-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathrewt_generic.py
411 lines (333 loc) · 15.5 KB
/
rewt_generic.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
import torch
import sys
import numpy as np
from logistic_regression import *
from deep_net import *
from weighted_cage import *
from sklearn.feature_extraction.text import TfidfVectorizer
from losses import *
import pickle
from torch.utils.data import TensorDataset, DataLoader
# CUDA for PyTorch
use_cuda = torch.cuda.is_available()
device = torch.device("cuda:0" if use_cuda else "cpu")
torch.backends.cudnn.benchmark = True
torch.set_default_dtype(torch.float64)
torch.set_printoptions(threshold=20)
objs = []
dset_directory = sys.argv[10]
n_classes = int(sys.argv[11])
feat_model = sys.argv[12]
qg_available = int(sys.argv[13])
batch_size = int(sys.argv[14])
lr_fnetwork = float(sys.argv[15])
lr_gm = float(sys.argv[16])
name_dset = dset_directory.split("/")[1].lower()
if name_dset =='youtube' or name_dset=='census':
from sklearn.metrics import accuracy_score as score
else:
from sklearn.metrics import f1_score as score
with open(dset_directory + '/d_processed.p', 'rb') as f:
while 1:
try:
o = pickle.load(f)
except EOFError:
break
objs.append(o)
x_supervised = torch.tensor(objs[0]).double()
y_supervised = torch.tensor(objs[3]).long()
l_supervised = torch.tensor(objs[2]).long()
s_supervised = torch.tensor(objs[2]).double()
objs = []
with open(dset_directory + '/U_processed.p', 'rb') as f:
while 1:
try:
o = pickle.load(f)
except EOFError:
break
objs.append(o)
excl= []
idx=0
for x in objs[1]:
if(all(x==int(n_classes))):
excl.append(idx)
idx+=1
print('no of excludings are ', len(excl))
x_unsupervised = torch.tensor(np.delete(objs[0],excl, axis=0)).double()
y_unsupervised = torch.tensor(np.delete(objs[3],excl, axis=0)).long()
l_unsupervised = torch.tensor(np.delete(objs[2],excl, axis=0)).long()
s_unsupervised = torch.tensor(np.delete(objs[2],excl, axis=0)).double()
print('Length of U is', len(x_unsupervised))
objs = []
with open(dset_directory + '/validation_processed.p', 'rb') as f:
while 1:
try:
o = pickle.load(f)
except EOFError:
break
objs.append(o)
x_valid = torch.tensor(objs[0]).double()
y_valid = objs[3]
l_valid = torch.tensor(objs[2]).long()
s_valid = torch.tensor(objs[2]).double()
objs1 = []
with open(dset_directory + '/test_processed.p', 'rb') as f:
while 1:
try:
o = pickle.load(f)
except EOFError:
break
objs1.append(o)
x_test = torch.tensor(objs1[0]).double()
y_test = objs1[3]
l_test = torch.tensor(objs1[2]).long()
s_test = torch.tensor(objs1[2]).double()
n_features = x_supervised.shape[1]
# Labeling Function Classes
# k = torch.from_numpy(np.array([0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 1, 0, 0, 1, 0, 0, 0, 1, 1, 1, 1, 0, 0, 1, 1, 0, 0, 0, 0, 1, 0, 0])).long()
#lf_classes_file = sys.argv[11]
k = torch.from_numpy(np.load(dset_directory + '/k.npy')).long()
n_lfs = len(k)
print('LFs are ',k)
print('no of lfs are ', n_lfs)
# a = torch.ones(n_lfs).double() * 0.9
# print('before ',a)
if qg_available:
a = torch.from_numpy(np.load(dset_directory+'/prec.npy')).double()
else:
# a = torch.ones(n_lfs).double() * 0.9
prec_lfs=[]
for i in range(n_lfs):
correct = 0
for j in range(len(y_valid)):
if y_valid[j] == l_valid[j][i]:
correct+=1
prec_lfs.append(correct/len(y_valid))
