.get_device())
self.optimizer = torch.optim.Adam(self.parameters(), lr=LEARNING_RATE)
def get_device(self):
if self.device == 'gpu':
return torch.device('cuda')
else:
return torch.device('cpu')
def read_data(self, path):
train_path = path+'/train.npz'
val_path = path+'/val.npz'
test_path = path+'/test.npz'
net_path = path+'/net.mat'
net_path2 = path+'/N.npy'
if os.path.exists(net_path):
net = loadmat(net_path)
support = net['S']
else:
support = np.load(net_path2)
N_NODE = support.shape[0]
self.n_node = support.shape[0]
support[range(N_NODE),range(N_NODE)] = np.ones(N_NODE)
self.support = support
train_npz = np.load(train_path)
train_x_ = train_npz['x']
train_y_ = train_npz['y']
train_y_ = np.split(train_y_, 2, axis=-1)[0]
# train_x_, (data_len, seq_len, n_node, input_dim)
# train_y_, (data_len, seq_len, n_node, out_dim)
train_x_ = train_x_.transpose(2,0,1,3)
train_y_ = train_y_.transpose(2,0,1,3)
#print(train_x_[0][0].shape)
# train_x_, (n_node, data_len, seq_len, input_dim)
# train_y_, (n_node, data_len, seq_len, out_dim)
val_npz = np.load(val_path)
val_x_ = val_npz['x']
val_y_ = val_npz['y']
val_y_ = np.split(val_y_, 2, axis=-1)[0]
val_x_ = val_x_.transpose(2,0,1,3)
val_y_ = val_y_.transpose(2,0,1,3)
test_npz = np.load(test_path)
test_x_ = test_npz['x']
test_y_ = test_npz['y']
test_y_ = np.split(test_y_, 2, axis=-1)[0]
test_x_ = test_x_.transpose(2,0,1,3)
test_y_ = test_y_.transpose(2,0,1,3)
self.train_ = [train_x_, train_y_]
self.val_ = [val_x_, val_y_]
self.test_ = [test_x_, test_y_]
print('Training size: {}'.format(train_x_.shape[1]))
print('Validating size: {}'.format(val_x_.shape[1]))
print('Testing size: {}'.format(test_x_.shape[1]))
def gen_walk(self, start_idx, walk_len):
walk = [start_idx]
for i in range(walk_len):
neigh_weight = self.support[walk[-1]]
neigh_nonzero = neigh_weight.nonzero()[0]
sum_neigh = sum([float(neigh_weight[_]) for _ in neigh_nonzero])
neigh_prob = [neigh_weight[_]/sum_neigh for _ in neigh_nonzero]
r = random.random()
for j in range(len(neigh_nonzero)):
if r < sum(neigh_prob[:j+1]):
walk.append(neigh_nonzero[j])
break
return np.array(walk[1:])
def get_neigh(self, src_node, walk_len):
neighs = []
neigh_nonzero = self.support[src_node].nonzero()[0]
for node in neigh_nonzero:
neighs.append(node)
while len(neighs) < walk_len:
neighs.append(self.n_node)
return np.array(neighs[:walk_len])
def get_batch(self, pdata):
curr_idx = 0
c_i, c_j = pdata[0].shape[0], pdata[0].shape[1]
while curr_idx < c_i*c_j:
start_idx = curr_idx
end_idx = min(curr_idx + BATCH_SIZE, c_i*c_j)
src = []
src_mask = []
trg = []
walks = []
walk_src = []
self_node = []
for i in range(end_idx-start_idx):
s_i = int((start_idx+i)/c_j) # idx of node
s_j = (start_idx+i)%c_j # idx of data seq
src.append(pdata[0][s_i][s_j])
trg_full = pdata[1][s_i][s_j]
trg_first = np.split(trg_full, SEQ_LEN, axis=0)[0]
trg.append(trg_full)
src_mask.append([1 for _ in range(SEQ_LEN)])
#walk = self.gen_walk(s_i, WALK_LEN)
walk = self.get_neigh(s_i, WALK_LEN)
walks.append(walk)
walk_src.append([])
walk_src[i] = []
for w_idx in walk:
if w_idx == self.n_node:
walk_src[i].append([[0.0, 0.0] for _ in range(SEQ_LEN)])
else:
walk_src[i].append(pdata[0][w_idx][s_j])
#walk_src[i] = [self.pdata[w_idx][s_j] for w_idx in walk]
self_node.append(s_i)
curr_idx = end_idx
yield (torch.tensor(src, dtype=torch.float, device=self.get_device()),
torch.tensor(src_mask, dtype=torch.int64, device=self.get_device()),
torch.tensor(trg, dtype=torch.float, device=self.get_device()),
torch.tensor(walks, dtype=torch.int64, device=self.get_device()),
torch.tensor(walk_src, dtype=torch.float, device=self.get_device()),
