-
Notifications
You must be signed in to change notification settings - Fork 32
/
strategy.py
375 lines (326 loc) · 16.2 KB
/
strategy.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
import numpy as np
from torch import nn
import sys
import torch
import torch.nn.functional as F
import torch.optim as optim
from torch.autograd import Variable
from torch.utils.data import DataLoader
from copy import deepcopy
import pdb
import resnet
from torch.distributions.categorical import Categorical
class Strategy:
def __init__(self, X, Y, idxs_lb, net, handler, args):
self.X = X
self.Y = Y
self.idxs_lb = idxs_lb
self.net = net
self.handler = handler
self.args = args
self.n_pool = len(Y)
use_cuda = torch.cuda.is_available()
def query(self, n):
pass
def update(self, idxs_lb):
self.idxs_lb = idxs_lb
def _train(self, epoch, loader_tr, optimizer):
self.clf.train()
accFinal = 0.
for batch_idx, (x, y, idxs) in enumerate(loader_tr):
x, y = Variable(x.cuda()), Variable(y.cuda())
optimizer.zero_grad()
out, e1 = self.clf(x)
loss = F.cross_entropy(out, y)
accFinal += torch.sum((torch.max(out,1)[1] == y).float()).data.item()
loss.backward()
# clamp gradients, just in case
for p in filter(lambda p: p.grad is not None, self.clf.parameters()): p.grad.data.clamp_(min=-.1, max=.1)
optimizer.step()
return accFinal / len(loader_tr.dataset.X), loss.item()
def train(self, reset=True, optimizer=0, verbose=True, data=[], net=[]):
def weight_reset(m):
newLayer = deepcopy(m)
if isinstance(m, nn.Conv2d) or isinstance(m, nn.Linear):
m.reset_parameters()
n_epoch = self.args['n_epoch']
if reset: self.clf = self.net.apply(weight_reset).cuda()
if type(net) != list: self.clf = net
if type(optimizer) == int: optimizer = optim.Adam(self.clf.parameters(), lr = self.args['lr'], weight_decay=0)
idxs_train = np.arange(self.n_pool)[self.idxs_lb]
loader_tr = DataLoader(self.handler(self.X[idxs_train], torch.Tensor(self.Y.numpy()[idxs_train]).long(), transform=self.args['transform']), shuffle=True, **self.args['loader_tr_args'])
if len(data) > 0:
loader_tr = DataLoader(self.handler(data[0], torch.Tensor(data[1]).long(), transform=self.args['transform']), shuffle=True, **self.args['loader_tr_args'])
epoch = 1
accCurrent = 0.
bestAcc = 0.
attempts = 0
while accCurrent < 0.99:
accCurrent, lossCurrent = self._train(epoch, loader_tr, optimizer)
if bestAcc < accCurrent:
bestAcc = accCurrent
attempts = 0
else: attempts += 1
epoch += 1
if verbose: print(str(epoch) + ' ' + str(attempts) + ' training accuracy: ' + str(accCurrent), flush=True)
# reset if not converging
if (epoch % 1000 == 0) and (accCurrent < 0.2) and (self.args['modelType'] != 'linear'):
self.clf = self.net.apply(weight_reset)
optimizer = optim.Adam(self.clf.parameters(), lr = self.args['lr'], weight_decay=0)
if attempts >= 50 and self.args['modelType'] == 'linear': break
#if attempts >= 50 and self.args['modelType'] != 'linear' and len(idxs_train) > 1000:
# self.clf = self.net.apply(weight_reset)
# optimizer = optim.Adam(self.clf.parameters(), lr = self.args['lr'], weight_decay=0)
# attempts = 0
def train_val(self, valFrac=0.1, opt='adam', verbose=False):
def weight_reset(m):
newLayer = deepcopy(m)
if isinstance(m, nn.Conv2d) or isinstance(m, nn.Linear):
newLayer.reset_parameters()
m.reset_parameters()
if verbose: print(' ',flush=True)
if verbose: print('getting validation minimizing number of epochs', flush=True)
self.clf = self.net.apply(weight_reset).cuda()
if opt == 'adam': optimizer = optim.Adam(self.clf.parameters(), lr=self.args['lr'], weight_decay=0)
if opt == 'sgd': optimizer = optim.SGD(self.clf.parameters(), lr=self.args['lr'], weight_decay=0)
idxs_train = np.arange(self.n_pool)[self.idxs_lb]
nVal = int(len(idxs_train) * valFrac)
idxs_train = idxs_train[np.random.permutation(len(idxs_train))]
idxs_val = idxs_train[:nVal]
idxs_train = idxs_train[nVal:]
loader_tr = DataLoader(self.handler(self.X[idxs_train], torch.Tensor(self.Y.numpy()[idxs_train]).long(), transform=self.args['transform']), shuffle=True, **self.args['loader_tr_args'])
epoch = 1
accCurrent = 0.