a = torch.tensor(prec_lfs).double()
# n_lfs = int(len(k))
# print('number of lfs ', n_lfs)
# a = torch.ones(n_lfs).double() * 0.9
continuous_mask = torch.zeros(n_lfs).double()
for i in range(s_supervised.shape[0]):
for j in range(s_supervised.shape[1]):
if s_supervised[i, j].item() > 0.999:
s_supervised[i, j] = 0.999
if s_supervised[i, j].item() < 0.001:
s_supervised[i, j] = 0.001
for i in range(s_unsupervised.shape[0]):
for j in range(s_unsupervised.shape[1]):
if s_unsupervised[i, j].item() > 0.999:
s_unsupervised[i, j] = 0.999
if s_unsupervised[i, j].item() < 0.001:
s_unsupervised[i, j] = 0.001
for i in range(s_valid.shape[0]):
for j in range(s_valid.shape[1]):
if s_valid[i, j].item() > 0.999:
s_valid[i, j] = 0.999
if s_valid[i, j].item() < 0.001:
s_valid[i, j] = 0.001
for i in range(s_test.shape[0]):
for j in range(s_test.shape[1]):
if s_test[i, j].item() > 0.999:
s_test[i, j] = 0.999
if s_test[i, j].item() < 0.001:
s_test[i, j] = 0.001
l = torch.cat([l_supervised, l_unsupervised])
s = torch.cat([s_supervised, s_unsupervised])
x_train = torch.cat([x_supervised, x_unsupervised])
y_train = torch.cat([y_supervised, y_unsupervised])
supervised_mask = torch.cat([torch.ones(l_supervised.shape[0]), torch.zeros(l_unsupervised.shape[0])])
## Quality Guides ##
## End Quality Quides##
# a = torch.tensor(np.load(dset_directory + '/precision_values.npy'))
# print('after ',a)
#Setting |validation|=|supevised|
x_valid = x_valid[0:len(x_supervised)]
y_valid = y_valid[0:len(x_supervised)]
s_valid = s_valid[0:len(x_supervised)]
l_valid = l_valid[0:len(x_supervised)]
# print(l_valid.shape)
# print(l_valid[0])
num_runs = int(sys.argv[9])
final_score_gm, final_score_lr, final_score_gm_val, final_score_lr_val = [],[],[],[]
for lo in range(0,num_runs):
pi = torch.ones((n_classes, n_lfs)).double()
pi.requires_grad = True
theta = torch.ones((n_classes, n_lfs)).double() * 1
theta.requires_grad = True
pi_y = torch.ones(n_classes).double()
pi_y.requires_grad = True
weights = torch.ones(k.shape[0])
if feat_model == 'lr':
lr_model = LogisticRegression(n_features, n_classes)
elif feat_model =='nn':
n_hidden = 512
lr_model = DeepNet(n_features, n_hidden, n_classes)
else:
print('Please provide feature based model : lr or nn')
exit()
optimizer = torch.optim.Adam([{"params": lr_model.parameters()}, {"params": [pi, pi_y, theta]}], lr=0.001)
optimizer_lr = torch.optim.Adam(lr_model.parameters(), lr=lr_fnetwork)
optimizer_gm = torch.optim.Adam([theta, pi, pi_y], lr=lr_gm, weight_decay=0)
# optimizer = torch.optim.Adam([theta, pi, pi_y], lr=0.01, weight_decay=0)
supervised_criterion = torch.nn.CrossEntropyLoss()
dataset = TensorDataset(x_train, y_train, l, s, supervised_mask)
loader = DataLoader(dataset, batch_size=batch_size, shuffle=True,pin_memory=True)
save_folder = sys.argv[1]
print('num runs are ', sys.argv[1], num_runs)
best_score_lr,best_score_gm,best_epoch_lr,best_epoch_gm,best_score_lr_val, best_score_gm_val = 0,0,0,0,0,0
stop_pahle, stop_pahle_gm = [], []
for epoch in range(100):
lr_model.train()
for batch_ndx, sample in enumerate(loader):
optimizer_lr.zero_grad()
optimizer_gm.zero_grad()
# Start entropy code
# unsupervised_indices = (1-sample[4]).nonzero().squeeze()
unsup = []
sup = []
supervised_indices = sample[4].nonzero().view(-1)
