-
Notifications
You must be signed in to change notification settings - Fork 1
/
utils.py
635 lines (549 loc) · 24.2 KB
/
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
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
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
import scipy.sparse
# import torch
from tqdm import tqdm
import numpy as np
import pandas as pd
import cvxpy as cp
import torch
import torch.nn as nn
import argparse
import utils
from scnn.metrics import Metrics
from scnn.models import ConvexGatedReLU, LinearModel
from scnn.activations import sample_gate_vectors
from scnn.solvers import RFISTA, CVXPYSolver
from scnn.optimize import optimize_model
from scnn.regularizers import NeuronGL1,L1
import time
from sklearn import linear_model
from torch.optim import AdamW
from transformers import get_linear_schedule_with_warmup
from torch.utils.data import TensorDataset, DataLoader, RandomSampler, SequentialSampler
def one_hot(a, num_classes):
return np.squeeze(np.eye(num_classes)[a.reshape(-1)])
def get_SJLT_matrix(m, n, s):
# This function returns SJLT sketching matrix in the form of a sparse matrix.
# m: sketch size, n: number of data samples, s: sparsity
nonzeros = 2*np.random.binomial(1, 0.5, size=(s*n)) - 1 # Rademacher random variables
K = int(np.ceil(s*n / m)) # number of repetitions
shuffled_row_indices = np.zeros((K*m), dtype=np.int32)
all_row_indices = np.linspace(0, m-1, m, dtype=np.int32)
for k in range(K):
shuffled_row_indices[k*m:(k+1)*m] = np.random.permutation(all_row_indices)
I = shuffled_row_indices[0:s*n]
J = np.repeat(np.linspace(0, n-1, n, dtype=np.int32), s)
V = nonzeros
S = scipy.sparse.coo_matrix((V,(I,J)), shape=(m,n), dtype=np.int8)
S = S.tocsr()
return S
def general_cross_product(X):
# Input:
# X: (d-1)*d matrix
# Output:
# v: d-dim vector
d = X.shape[1]
v = (-1*np.ones(d))**np.arange(d)
base = np.arange(d)
for i in range(d):
X_i = X[:,np.delete(base,i,0)]
v[i]*=np.linalg.det(X_i)
return v
def general_cross_product_v2(X,max_trial=20):
d = X.shape[1]
iter_num = 0
while iter_num<max_trial:
v_aug = np.random.randn(d)
X_aug = np.concatenate([X,v_aug.reshape([1,-1])],axis=0)
if np.linalg.det(X_aug)!=0:
break
iter_num+=1
# print(iter_num)
if iter_num<max_trial:
q = np.zeros(d)
q[-1]=1
v = np.linalg.solve(X_aug,q)
return v, 0
else:
return None, -1
def general_cross_product_v3(X):
# using svd
U, s, Vt = np.linalg.svd(X.T)
return U[:,-1]
# s = np.sum(v)/np.linalg.det(X_aug)
# v = s*v
def test_general_cross_product(d=3, gcp_ver='v3', max_trial=10):
X = np.random.randn(d-1,d)
if gcp_ver=='v2':
v, flag = general_cross_product_v2(X, max_trial=max_trial)
elif gcp_ver=='std':
v = general_cross_product(X)
flag = 0
elif gcp_ver=='v3':
v = general_cross_product_v3(X)
flag = 0
if flag==0:
x = np.random.randn(1,d)
print(X@v)
print((x@v).item())
print(np.linalg.det(np.concatenate([x,X],axis=0)).item())
def relu(x):
return np.maximum(0,x)
def drelu(x):
return x>=0
def cvx_sample_vectors(training_data_np, max_neurons, arr_select = 'Gaussian', with_bias = True, sketch=True, sdim=50,
gcp_ver='v3', max_trial=20, aug_sym = False):
n, Embedding_Size = np.shape(training_data_np)
if arr_select == 'Gaussian':
if with_bias:
G = np.random.randn(Embedding_Size+1,max_neurons)
else:
G = np.random.randn(Embedding_Size,max_neurons)