bestLoss = np.inf
attempts = 0
ce = nn.CrossEntropyLoss()
valTensor = torch.Tensor(self.Y.numpy()[idxs_val]).long()
attemptThresh = 10
while True:
accCurrent, lossCurrent = self._train(epoch, loader_tr, optimizer)
valPreds = self.predict_prob(self.X[idxs_val], valTensor, exp=False)
loss = ce(valPreds, valTensor).item()
if loss < bestLoss:
bestLoss = loss
attempts = 0
bestEpoch = epoch
else:
attempts += 1
if attempts == attemptThresh: break
if verbose: print(epoch, attempts, loss, bestEpoch, bestLoss, flush=True)
epoch += 1
return bestEpoch
def get_dist(self, epochs, nEns=1, opt='adam', verbose=False):
def weight_reset(m):
newLayer = deepcopy(m)
if isinstance(m, nn.Conv2d) or isinstance(m, nn.Linear):
newLayer.reset_parameters()
m.reset_parameters()
if verbose: print(' ',flush=True)
if verbose: print('training to indicated number of epochs', flush=True)
ce = nn.CrossEntropyLoss()
idxs_train = np.arange(self.n_pool)[self.idxs_lb]
loader_tr = DataLoader(self.handler(self.X[idxs_train], torch.Tensor(self.Y.numpy()[idxs_train]).long(), transform=self.args['transform']), shuffle=True, **self.args['loader_tr_args'])
dataSize = len(idxs_train)
N = np.round((epochs * len(loader_tr)) ** 0.5)
allAvs = []
allWeights = []
for m in range(nEns):
# initialize new model and optimizer
net = self.net.apply(weight_reset).cuda()
if opt == 'adam': optimizer = optim.Adam(net.parameters(), lr=self.args['lr'], weight_decay=0)
if opt == 'sgd': optimizer = optim.SGD(net.parameters(), lr=self.args['lr'], weight_decay=0)
nUpdates = k = 0
ek = (k + 1) * N
pVec = torch.cat([torch.zeros_like(p).cpu().flatten() for p in self.clf.parameters()])
avIterates = []
for epoch in range(epochs + 1):
correct = lossTrain = 0.