# unsupervised_indices = indices ## Uncomment for entropy
unsupervised_indices = (1-sample[4]).nonzero().squeeze()
if(sys.argv[2] =='l1'):
if len(supervised_indices) > 0:
loss_1 = supervised_criterion(lr_model(sample[0][supervised_indices]), sample[1][supervised_indices])
else:
loss_1 = 0
else:
loss_1=0
if(sys.argv[3] =='l2'):
unsupervised_lr_probability = torch.nn.Softmax()(lr_model(sample[0][unsupervised_indices]))
loss_2 = entropy(unsupervised_lr_probability)
else:
loss_2=0
if(sys.argv[4] =='l3'):
y_pred_unsupervised = np.argmax(probability(theta, pi_y, pi, sample[2][unsupervised_indices], sample[3][unsupervised_indices], k, n_classes,
continuous_mask, weights).detach().numpy(), 1)
loss_3 = supervised_criterion(lr_model(sample[0][unsupervised_indices]), torch.tensor(y_pred_unsupervised))
else:
loss_3 = 0
if (sys.argv[5] == 'l4' and len(supervised_indices) > 0):
loss_4 = log_likelihood_loss_supervised(theta, pi_y, pi, sample[1][supervised_indices], sample[2][supervised_indices],\
sample[3][supervised_indices], k, n_classes, continuous_mask, weights)
else:
loss_4 = 0
if(sys.argv[6] =='l5'):
loss_5 = log_likelihood_loss(theta, pi_y, pi, sample[2][unsupervised_indices], \
sample[3][unsupervised_indices], k, n_classes, continuous_mask, weights)
else:
loss_5 =0
if(sys.argv[7] =='l6'):
if(len(supervised_indices) >0):
supervised_indices = supervised_indices.tolist()
probs_graphical = probability(theta, pi_y, pi, torch.cat([sample[2][unsupervised_indices], \
sample[2][supervised_indices]]), torch.cat([sample[3][unsupervised_indices],\
sample[3][supervised_indices]]), k, n_classes, continuous_mask, weights)
else:
probs_graphical = probability(theta, pi_y, pi,sample[2][unsupervised_indices],\
sample[3][unsupervised_indices],k, n_classes, continuous_mask, weights)
probs_graphical = (probs_graphical.t() / probs_graphical.sum(1)).t()
probs_lr = torch.nn.Softmax()(lr_model(sample[0]))
loss_6 = kl_divergence(probs_lr, probs_graphical)
# loss_6 = kl_divergence(probs_graphical, probs_lr) #original version
else:
loss_6= 0
# loss_6 = - torch.log(1 - probs_graphical * (1 - probs_lr)).sum(1).mean()
if(sys.argv[8] =='qg'):
prec_loss = precision_loss(theta, k, n_classes, a)
else:
prec_loss =0
loss = loss_1 + loss_2 + loss_3 + loss_4 + loss_6+loss_5 + prec_loss
# print('loss is',loss_1, loss_2, loss_3, loss_4, loss_5, loss_6, prec_loss)
if loss != 0:
loss.backward()
optimizer_gm.step()
optimizer_lr.step()
y_pred = np.argmax(probability(theta, pi_y, pi, l_test, s_test, k, n_classes, \
continuous_mask ,weights).detach().numpy(), 1)
gm_acc = score(y_test, y_pred)
#Valid
y_pred = np.argmax(probability(theta, pi_y, pi, l_valid, s_valid, k, n_classes, \
continuous_mask, weights).detach().numpy(), 1)
gm_valid_acc = score(y_valid, y_pred)
#LR Test
probs = torch.nn.Softmax()(lr_model(x_test))
y_pred = np.argmax(probs.detach().numpy(), 1)
lr_acc =score(y_test, y_pred)
#LR Valid
probs = torch.nn.Softmax()(lr_model(x_valid))
y_pred = np.argmax(probs.detach().numpy(), 1)
lr_valid_acc = score(y_valid, y_pred)
print("Epoch: {}\t Test GM accuracy_score: {}".format(epoch, gm_acc ))
print("Epoch: {}\tGM accuracy_score(Valid): {}".format(epoch, gm_valid_acc))
print("Epoch: {}\tTest LR accuracy_score: {}".format(epoch, lr_acc ))
print("Epoch: {}\tLR accuracy_score(Valid): {}".format(epoch, lr_valid_acc))
