bonjour
Lors de l'exécution du code sur google colab, je rencontre un problème lié à theano.
voici le lien du projet complet sur github: https://github.com/EmotionalMinors/p...-detection.git
voici la classe conv_net_train_gpu.py executé:
Code : Sélectionner tout - Visualiser dans une fenêtre à part
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20 loading data...: floatx:float64 data loaded! attr: 3 model architecture: CNN-static using: word2vec vectors [('image shape', 153, 300), ('filter shape', [(200, 1, 1, 300), (200, 1, 2, 300), (200, 1, 3, 300)]), ('hidden_units', [200, 200, 2]), ('dropout', [0.5, 0.5, 0.5]), ('batch_size', 50), ('non_static', False), ('learn_decay', 0.95), ('conv_non_linear', 'relu'), ('non_static', False), ('sqr_norm_lim', 9), ('shuffle_batch', True)] Traceback (most recent call last): File "/content/drive/MyDrive/Colab Notebooks/conv_net_train_gpu.py", line 515, in <module> activations=[Sigmoid]) File "/content/drive/MyDrive/Colab Notebooks/conv_net_train_gpu.py", line 106, in train_conv_net allow_input_downcast=True) File "/usr/local/lib/python3.7/dist-packages/theano/compile/function/__init__.py", line 350, in function output_keys=output_keys, File "/usr/local/lib/python3.7/dist-packages/theano/compile/function/pfunc.py", line 427, in pfunc _pfunc_param_to_in(p, allow_downcast=allow_input_downcast) for p in params File "/usr/local/lib/python3.7/dist-packages/theano/compile/function/pfunc.py", line 427, in <listcomp> _pfunc_param_to_in(p, allow_downcast=allow_input_downcast) for p in params File "/usr/local/lib/python3.7/dist-packages/theano/compile/function/pfunc.py", line 543, in _pfunc_param_to_in raise TypeError(f"Unknown parameter type: {type(param)}") TypeError: Unknown parameter type: <class 'theano.tensor.var.TensorVariable'>
Code Python : Sélectionner tout - Visualiser dans une fenêtre à part
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 """ Sample code for Convolutional Neural Networks for Sentence Classification http://arxiv.org/pdf/1408.5882v2.pdf Much of the code is modified from - deeplearning.net (for ConvNet classes) - https://github.com/mdenil/dropout (for dropout) - https://groups.google.com/forum/#!topic/pylearn-dev/3QbKtCumAW4 (for Adadelta) """ # from conv_net_classes_gpu import LeNetConvPoolLayer, MLPDropout # import _pickle as cPickle try: import cPickle as pickle except: import pickle as pickle import numpy as np from collections import defaultdict, OrderedDict import theano import theano.tensor as T from theano.ifelse import ifelse import os import warnings import sys import time import getpass import csv from conv_net_classes_gpu import LeNetConvPoolLayer, MLPDropout warnings.filterwarnings("ignore") # different non-linearities def ReLU(x): y = T.maximum(0.0, x) return (y) def Sigmoid(x): y = T.nnet.sigmoid(x) return (y) def Tanh(x): y = T.tanh(x) return (y) def Iden(x): y = x return (y) def train_conv_net(datasets, U, ofile, cv=0, attr=0, img_w=300, filter_hs=[3, 4, 5], hidden_units=[100, 2], dropout_rate=[0.5], shuffle_batch=True, n_epochs=25, batch_size=50, lr_decay=0.95, conv_non_linear="relu", activations=[Iden], sqr_norm_lim=9, non_static=True): """ Train a simple conv net img_h = sentence length (padded where necessary) img_w = word vector length (300 for word2vec) filter_hs = filter window sizes hidden_units = [x,y] x is the number of feature maps (per filter window), and y is the penultimate layer sqr_norm_lim = s^2 in the paper lr_decay = adadelta decay parameter """ rng = np.random.RandomState(3435) img_h = len(datasets[0][0][0]) filter_w = img_w feature_maps = hidden_units[0] filter_shapes = [] pool_sizes = [] for filter_h in filter_hs: filter_shapes.append((feature_maps, 1, filter_h, filter_w)) pool_sizes.append((img_h - filter_h + 1, img_w - filter_w + 1)) parameters = [("image shape", img_h, img_w), ("filter shape", filter_shapes), ("hidden_units", hidden_units), ("dropout", dropout_rate), ("batch_size", batch_size), ("non_static", non_static), ("learn_decay", lr_decay), ("conv_non_linear", conv_non_linear), ("non_static", non_static) , ("sqr_norm_lim", sqr_norm_lim), ("shuffle_batch", shuffle_batch)] print(parameters) # define model architecture index = T.iscalar() x = T.tensor3('x', dtype=theano.config.floatX) y = T.ivector('y') mair = T.matrix('mair') Words = theano.shared(value=U, name="Words") zero_vec_tensor = T.vector(dtype=theano.config.floatX) zero_vec = np.zeros(img_w, dtype=theano.config.floatX) set_zero = theano.function([zero_vec_tensor], updates=[(Words, T.set_subtensor(Words[0, :], zero_vec_tensor))], allow_input_downcast=True) conv_layers = [] for i in range(len(filter_hs)): filter_shape = filter_shapes[i] pool_size = pool_sizes[i] conv_layer = LeNetConvPoolLayer(rng, image_shape=None, filter_shape=filter_shape, poolsize=pool_size, non_linear=conv_non_linear) conv_layers.append(conv_layer) layer0_input = Words[T.cast(x.flatten(), dtype="int32")].reshape( (x.shape[0], x.shape[1], x.shape[2], Words.shape[1])) def convolve_user_statuses(statuses): layer1_inputs = [] def sum_mat(mat, out): z = ifelse(T.neq(T.sum(mat, dtype=theano.config.floatX), T.constant(0, dtype=theano.config.floatX)), T.constant(1, dtype=theano.config.floatX), T.constant(0, dtype=theano.config.floatX)) return out + z, theano.scan_module.until(T.eq(z, T.constant(0, dtype=theano.config.floatX))) status_count, _ = theano.scan(fn=sum_mat, sequences=statuses, outputs_info=T.constant(0, dtype=theano.config.floatX)) # Slice-out dummy (zeroed) sentences relv_input = statuses[:T.cast(status_count[-1], dtype='int32')].dimshuffle(0, 'x', 1, 2) for conv_layer in conv_layers: layer1_inputs.append(conv_layer.set_input(input=relv_input).flatten(2)) features = T.concatenate(layer1_inputs, axis=1) avg_feat = T.max(features, axis=0) return avg_feat conv_feats, _ = theano.scan(fn=convolve_user_statuses, sequences=layer0_input) # Add Mairesse features layer1_input = T.concatenate([conv_feats, mair], axis=1) ##mairesse_change hidden_units[0] = feature_maps * len(filter_hs) + datasets[4].shape[1] ##mairesse_change classifier = MLPDropout(rng, input=layer1_input, layer_sizes=hidden_units, activations=activations, dropout_rates=dropout_rate) svm_data = T.concatenate([classifier.layers[0].output, y.dimshuffle(0, 'x')], axis=1) # define parameters of the model and update functions using adadelta params = classifier.params for conv_layer in conv_layers: params += conv_layer.params if non_static: # if word vectors are allowed to change, add them as model parameters params += [Words] cost = classifier.negative_log_likelihood(y) dropout_cost = classifier.dropout_negative_log_likelihood(y) grad_updates = sgd_updates_adadelta(params, dropout_cost, lr_decay, 1e-6, sqr_norm_lim) # shuffle dataset and assign to mini batches. if dataset size is not a multiple of mini batches, replicate # extra data (at random) np.random.seed(3435) if datasets[0].shape[0] % batch_size > 0: extra_data_num = batch_size - datasets[0].shape[0] % batch_size rand_perm = np.random.permutation(range(len(datasets[0]))) train_set_x = datasets[0][rand_perm] train_set_y = datasets[1][rand_perm] train_set_m = datasets[4][rand_perm] extra_data_x = train_set_x[:extra_data_num] extra_data_y = train_set_y[:extra_data_num] extra_data_m = train_set_m[:extra_data_num] new_data_x = np.append(datasets[0], extra_data_x, axis=0) new_data_y = np.append(datasets[1], extra_data_y, axis=0) new_data_m = np.append(datasets[4], extra_data_m, axis=0) else: new_data_x = datasets[0] new_data_y = datasets[1] new_data_m = datasets[4] rand_perm = np.random.permutation(range(len(new_data_x))) new_data_x = new_data_x[rand_perm] new_data_y = new_data_y[rand_perm] new_data_m = new_data_m[rand_perm] n_batches = new_data_x.shape[0] / batch_size n_train_batches = int(np.round(n_batches * 0.9)) # divide train set into train/val sets test_set_x = datasets[2] test_set_y = np.asarray(datasets[3], "int32") test_set_m = datasets[5] train_set_x, train_set_y, train_set_m = shared_dataset((new_data_x[:n_train_batches * batch_size], new_data_y[:n_train_batches * batch_size], new_data_m[:n_train_batches * batch_size])) val_set_x, val_set_y, val_set_m = shared_dataset((new_data_x[n_train_batches * batch_size:], new_data_y[n_train_batches * batch_size:], new_data_m[n_train_batches * batch_size:])) n_val_batches = n_batches - n_train_batches val_model = theano.function([index], classifier.errors(y), givens={ x: val_set_x[index * batch_size: (index + 1) * batch_size], y: val_set_y[index * batch_size: (index + 1) * batch_size], mair: val_set_m[index * batch_size: (index + 1) * batch_size]}, ##mairesse_change allow_input_downcast=False) # compile theano functions to get train/val/test errors test_model = theano.function([index], [classifier.errors(y), svm_data], givens={ x: train_set_x[index * batch_size: (index + 1) * batch_size], y: train_set_y[index * batch_size: (index + 1) * batch_size], mair: train_set_m[index * batch_size: (index + 1) * batch_size]}, ##mairesse_change allow_input_downcast=True) train_model = theano.function([index], cost, updates=grad_updates, givens={ x: train_set_x[index * batch_size:(index + 1) * batch_size], y: train_set_y[index * batch_size:(index + 1) * batch_size], mair: train_set_m[index * batch_size: (index + 1) * batch_size]}, ##mairesse_change allow_input_downcast=True) test_y_pred = classifier.predict(layer1_input) test_error = T.sum(T.neq(test_y_pred, y), dtype=theano.config.floatX) true_p = T.sum(test_y_pred * y, dtype=theano.config.floatX) false_p = T.sum(test_y_pred * T.mod(y + T.ones_like(y, dtype=theano.config.floatX), T.constant(2, dtype='int32'))) false_n = T.sum(y * T.mod(test_y_pred + T.ones_like(y, dtype=theano.config.floatX), T.constant(2, dtype='int32'))) test_model_all = theano.function([x, y, mair ##mairesse_change ] , [test_error, true_p, false_p, false_n, svm_data], allow_input_downcast=True) test_batches = test_set_x.shape[0] / batch_size; # start training over mini-batches print('... training') epoch = 0 best_val_perf = 0 val_perf = 0 test_perf = 0 fscore = 0 cost_epoch = 0 while (epoch < n_epochs): start_time = time.time() epoch = epoch + 1 if shuffle_batch: for minibatch_index in np.random.permutation(range(n_train_batches)): cost_epoch = train_model(minibatch_index) set_zero(zero_vec) else: for minibatch_index in range(int(n_train_batches)): cost_epoch = train_model(minibatch_index) set_zero(zero_vec) train_losses = [test_model(i) for i in range(int(n_train_batches))] train_perf = 1 - np.mean([loss[0] for loss in train_losses]) val_losses = [val_model(i) for i in range(int(n_val_batches))] val_perf = 1 - np.mean(val_losses) epoch_perf = 'epoch: %i, training time: %.2f secs, train perf: %.2f %%, val perf: %.2f %%' % ( epoch, time.time() - start_time, train_perf * 100., val_perf * 100.) print(epoch_perf) ofile.write(epoch_perf + "\n") ofile.flush() if val_perf >= best_val_perf: best_val_perf = val_perf test_loss_list = [test_model_all(test_set_x[idx * batch_size:(idx + 1) * batch_size], test_set_y[idx * batch_size:(idx + 1) * batch_size], test_set_m[idx * batch_size:(idx + 1) * batch_size] ##mairesse_change ) for idx in range(int(test_batches))] if test_set_x.shape[0] > test_batches * batch_size: test_loss_list.append( test_model_all(test_set_x[int(test_batches * batch_size):], test_set_y[int(test_batches * batch_size):], test_set_m[int(test_batches * batch_size):] ##mairesse_change )) test_loss_list_temp = test_loss_list test_loss_list = np.asarray([t[:-1] for t in test_loss_list]) test_loss = np.sum(test_loss_list[:, 0]) / float(test_set_x.shape[0]) test_perf = 1 - test_loss tp = np.sum(test_loss_list[:, 1]) fp = np.sum(test_loss_list[:, 2]) fn = np.sum(test_loss_list[:, 3]) tn = test_set_x.shape[0] - (tp + fp + fn) fscore = np.mean([2 * tp / float(2 * tp + fp + fn), 2 * tn / float(2 * tn + fp + fn)]) svm_test = np.concatenate([t[-1] for t in test_loss_list_temp], axis=0) svm_train = np.concatenate([t[1] for t in train_losses], axis=0) output = "Test result: accu: " + str(test_perf) + ", macro_fscore: " + str(fscore) + "\ntp: " + str( tp) + " tn:" + str(tn) + " fp: " + str(fp) + " fn: " + str(fn) print(output) ofile.write(output + "\n") ofile.flush() # dump train and test features pickle.dump(svm_test, open("cvte" + str(attr) + str(cv) + ".p", "wb")) pickle.dump(svm_train, open("cvtr" + str(attr) + str(cv) + ".p", "wb")) updated_epochs = refresh_epochs() if updated_epochs != None and n_epochs != updated_epochs: n_epochs = updated_epochs print('Epochs updated to ' + str(n_epochs)) return test_perf, fscore def refresh_epochs(): try: f = open('n_epochs', 'r') except Exception: return None try: n = int(f.readline().strip()) except Exception: f.close() return None f.close() return n def shared_dataset(data_xy, borrow=True): """ Function that loads the dataset into shared variables The reason we store our dataset in shared variables is to allow Theano to copy it into the GPU memory (when code is run on GPU). Since copying data into the GPU is slow, copying a minibatch everytime is needed (the default behaviour if the data is not in a shared variable) would lead to a large decrease in performance. """ data_x, data_y, data_m = data_xy shared_x = theano.shared(np.asarray(data_x, dtype=theano.config.floatX), borrow=borrow) shared_y = theano.shared(np.asarray(data_y, dtype=theano.config.floatX), borrow=borrow) shared_m = theano.shared(np.asarray(data_m, dtype=theano.config.floatX), borrow=borrow) return shared_x, T.cast(shared_y, 'int32'), shared_m def sgd_updates_adadelta(params, cost, rho=0.95, epsilon=1e-6, norm_lim=9, word_vec_name='Words'): """ adadelta update rule, mostly from https://groups.google.com/forum/#!topic/pylearn-dev/3QbKtCumAW4 (for Adadelta) """ updates = OrderedDict({}) exp_sqr_grads = OrderedDict({}) exp_sqr_ups = OrderedDict({}) gparams = [] for param in params: empty = np.zeros_like(param.get_value(), dtype=theano.config.floatX) exp_sqr_grads[param] = theano.shared(value=as_floatX(empty), name="exp_grad_%s" % param.name) gp = T.grad(cost, param) exp_sqr_ups[param] = theano.shared(value=as_floatX(empty), name="exp_grad_%s" % param.name) gparams.append(gp) for param, gp in list(zip(params, gparams)): exp_sg = exp_sqr_grads[param] exp_su = exp_sqr_ups[param] up_exp_sg = rho * exp_sg + (1 - rho) * T.sqr(gp) updates[exp_sg] = up_exp_sg step = -(T.sqrt(exp_su + epsilon) / T.sqrt(up_exp_sg + epsilon)) * gp updates[exp_su] = rho * exp_su + (1 - rho) * T.sqr(step) stepped_param = param + step updates[param] = stepped_param return updates def as_floatX(variable): if isinstance(variable, float): return np.cast[theano.config.floatX](variable) if isinstance(variable, np.ndarray): return np.cast[theano.config.floatX](variable) return theano.tensor.cast(variable, theano.config.floatX) def safe_update(dict_to, dict_from): """ re-make update dictionary for safe updating """ for key, val in dict(dict_from).iteritems(): if key in dict_to: raise KeyError(key) dict_to[key] = val return dict_to def get_idx_from_sent(status, word_idx_map, charged_words, max_l=51, max_s=200, k=300, filter_h=5): """ Transforms sentence into a list of indices. Pad with zeroes. """ x = [] pad = filter_h - 1 length = len(status) pass_one = True while len(x) == 0: for i in range(length): words = status[i].split() if pass_one: words_set = set(words) if len(charged_words.intersection(words_set)) == 0: continue else: if np.random.randint(0, 2) == 0: continue y = [] for i in range(int(pad)): y.append(0) for word in words: if word in word_idx_map: y.append(word_idx_map[word]) while len(y) < max_l + 2 * pad: y.append(0) x.append(y) pass_one = False if len(x) < max_s: x.extend([[0] * (max_l + 2 * pad)] * (max_s - len(x))) return x def make_idx_data_cv(revs, word_idx_map, mairesse, charged_words, cv, per_attr=0, max_l=51, max_s=200, k=300, filter_h=5): """ Transforms sentences into a 2-d matrix. """ trainX, testX, trainY, testY, mTrain, mTest = [], [], [], [], [], [] for rev in revs: sent = get_idx_from_sent(rev["text"], word_idx_map, charged_words, max_l, max_s, k, filter_h) if rev["split"] == cv: testX.append(sent) testY.append(rev['y' + str(per_attr)]) mTest.append(mairesse[rev["user"]]) else: trainX.append(sent) trainY.append(rev['y' + str(per_attr)]) mTrain.append(mairesse[rev["user"]]) trainX = np.array(trainX, dtype="float32") testX = np.array(testX, dtype="float32") trainY = np.array(trainY, dtype="float32") testY = np.array(testY, dtype="float32") mTrain = np.array(mTrain, dtype=theano.config.floatX) mTest = np.array(mTest, dtype=theano.config.floatX) return [trainX, trainY, testX, testY, mTrain, mTest] if __name__ == "__main__": print("loading data...: floatx:" + theano.config.floatX), x = pickle.load(open("essays_mairesse.p", "rb")) revs, W, W2, word_idx_map, vocab, mairesse = x[0], x[1], x[2], x[3], x[4], x[5] print("data loaded!") mode = "-static" word_vectors = "-word2vec" attr = int(3) # mode = "-static" # word_vectors = "-word2vec" # attr = int("2") print ("attr: " + str(attr)) if mode == "-nonstatic": print("model architecture: CNN-non-static") non_static = True elif mode == "-static": print("model architecture: CNN-static") non_static = False exec (open("/content/drive/MyDrive/Colab Notebooks/conv_net_classes_gpu.py").read()) if word_vectors == "-rand": print("using: random vectors") U = np.float32(W2) elif word_vectors == "-word2vec": print("using: word2vec vectors") U = np.float32(W) r = range(0, 10) ofile = open('perf_output_' + str(attr) + '.txt', 'w') charged_words = [] emof = open("/content/drive/MyDrive/Colab Notebooks/Emotion_Lexicon.csv", "r", encoding='cp1252') csvf = csv.reader(emof, delimiter=',', quotechar='"') first_line = True for line in csvf: if first_line: first_line = False continue if line[11] == "1": charged_words.append(line[0]) emof.close() charged_words = set(charged_words) results = [] for i in r: datasets = make_idx_data_cv(revs, word_idx_map, mairesse, charged_words, i, attr, max_l=149, max_s=312, k=300, filter_h=3) perf, fscore = train_conv_net(datasets, U, ofile, cv=i, attr=attr, lr_decay=0.95, filter_hs=[1, 2, 3], conv_non_linear="relu", hidden_units=[200, 200, 2], shuffle_batch=True, n_epochs=50, sqr_norm_lim=9, non_static=non_static, batch_size=50, dropout_rate=[0.5, 0.5, 0.5], activations=[Sigmoid]) output = "cv: " + str(i) + ", perf: " + str(perf) + ", macro_fscore: " + str(fscore) print(output) ofile.write(output + "\n") ofile.flush() results.append([perf, fscore]) results = np.asarray(results) perf_out = 'Perf : ' + str(np.mean(results[:, 0])) fscore_out = 'Macro_Fscore : ' + str(np.mean(results[:, 1])) print(perf_out) print(fscore_out) ofile.write(perf_out + "\n" + fscore_out) ofile.close()
voici la classe conv_net_classes_gpu.py :
Code Python : Sélectionner tout - Visualiser dans une fenêtre à part