elif arr_select == 'Geometric_Algebra':
if sketch:
S = get_SJLT_matrix(Embedding_Size,sdim,1)
training_data_np = training_data_np@S
d = sdim
else:
d = Embedding_Size
S = np.identity(d)
if with_bias:
G = np.zeros([Embedding_Size+1,max_neurons])
for i in range(max_neurons):
# sample x_1,...x_{d}
iter_num = 0
while iter_num<max_trial:
iter_num+=1
index = np.random.choice(n,d,replace=False)
x_d = training_data_np[index[-1],:].reshape([1,-1])
A = training_data_np[index[:-1],:]-x_d
# A = A/np.linalg.norm(A)*10
if gcp_ver=='v2':
v, flag = general_cross_product_v2(A, max_trial=max_trial)
elif gcp_ver=='std':
v = general_cross_product(A)
flag = 0
elif gcp_ver=='v3':
v = general_cross_product_v3(A)
flag = 0
if flag==0:
v = v.reshape([1,-1])
break
# print(np.linalg.norm(v))
if iter_num<max_trial:
G[:,i] = np.concatenate([[email protected],[email protected]],axis=1)/np.linalg.norm([email protected])
else:
print('Warning: reach max trial')
v = np.random.randn(d).reshape([1,-1])
G[:,i] = np.concatenate([[email protected],[email protected]],axis=1)/np.linalg.norm([email protected])
else:
G = np.zeros([Embedding_Size,max_neurons])
for i in range(max_neurons):
index = np.random.choice(n,d-1,replace=False)
if gcp_ver=='v2':
v = general_cross_product_v2(training_data_np[index,:], max_trial=max_trial)
elif gcp_ver == 'std':
v = general_cross_product(training_data_np[index,:])
elif gcp_ver == 'std':
v = general_cross_product(training_data_np[index,:])
G[:,i] = [email protected]/np.linalg.norm([email protected])
if aug_sym == True:
G = np.concatenate([G,-G],axis=1)
elif arr_select == 'Polished_Gaussian':
if sketch:
S = get_SJLT_matrix(Embedding_Size,sdim,1)
training_data_np = training_data_np@S
d = sdim
else:
d = Embedding_Size
S = np.identity(d)
G_aug = np.random.randn(d,max_neurons)
if with_bias:
G = np.zeros([Embedding_Size+1,max_neurons])
for i in range(max_neurons):
index_x_d = np.random.choice(n,1)
x_d = training_data_np[index_x_d,:].reshape([1,-1])
product = np.abs((training_data_np-x_d)@(G_aug[:,i].reshape(-1)-x_d.reshape(-1)))
index = product.argsort()[1:d] # remove index_x_d from indexing
# print(index)
A = training_data_np[index,:]-x_d
# print(A.shape)
if fast_comp:
v = general_cross_product_v2(A, max_trial=max_trial).reshape([1,-1])
else:
v = general_cross_product(A).reshape([1,-1])
G[:,i] = np.concatenate([[email protected],[email protected]],axis=1)/np.linalg.norm([email protected])
else:
G = np.zeros([Embedding_Size,max_neurons])
for i in range(max_neurons):
product = np.abs(training_data_np@G_aug[:,i])
index = product.argsort()[1:d]
if fast_comp:
v = general_cross_product_v2(training_data_np[index,:], max_trial=max_trial)
else:
v = general_cross_product(training_data_np[index,:])
G[:,i] = [email protected]/np.linalg.norm([email protected])
return G
def cvx_solver_mosek(training_data_np,training_labels_np,beta=1e-3,Hidden=50,
arr_select = 'Gaussian', with_bias = True, activation='gReLU',
reg_p=2,sdim=100,gcp_ver='v3',verbose=False,add_eps=True,eps=1e-8,
aug_sym=False,solver='mosek', cp_verbose=False):
n, Embedding_Size = np.shape(training_data_np)
if arr_select == 'GA_enum':
assert Embedding_Size==2, 'wrong input dimension'
G = np.zeros([3,n*(n-1)])
count = 0
for i in range(n):
for j in range(i+1,n):
xi = training_data_np[i]
xj = training_data_np[j]