net = net.train()
for ind, (x, y, _) in enumerate(loader_tr):
x, y = x.cuda(), y.cuda()
optimizer.zero_grad()
output, _ = net(x)
correct += torch.sum(output.argmax(1) == y).item()
loss = ce(output, y)
loss.backward()
lossTrain += loss.item() * len(y)
optimizer.step()
flat = torch.cat([deepcopy(p.detach().cpu()).flatten() for p in net.parameters()])
pVec = pVec + flat
nUpdates += 1
if nUpdates > ek:
avIterates.append(pVec / N)
pVec = torch.cat([torch.zeros_like(p).cpu().flatten() for p in net.parameters()])
k += 1
ek = (k + 1) * N
lossTrain /= dataSize
accuracy = correct / dataSize
if verbose: print(m, epoch, nUpdates, accuracy, lossTrain, flush=True)
allAvs.append(avIterates)
allWeights.append(torch.cat([deepcopy(p.detach().cpu()).flatten() for p in net.parameters()]))
for m in range(nEns):
avIterates = torch.stack(allAvs[m])
if k > 1: avIterates = torch.stack(allAvs[m][1:])
avIterates = avIterates - torch.mean(avIterates, 0)
allAvs[m] = avIterates
return allWeights, allAvs, optimizer, net
def getNet(self, params):
i = 0
model = deepcopy(self.clf).cuda()
for p in model.parameters():
L = len(p.flatten())
param = params[i:(i + L)]
p.data = param.view(p.size())
i += L
return model
def fitBatchnorm(self, model):
idxs_train = np.arange(self.n_pool)[self.idxs_lb]
loader_tr = DataLoader(self.handler(self.X[idxs_train], torch.Tensor(self.Y.numpy()[idxs_train]).long(), transform=self.args['transform']), shuffle=True, **self.args['loader_tr_args'])
model = model.cuda()
for ind, (x, y, _) in enumerate(loader_tr):
x, y = x.cuda(), y.cuda()
output = model(x)
return model
def sampleNet(self, weights, iterates):
nEns = len(weights)
k = len(iterates[0])
i = np.random.randint(nEns)
z = torch.randn(k, 1)
weightSample = weights[i].view(-1) - torch.mm(iterates[i].t(), z).view(-1) / np.sqrt(k)
sampleNet = self.getNet(weightSample).cuda()
sampleNet = self.fitBatchnorm(sampleNet)
return sampleNet
def getPosterior(self, weights, iterates, X, Y, nSamps=50):
net = self.fitBatchnorm(self.sampleNet(weights, iterates))
output = self.predict_prob(X, Y, model=net) / nSamps
print(' ', flush=True)
ce = nn.CrossEntropyLoss()
print('sampling models', flush=True)
for i in range(nSamps - 1):
net = self.fitBatchnorm(self.sampleNet(weights, iterates))
output = output + self.predict_prob(X, Y, model=net) / nSamps
print(i+2, torch.sum(torch.argmax(output, 1) == Y).item() / len(Y), flush=True)
return output.numpy()
def predict(self, X, Y):
if type(X) is np.ndarray:
loader_te = DataLoader(self.handler(X, Y, transform=self.args['transformTest']),
shuffle=False, **self.args['loader_te_args'])
else:
loader_te = DataLoader(self.handler(X.numpy(), Y, transform=self.args['transformTest']),
shuffle=False, **self.args['loader_te_args'])
self.clf.eval()
P = torch.zeros(len(Y)).long()
with torch.no_grad():
for x, y, idxs in loader_te:
x, y = Variable(x.cuda()), Variable(y.cuda())
out, e1 = self.clf(x)
pred = out.max(1)[1]
P[idxs] = pred.data.cpu()
return P
def predict_prob(self, X, Y, model=[], exp=True):
if type(model) == list: model = self.clf
loader_te = DataLoader(self.handler(X, Y, transform=self.args['transformTest']), shuffle=False, **self.args['loader_te_args'])
model = model.eval()
probs = torch.zeros([len(Y), len(np.unique(self.Y))])
with torch.no_grad():
for x, y, idxs in loader_te:
x, y = Variable(x.cuda()), Variable(y.cuda())
out, e1 = model(x)
if exp: out = F.softmax(out, dim=1)
probs[idxs] = out.cpu().data
return probs
def predict_prob_dropout(self, X, Y, n_drop):
loader_te = DataLoader(self.handler(X, Y, transform=self.args['transformTest']),