torch.tensor(self_node, dtype=torch.int64, device=self.get_device()))
def get_train_batch(self):
curr_idx = 0
c_i, c_j = self.train_[0].shape[0], self.train_[0].shape[1]
while curr_idx < c_i*c_j:
start_idx = curr_idx
end_idx = min(curr_idx + BATCH_SIZE, c_i*c_j)
src = []
src_mask = []
trg = []
walks = []
walk_src = []
self_node = []
for i in range(end_idx-start_idx):
s_i = int((start_idx+i)/c_j) # idx of node
s_j = (start_idx+i)%c_j # idx of data seq
src.append(self.train_[0][s_i][s_j])
trg_full = self.train_[1][s_i][s_j]
trg_first = np.split(trg_full, SEQ_LEN, axis=0)[0]
trg.append(trg_first)
src_mask.append([1 for _ in range(SEQ_LEN)])
#walk = self.gen_walk(s_i, WALK_LEN)
walk = self.get_neigh(s_i, WALK_LEN)
walks.append(walk)
walk_src.append([])
walk_src[i] = [self.train_[0][w_idx][s_j] for w_idx in walk]
self_node.append(s_i)
curr_idx = end_idx
yield (torch.tensor(src, dtype=torch.float, device=self.get_device()),
torch.tensor(src_mask, dtype=torch.int64, device=self.get_device()),
torch.tensor(trg, dtype=torch.float, device=self.get_device()),
torch.tensor(walks, dtype=torch.int64, device=self.get_device()),
torch.tensor(walk_src, dtype=torch.float, device=self.get_device()),
torch.tensor(self_node, dtype=torch.int64, device=self.get_device()))
def get_val_batch(self):
curr_idx = 0
c_i, c_j = self.val_[0].shape[0], self.val_[0].shape[1]
while curr_idx < c_i*c_j:
start_idx = curr_idx
end_idx = min(curr_idx + BATCH_SIZE, c_i*c_j)
src = []
src_mask = []
trg = []
walks = []
walk_src = []
self_node = []
for i in range(end_idx-start_idx):
s_i = int((start_idx+i)/c_j) # idx of node
s_j = (start_idx+i)%c_j # idx of data seq
src.append(self.val_[0][s_i][s_j])
trg_full = self.val_[1][s_i][s_j]
trg_first = np.split(trg_full, SEQ_LEN, axis=0)[0]
trg.append(trg_first)
src_mask.append([1 for _ in range(SEQ_LEN)])
#walk = self.gen_walk(s_i, WALK_LEN)
walk = self.get_neigh(s_i, WALK_LEN)
walks.append(walk)
walk_src.append([])
walk_src[i] = [self.val_[0][w_idx][s_j] for w_idx in walk]
self_node.append(s_i)
curr_idx = end_idx
yield (torch.tensor(src, dtype=torch.float, device=self.get_device()),
torch.tensor(src_mask, dtype=torch.int64, device=self.get_device()),
torch.tensor(trg, dtype=torch.float, device=self.get_device()),
torch.tensor(walks, dtype=torch.int64, device=self.get_device()),
torch.tensor(walk_src, dtype=torch.float, device=self.get_device()),
torch.tensor(self_node, dtype=torch.int64, device=self.get_device()))
def get_test_batch(self):
curr_idx = 0
c_i, c_j = self.test_[0].shape[0], self.test_[0].shape[1]
while curr_idx < c_i*c_j:
start_idx = curr_idx
end_idx = min(curr_idx + BATCH_SIZE, c_i*c_j)
src = []
src_mask = []
trg = []
walks = []
walk_src = []
self_node = []
for i in range(end_idx-start_idx):
s_i = int((start_idx+i)/c_j) # idx of node
s_j = (start_idx+i)%c_j # idx of data seq
src.append(self.test_[0][s_i][s_j])
trg_full = self.test_[1][s_i][s_j]
trg_first = np.split(trg_full, SEQ_LEN, axis=0)[0]
trg.append(trg_first)
src_mask.append([1 for _ in range(SEQ_LEN)])
#walk = self.gen_walk(s_i, WALK_LEN)
walk = self.get_neigh(s_i, WALK_LEN)
walks.append(walk)
walk_src.append([])
walk_src[i] = [self.test_[0][w_idx][s_j] for w_idx in walk]
self_node.append(s_i)
curr_idx = end_idx
yield (torch.tensor(src, dtype=torch.float, device=self.get_device()),
torch.tensor(src_mask, dtype=torch.int64, device=self.get_device()),
torch.tensor(trg, dtype=torch.float, device=self.get_device()),
torch.tensor(walks, dtype=torch.int64, device=self.get_device()),
torch.tensor(walk_src, dtype=torch.float, device=self.get_device()),