if gm_valid_acc >= best_score_gm_val and gm_valid_acc >= best_score_lr_val:
# print("Inside Best hu Epoch: {}\t Test GM accuracy_score: {}".format(epoch, gm_acc ))
# print("Inside Best hu Epoch: {}\tGM accuracy_score(Valid): {}".format(epoch, gm_valid_acc))
if gm_valid_acc == best_score_gm_val or gm_valid_acc == best_score_lr_val:
if best_score_gm < gm_acc or best_score_lr < lr_acc:
best_epoch_gm = epoch
best_score_gm_val = gm_valid_acc
best_score_gm = gm_acc
best_epoch_lr = epoch
best_score_lr_val = lr_valid_acc
best_score_lr = lr_acc
else:
best_epoch_gm = epoch
best_score_gm_val = gm_valid_acc
best_score_gm = gm_acc
best_epoch_lr = epoch
best_score_lr_val = lr_valid_acc
best_score_lr = lr_acc
stop_pahle = []
stop_pahle_gm = []
checkpoint = {'theta': theta,'pi': pi}
# torch.save(checkpoint, save_folder+"/gm_"+str(epoch) +".pt")
checkpoint = {'params': lr_model.state_dict()}
# torch.save(checkpoint, save_folder+"/lr_"+ str(epoch)+".pt")
if lr_valid_acc >= best_score_lr_val and lr_valid_acc >= best_score_gm_val:
# print("Inside Best hu Epoch: {}\tTest LR accuracy_score: {}".format(epoch, lr_acc ))
# print("Inside Best hu Epoch: {}\tLR accuracy_score(Valid): {}".format(epoch, lr_valid_acc))
if lr_valid_acc == best_score_lr_val or lr_valid_acc == best_score_gm_val:
if best_score_lr < lr_acc or best_score_gm < gm_acc:
best_epoch_lr = epoch
best_score_lr_val = lr_valid_acc
best_score_lr = lr_acc
best_epoch_gm = epoch
best_score_gm_val = gm_valid_acc
best_score_gm = gm_acc
else:
best_epoch_lr = epoch
best_score_lr_val = lr_valid_acc
best_score_lr = lr_acc
best_epoch_gm = epoch
best_score_gm_val = gm_valid_acc
best_score_gm = gm_acc
stop_pahle = []
stop_pahle_gm = []
checkpoint = {'theta': theta,'pi': pi}
# torch.save(checkpoint, save_folder+"/gm_"+str(epoch) +".pt")
checkpoint = {'params': lr_model.state_dict()}
# torch.save(checkpoint, save_folder+"/lr_"+ str(epoch)+".pt")
if len(stop_pahle) > 20 and len(stop_pahle_gm) > 20 and (all(best_score_lr_val >= k for k in stop_pahle) or \
all(best_score_gm_val >= k for k in stop_pahle_gm)):
print('Early Stopping at', best_epoch_gm, best_score_gm, best_score_lr)
print('Validation score Early Stopping at', best_epoch_gm, best_score_lr_val, best_score_gm_val)
break
else:
# print('inside else stop pahle epoch', epoch)
stop_pahle.append(lr_valid_acc)
stop_pahle_gm.append(gm_valid_acc)
# print("Run \t",lo, "Epoch Gm, Epoch LR, GM, LR \t", best_epoch_gm, best_epoch_lr,best_score_gm, best_score_lr)
# print("Run \t",lo, "GM Val, LR Val \t", best_score_gm_val, best_score_lr_val)
print("Run \t",lo, "Epoch, GM, LR \t",best_epoch_lr, best_score_gm, best_score_lr)
print("Run \t",lo, "GM Val, LR Val \t", best_score_gm_val, best_score_lr_val)
final_score_gm.append(best_score_gm)
final_score_lr.append(best_score_lr)
final_score_gm_val.append(best_score_gm_val)
final_score_lr_val.append(best_score_lr_val)
print("Test - Averaged scores are for GM,LR", np.mean(final_score_gm), np.mean(final_score_lr))
print("VALIDATION Averaged scores are for GM,LR", np.mean(final_score_gm_val), np.mean(final_score_lr_val))
print("Test Standard deviation are for GM,LR", np.std(final_score_gm), np.std(final_score_lr))
print("VALIDATION Standard deviation are for GM,LR", np.std(final_score_gm_val), np.std(final_score_lr_val))