shuffle=False, **self.args['loader_te_args'])
self.clf.train()
probs = torch.zeros([len(Y), len(np.unique(Y))])
with torch.no_grad():
for i in range(n_drop):
print('n_drop {}/{}'.format(i+1, n_drop))
for x, y, idxs in loader_te:
x, y = Variable(x.cuda()), Variable(y.cuda())
out, e1 = self.clf(x)
prob = F.softmax(out, dim=1)
probs[idxs] += out.cpu().data
probs /= n_drop
return probs
def predict_prob_dropout_split(self, X, Y, n_drop):
loader_te = DataLoader(self.handler(X, Y, transform=self.args['transformTest']),
shuffle=False, **self.args['loader_te_args'])
self.clf.train()
probs = torch.zeros([n_drop, len(Y), len(np.unique(Y))])
with torch.no_grad():
for i in range(n_drop):
print('n_drop {}/{}'.format(i+1, n_drop))
for x, y, idxs in loader_te:
x, y = Variable(x.cuda()), Variable(y.cuda())
out, e1 = self.clf(x)
probs[i][idxs] += F.softmax(out, dim=1).cpu().data
return probs
def get_embedding(self, X, Y, return_probs=False):
loader_te = DataLoader(self.handler(X, Y, transform=self.args['transformTest']),
shuffle=False, **self.args['loader_te_args'])
self.clf.eval()
embedding = torch.zeros([len(Y), self.clf.get_embedding_dim()])
probs = torch.zeros(len(Y), self.clf.linear.out_features)
with torch.no_grad():
for x, y, idxs in loader_te:
x, y = Variable(x.cuda()), Variable(y.cuda())
out, e1 = self.clf(x)
embedding[idxs] = e1.data.cpu()
if return_probs:
pr = F.softmax(out,1)
probs[idxs] = pr.data.cpu()
if return_probs: return embedding, probs
return embedding
# gradient embedding for badge (assumes cross-entropy loss)
def get_grad_embedding(self, X, Y, model=[]):
if type(model) == list:
model = self.clf
embDim = model.get_embedding_dim()
model.eval()
nLab = len(np.unique(Y))
embedding = np.zeros([len(Y), embDim * nLab])
loader_te = DataLoader(self.handler(X, Y, transform=self.args['transformTest']),
shuffle=False, **self.args['loader_te_args'])
with torch.no_grad():
for x, y, idxs in loader_te:
x, y = Variable(x.cuda()), Variable(y.cuda())
cout, out = model(x)
out = out.data.cpu().numpy()
batchProbs = F.softmax(cout, dim=1).data.cpu().numpy()
maxInds = np.argmax(batchProbs,1)
for j in range(len(y)):
for c in range(nLab):
if c == maxInds[j]:
embedding[idxs[j]][embDim * c : embDim * (c+1)] = deepcopy(out[j]) * (1 - batchProbs[j][c])
else:
embedding[idxs[j]][embDim * c : embDim * (c+1)] = deepcopy(out[j]) * (-1 * batchProbs[j][c])
return torch.Tensor(embedding)
# fisher embedding for bait (assumes cross-entropy loss)
def get_exp_grad_embedding(self, X, Y, probs=[], model=[]):
if type(model) == list:
model = self.clf
embDim = model.get_embedding_dim()
model.eval()
nLab = len(np.unique(Y))
embedding = np.zeros([len(Y), nLab, embDim * nLab])
for ind in range(nLab):
loader_te = DataLoader(self.handler(X, Y, transform=self.args['transformTest']),
shuffle=False, **self.args['loader_te_args'])
with torch.no_grad():
for x, y, idxs in loader_te:
x, y = Variable(x.cuda()), Variable(y.cuda())
cout, out = model(x)
out = out.data.cpu().numpy()
batchProbs = F.softmax(cout, dim=1).data.cpu().numpy()
for j in range(len(y)):
for c in range(nLab):
if c == ind:
embedding[idxs[j]][ind][embDim * c : embDim * (c+1)] = deepcopy(out[j]) * (1 - batchProbs[j][c])
else:
embedding[idxs[j]][ind][embDim * c : embDim * (c+1)] = deepcopy(out[j]) * (-1 * batchProbs[j][c])
if len(probs) > 0: embedding[idxs[j]][ind] = embedding[idxs[j]][ind] * np.sqrt(probs[idxs[j]][ind])
else: embedding[idxs[j]][ind] = embedding[idxs[j]][ind] * np.sqrt(batchProbs[j][ind])
return torch.Tensor(embedding)