v = (xi-xj)/np.linalg.norm(xi-xj)
G[:2,count] = v
G[2,count] = -xj@v
G[:2,count+1] = -v
G[2,count+1] = xi@v
count+=2
else:
if sdim<=0:
sketch=False
else:
sketch=True
G = utils.cvx_sample_vectors(training_data_np, Hidden, arr_select=arr_select, with_bias=True,sketch=sketch,
sdim=sdim,gcp_ver=gcp_ver,max_trial=20,aug_sym=aug_sym)
if arr_select=='Geometric_Algebra' and add_eps:
G[-1,:]+=eps
training_data_np = np.concatenate([training_data_np,np.ones([n,1])],axis=1)
Embedding_Size += 1
dmat= drelu(np.matmul(training_data_np,G))
if activation=='gReLU':
# Optimal CVX
m1=dmat.shape[1]
Uopt1=cp.Variable((Embedding_Size,m1))
## Below we use squared loss as a performance metric for binary classification
yopt1=cp.sum(cp.multiply(dmat,(training_data_np@Uopt1)),axis=1)
if with_bias:
regularization = cp.mixed_norm(Uopt1[:-1,:].T,reg_p,1)
else:
regularization = cp.mixed_norm(Uopt1.T,reg_p,1)
cost=cp.sum_squares(training_labels_np-yopt1)+beta*regularization
constraints=[]
prob=cp.Problem(cp.Minimize(cost))
prob.solve(solver=cp.MOSEK,verbose=verbose)
cvx_opt=prob.value
if verbose:
print("Convex program objective value: ",cvx_opt)
return G, Uopt1._value
elif activation=='ReLU':
dmat=(np.unique(dmat,axis=1))
m1=dmat.shape[1]
Uopt1=cp.Variable((Embedding_Size,m1))
Uopt2=cp.Variable((Embedding_Size,m1))
yopt1=cp.sum(cp.multiply(dmat,(training_data_np@Uopt1)),axis=1)
yopt2=cp.sum(cp.multiply(dmat,(training_data_np@Uopt2)),axis=1)
if with_bias:
regularization = cp.mixed_norm(Uopt1[:-1,:].T,reg_p,1)+cp.mixed_norm(Uopt2[:-1,:].T,reg_p,1)
else:
regularization = cp.mixed_norm(Uopt1.T,2,1)+cp.mixed_norm(Uopt1.T,2,1)
cost=cp.sum_squares(yopt1-yopt2-training_labels_np)+beta*regularization
constraints=[]
constraints+=[cp.multiply((2*dmat-np.ones((n,m1))),(training_data_np@Uopt1))>=0]
constraints+=[cp.multiply((2*dmat-np.ones((n,m1))),(training_data_np@Uopt2))>=0]
prob=cp.Problem(cp.Minimize(cost),constraints)
prob.solve(solver=cp.MOSEK,verbose=verbose)
cvx_opt=prob.value
if verbose:
print("Convex program objective value: ",cvx_opt)
U1, U2 = Uopt1._value, Uopt2._value
return U1, U2
if activation=='Lasso':
G = G/np.linalg.norm(G[:-1,:],2,axis=0)
m1=G.shape[1]
z =cp.Variable(m1)
t = cp.Variable(1)
yopt = relu(training_data_np@G)@z+t
regularization = cp.norm(z,1)
cost=cp.sum_squares(yopt-training_labels_np)+beta*regularization
prob=cp.Problem(cp.Minimize(cost))
if solver == 'mosek':
cvx_solver = cp.MOSEK
elif solver == 'scs':
cvx_solver = cp.SCS
prob.solve(solver=cvx_solver,verbose=cp_verbose)
cvx_opt=prob.value
if verbose:
print("Convex program objective value: ",cvx_opt)
z_v, t_v = z._value, t._value
return G, z_v, t_v
def cvx_solver_evaluate(test_data_np,test_labels_np,params,activation='gReLU', with_bias=True):
if with_bias:
n = np.shape(test_data_np)[0]
test_data_np = np.concatenate([test_data_np,np.ones([n,1])],axis=1)
if activation=='gReLU':
G, U = params
preds = np.sum(drelu(test_data_np@G)*(test_data_np@U),axis=1)
elif activation=='ReLU':
U1, U2 = params
preds = np.sum(relu(test_data_np@U1)-relu(test_data_np@U2),axis=1)
elif activation=='Lasso':
G, z, t = params
preds = relu(test_data_np@G)@z+t
acc = accuracy(preds,test_labels_np)*100
# print("Convex accuacy:{:.2f}%".format(acc))
return acc
def set_seed(seed_value=42):
"""Set seed for reproducibility."""