torch.tensor(self_node, dtype=torch.int64, device=self.get_device()))
def forward(self, src, src_mask, walk, walk_src, self_node):
src, src_mask = self.encoder(src, src_mask, walk, walk_src, self_node)
src = self.decoder(src)
return src
def calculate_loss(self, this_batch):
src, src_mask, trg, walk, walk_src, self_node = this_batch
src = self.forward(src, src_mask, walk, walk_src, self_node)
loss_func = nn.MSELoss()
src = src.view(src.shape[0], SEQ_LEN)
trg = trg.view(trg.shape[0], SEQ_LEN)
mse = loss_func(src, trg)
return torch.sqrt(mse)
def run_step(self, this_batch):
self.optimizer.zero_grad()
loss = self.calculate_loss(this_batch)
loss.backward()
torch.nn.utils.clip_grad_norm_(self.parameters(), CLIP)
self.optimizer.step()
return float(loss.cpu())
def run_training(self):
best_loss = float('inf')
start_time = time.time()
ready_to_stop = 0
for epoch in range(MAX_EPOCHES):
if ready_to_stop > PATIENCE:
print('Terminating... No more improvement...')
break
self.train()
batch_data = self.get_batch(self.train_)
train_loss = 0
batch_num = 0
for this_batch in batch_data:
batch_num += 1
train_loss += self.run_step(this_batch)
train_loss /= batch_num
self.eval()
batch_data = self.get_batch(self.val_)
val_loss = 0
batch_num = 0
for this_batch in batch_data:
batch_num += 1
val_loss += self.calculate_loss(this_batch)
val_loss /= batch_num
end_time = time.time()
epoch_mins, epoch_secs = epoch_time(start_time, end_time)
if val_loss <= best_loss:
best_loss = val_loss
torch.save(self, self.model_path)
ready_to_stop = 0
else:
ready_to_stop += 1
print('Training...')
print('Epoch: {} ---- Time: {}m {}s'.format(epoch, epoch_mins, epoch_secs))
print('Training loss: {}'.format(train_loss))
print('Validating loss: {}'.format(val_loss))
print('------------------------------------------')
# del
def get_test_loss(self, this_batch):
src, src_mask, trg, walk, walk_src, self_node = this_batch
src = self.forward(src, src_mask, walk, walk_src, self_node)
loss_func = nn.MSELoss()
src = torch.split(src, 1, dim=1)
trg = torch.split(trg, 1, dim=1)
mse = []
for i in range(SEQ_LEN):
mse.append(torch.sqrt(loss_func(src[i].view(src[i].shape[0],1), trg[i].view(trg[i].shape[0],1))))
return [_.cpu() for _ in mse]
def run_testing(self):
self = torch.load(self.model_path)
start_time = time.time()
self.eval()
batch_data = self.get_batch(self.test_)
test_loss = [0.0 for _ in range(SEQ_LEN)]
batch_num = 0
for this_batch in batch_data:
batch_num += 1
test_losses = self.get_test_loss(this_batch)
for i in range(SEQ_LEN):
test_loss[i] += float(test_losses[i])
#mse_loss = self.get_test_loss(this_batch)
#test_loss = [test_loss[i] + mse_loss[i] for i in range(SEQ_LEN)]
#test_loss = [test_loss[i]/batch_num for i in range(SEQ_LEN)]
test_loss = [_/batch_num for _ in test_loss]
end_time = time.time()
epoch_mins, epoch_secs = epoch_time(start_time, end_time)
print('Testing...')
print('Time: {}m {}s'.format(epoch_mins, epoch_secs))
#print('Testing loss: {}'.format(test_loss))
for ts in range(SEQ_LEN):
print('Testing loss on SEQ {}: {}'.format(ts,test_loss[ts]))
print('-------------------------------------------------')
if __name__ == '__main__':
data_path = 'traffic'
os.environ["CUDA_VISIBLE_DEVICES"] = "10"
random.seed(SEED)
np.random.seed(SEED)
torch.manual_seed(SEED)
torch.cuda.manual_seed(SEED)
torch.backends.cudnn.deterministic = True
if torch.cuda.is_available():
model = Sees_model('gpu')
model.INIT(data_path)
model.cuda()
for c in model.children():
c.cuda()
else:
model = Sees_model('cpu')
model.INIT(data_path)
model.cpu()
for c in model.children():
c.cpu()
model.run_training()
model.run_testing()