np.random.seed(seed_value)
torch.manual_seed(seed_value)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(seed_value)
# Create the NNClassifier class
class NNClassifier(nn.Module):
"""2 layer NN Model for Classification Tasks.
"""
def __init__(self, hidden, D_in = 768*2, add_skip=False):
"""
@param bert: a BertModel object
@param classifier: a torch.nn.Module classifier
"""
super(NNClassifier, self).__init__()
# Specify hidden size of BERT, hidden size of our classifier, and number of labels
D_in, H, D_out = D_in, hidden, 1
self.add_skip = add_skip
# Instantiate a two-layer feed-forward classifier
self.classifier = nn.Sequential(
nn.Linear(D_in, H),
nn.ReLU(),
nn.Linear(H, D_out)
)
if add_skip:
self.skip_layer = nn.Sequential(
nn.Linear(D_in, 1),
nn.Linear(1, D_out)
)
def forward(self, input_ids):
# BY LOADING THE CSV FILE OF THE OUTPUT EMBEDDING FROM BERT
logits = self.classifier(input_ids)
if self.add_skip:
logits+= self.skip_layer(input_ids)
logits = logits.squeeze()
return logits
def evaluate(model, val_dataloader, device):
loss_fn = nn.MSELoss()
"""Measure model's performance on the validation set."""
model.eval() # Evaluation mode
val_accuracy_list = []
val_loss_list = []
for batch in val_dataloader:
b_input_ids, b_labels = tuple(t.to(device) for t in batch)
with torch.no_grad():
logits = model(b_input_ids).squeeze()
loss = loss_fn(logits, b_labels)
val_loss_list.append(loss.item())
preds = torch.sign(logits)
# preds = torch.argmax(logits, dim=1).flatten()
accuracy = (preds == b_labels).cpu().numpy().mean() * 100
val_accuracy_list.append(accuracy)
val_loss = np.mean(val_loss_list)
val_accuracy = np.mean(val_accuracy_list)
return val_loss, val_accuracy
def transform_U(U, X):
"""
For each row in U, find the (d-1) rows in X with the lowest inner-product magnitude,
then replace the row in U with the normal of the hyperplane through the origin and those points.
"""
d = U.shape[1] # Dimensions
transformed_U = np.zeros_like(U)
#for i, u_row in enumerate(U):
for i, u_row in tqdm(enumerate(U), total=U.shape[0], desc="Processing neurons"):
u_row_normalized = u_row# / np.linalg.norm(u_row)
inner_products = [np.abs(np.dot(u_row_normalized, x_row)) for x_row in X]#/ np.linalg.norm(x_row)
#add debug point
# Get indices of the (d-1) smallest inner products
smallest_indices = np.argsort(inner_products)[:d-1]
# Get the (d-1) rows from X
selected_rows = X[smallest_indices]
# Find normal of hyperplane passing through the origin and these rows
normal = find_normal_of_hyperplane(selected_rows)
# Replace the current row of U with this normal
transformed_U[i] = normal
return transformed_U
def find_normal_of_hyperplane(points):
"""
Find the normal vector of the hyperplane passing through the origin and the given (d-1) points.
Assumes points is a (d-1)xN matrix, where each row is a point in d-dimensional space.
"""
method = 'eigh'
# Use SVD to find the null space of the matrix formed by points
if method == 'svd':
u, s, vh = np.linalg.svd(points, full_matrices=True)
#d = u.shape[1]
# MP: change this to rank k svd not the full svd.
# problem: smallest singular value is zero for MNIST.
# The normal vector is the last column of vh, corresponding to the smallest singular value
normal = vh[-1]
else:
G = np.dot(points.T, points)
# Compute the eigenvalues and eigenvectors
eigenvalues, eigenvectors = np.linalg.eigh(G)
# The smallest eigenvalue's corresponding eigenvector
normal = eigenvectors[:, 0]
normalized_normal = normal / np.linalg.norm(normal)
return normalized_normal
def train(model, optimizer, scheduler, train_dataloader, val_dataloader=None, epochs=4, evaluation=False, freq_batch=5,
polish=False,polish_freq=1, sdim=100):
"""Train the BertClassifier model."""
train_loss_set = []
val_loss_set = []
train_accuracy_set = []
val_accuracy_set = []
cumulative_time = []
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
total_time_elapsed = 0
print("Start training...\n")
t0_start = time.time()
loss_fn = nn.MSELoss()
for epoch_i in range(epochs):
total_loss, total_correct, total_preds, total_num = 0, 0, 0, 0
model.train()
data_list = []
labels_list = []
for step, batch in enumerate(train_dataloader):
b_input_ids, b_labels = tuple(t.to(device) for t in batch)
data_list.append(b_input_ids.cpu())
labels_list.append(b_labels.cpu().numpy())
model.zero_grad()
logits = model(b_input_ids)
loss = loss_fn(logits, b_labels)
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
optimizer.step()
scheduler.step()
total_loss += loss.item()*int(b_labels.size(0))
total_num += b_labels.size(0)
preds = torch.sign(logits)
total_correct += (preds == b_labels).sum().item()
total_preds += b_labels.size(0)
# Print training results for every freq_batch
if (step % freq_batch == 0 and step != 0) or (step == len(train_dataloader) - 1):
#print(f"Epoch: {epoch_i + 1}, Batch: {step}, Batch Loss: {batch_loss:.6f}, "
#f"Batch Accuracy: {batch_accuracy:.2f}%, Time: {time_elapsed:.2f}s")
avg_train_loss = total_loss / total_num
total_loss = 0
total_num = 0
train_accuracy = (total_correct / total_preds) * 100
total_correct = 0
total_preds = 0
train_loss_set.append(avg_train_loss)
train_accuracy_set.append(train_accuracy)
time_elapsed = time.time() - t0_start
cumulative_time.append(time_elapsed)
'''
print(f"Epoch: {epoch_i + 1}, Batch: {step}, Train Loss: {avg_train_loss:.6f}, "
f"Train Accuracy: {train_accuracy:.2f}%, Time: {time_elapsed:.2f}s")
'''
if evaluation:
val_loss, val_accuracy = evaluate(model, val_dataloader, device)
val_loss_set.append(val_loss)
val_accuracy_set.append(val_accuracy)
if polish:
if (epoch_i)%polish_freq==0:
Uorg = model.classifier[0].weight.detach().numpy().T
bias = model.classifier[0].bias.detach().numpy()
Xdata_org = torch.cat(data_list, dim=0)
if sdim>0:
onestrain = np.ones((Xdata_org.shape[0], 1))
S = get_SJLT_matrix(Xdata_org.shape[1],sdim,1)
Xdata_np = Xdata_org.detach().numpy()
Xdata_np_sketch = Xdata_np@S
Xdata = np.hstack((Xdata_np_sketch, onestrain))
ydata = np.concatenate(labels_list)
Uorg = np.vstack((S.T@Uorg, bias))
new_U = transform_U(np.copy(Uorg).T,Xdata).T
else:
onestrain = np.ones((Xdata_org.shape[0], 1))
Uorg = np.vstack((Uorg, bias))
Xdata = np.hstack((Xdata_org.detach().numpy(), onestrain))
ydata = np.concatenate(labels_list)
new_U = transform_U(np.copy(Uorg).T,Xdata).T
beta = 1e-3
A = relu(Xdata@new_U)
AtA = A.T@A
w_ls = np.linalg.solve(AtA + beta*np.identity(AtA.shape[0]),A.T@ydata)
for col in range(new_U.shape[1]):
if np.abs(w_ls[col]) > 1e-10:
scale = np.sqrt(np.linalg.norm(new_U[:,col])/np.abs(w_ls[col]))
new_U[:,col] = new_U[:,col]/scale
w_ls[col] = w_ls[col]*scale
else:
new_U[:,col] = np.zeros_like(new_U[:,col])
w_ls[col] = 0
print('dropped neuron')
if sdim>0:
model.classifier[0].weight.data = torch.tensor(new_U[:-1,:][email protected],requires_grad=True)
else:
model.classifier[0].weight.data = torch.tensor(new_U[:-1,:].T,requires_grad=True)
model.classifier[0].bias.data = torch.tensor(new_U[-1,:],requires_grad=True)
model.classifier[2].weight.data = torch.tensor(w_ls,requires_grad=True).float().reshape([1,-1])
model.to(device)
if evaluation:
print('Epoch: {} Train_acc: {} Test_acc: {}'.format(epoch_i+1, train_accuracy, val_accuracy))
return cumulative_time, train_loss_set, val_loss_set, train_accuracy_set, val_accuracy_set
def initialize_model(hidden, epochs=4, lr=1e-5, beta = 1e-2, add_skip=False, D_in=768,train_dataloader=None):
"""Initialize the Bert Classifier, the optimizer and the learning rate scheduler.
"""
# Instantiate Bert Classifier
bert_classifier = NNClassifier(hidden, D_in=D_in, add_skip = add_skip)
# Tell PyTorch to run the model on GPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
#bert_classifier = bert_classifier.to(device)
bert_classifier.to(device)
# Create the optimizer
optimizer = AdamW(bert_classifier.parameters(),
lr=lr, # Default learning rate
eps=1e-8, # Default epsilon value
weight_decay = beta #weight decay
)
# Total number of training steps
total_steps = len(train_dataloader) * epochs
# Set up the learning rate scheduler
scheduler = get_linear_schedule_with_warmup(optimizer,
num_warmup_steps=0, # Default value
num_training_steps=total_steps)
return bert_classifier, optimizer, scheduler
def accuracy(logits, y):
return np.sum((np.sign(logits) == y)) / len(y)
def scnn_inner(training_data_np,training_labels_np,test_data_np, test_labels_np,Hidden,method,c,beta=1e-3,
verbose=True,tol=1e-7,add_skip=False,sdim=50,activation='grelu',gcp_ver='v3',solver='std',
add_eps=False, eps=1e-8):
if sdim<=0:
sketch=False
else:
sketch=True
G = utils.cvx_sample_vectors(training_data_np, Hidden, arr_select=method, with_bias=True,sketch=sketch,sdim=sdim,gcp_ver=gcp_ver,max_trial=20)
G_bias = G[-1,:].reshape(-1)
G = G[:-1,:]
if add_eps:
G_bias = G_bias + eps
if add_skip:
G_bias = np.concatenate([G_bias,np.array([1])],axis=0)
d, m = G.shape
G = np.concatenate([G,np.zeros(d).reshape([-1,1])],axis=1)
if activation == 'grelu':
model = ConvexGatedReLU(G, c=c, bias=True, G_bias=G_bias)
if solver=='cvxpy':
solver = CVXPYSolver(model,'mosek')
else:
solver = RFISTA(model,tol=tol)
elif activation == 'relu':
model = ConvexReLU(G, c=1, bias=True, G_bias=G_bias)
solver = AL(model,tol=tol,max_primal_iters=100,constraint_tol=tol)
beta = beta
metrics = Metrics(train_accuracy=True, train_mse = True, test_accuracy = True, test_mse = True)
cvx_model, metrics = optimize_model(model, solver, metrics, training_data_np, training_labels_np, test_data_np, test_labels_np, regularizer = NeuronGL1(beta), verbose=verbose)
return cvx_model, metrics
def eval_model(training_data_np,training_labels_np,test_data_np, test_labels_np,cvx_model):
preds = cvx_model(test_data_np)
open_AI_test_accuracy = accuracy(np.squeeze(preds), np.squeeze(test_labels_np))
preds = cvx_model(training_data_np)
open_AI_train_accuracy = accuracy(np.squeeze(preds), np.squeeze(training_labels_np))
return open_AI_train_accuracy, open_AI_test_accuracy