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 """ Sample code for Convolutional Neural Networks for Sentence Classification http://arxiv.org/pdf/1408.5882v2.pdf Much of the code is modified from - deeplearning.net (for ConvNet classes) - https://github.com/mdenil/dropout (for dropout) - https://groups.google.com/forum/#!topic/pylearn-dev/3QbKtCumAW4 (for Adadelta) """ import numpy import theano.tensor.shared_randomstreams import theano import theano.tensor as T # from theano.tensor.signal import downsample from theano.tensor.signal import pool from theano.tensor.nnet import conv2d as conv def ReLU(x): y = T.maximum(0.0, x) return(y) def Sigmoid(x): y = T.nnet.sigmoid(x) return(y) def Tanh(x): y = T.tanh(x) return(y) def Iden(x): y = x return(y) class HiddenLayer(object): """ Class for HiddenLayer """ def __init__(self, rng, input, n_in, n_out, activation, W=None, b=None, use_bias=False): self.input = input self.activation = activation if W is None: if activation.__name__ == "ReLU": W_values = numpy.asarray(0.01 * rng.standard_normal(size=(n_in, n_out)), dtype=theano.config.floatX) else: W_values = numpy.asarray(rng.uniform(low=-numpy.sqrt(6. / (n_in + n_out)), high=numpy.sqrt(6. / (n_in + n_out)), size=(n_in, n_out)), dtype=theano.config.floatX) W = theano.shared(value=W_values, name='W') if b is None: b_values = numpy.zeros((n_out,), dtype=theano.config.floatX) b = theano.shared(value=b_values, name='b') self.W = W self.b = b if use_bias: lin_output = T.dot(input, self.W) + self.b else: lin_output = T.dot(input, self.W) self.output = (lin_output if activation is None else activation(lin_output)) # parameters of the model if use_bias: self.params = [self.W, self.b] else: self.params = [self.W] def _dropout_from_layer(rng, layer, p): """p is the probablity of dropping a unit """ srng = theano.tensor.shared_randomstreams.RandomStreams(rng.randint(999999)) # p=1-p because 1's indicate keep and p is prob of dropping mask = srng.binomial(n=1, p=1-p, size=layer.shape) # The cast is important because # int * float32 = float64 which pulls things off the gpu output = layer * T.cast(mask, theano.config.floatX) return output class DropoutHiddenLayer(HiddenLayer): def __init__(self, rng, input, n_in, n_out, activation, dropout_rate, use_bias, W=None, b=None): super(DropoutHiddenLayer, self).__init__( rng=rng, input=input, n_in=n_in, n_out=n_out, W=W, b=b, activation=activation, use_bias=use_bias) self.output = _dropout_from_layer(rng, self.output, p=dropout_rate) class MLPDropout(object): """A multilayer perceptron with dropout""" def __init__(self,rng,input,layer_sizes,dropout_rates,activations,use_bias=True): #rectified_linear_activation = lambda x: T.maximum(0.0, x) # Set up all the hidden layers self.weight_matrix_sizes = list(zip(layer_sizes, layer_sizes[1:])) self.layers = [] self.dropout_layers = [] self.activations = activations next_layer_input = input #first_layer = True # dropout the input next_dropout_layer_input = _dropout_from_layer(rng, input, p=dropout_rates[0]) layer_counter = 0 for n_in, n_out in self.weight_matrix_sizes[:-1]: next_dropout_layer = DropoutHiddenLayer(rng=rng, input=next_dropout_layer_input, activation=activations[layer_counter], n_in=n_in, n_out=n_out, use_bias=use_bias, dropout_rate=dropout_rates[layer_counter]) self.dropout_layers.append(next_dropout_layer) next_dropout_layer_input = next_dropout_layer.output # Reuse the parameters from the dropout layer here, in a different # path through the graph. next_layer = HiddenLayer(rng=rng, input=next_layer_input, activation=activations[layer_counter], # scale the weight matrix W with (1-p) W=next_dropout_layer.W * (1 - dropout_rates[layer_counter]), b=next_dropout_layer.b, n_in=n_in, n_out=n_out, use_bias=use_bias) self.layers.append(next_layer) next_layer_input = next_layer.output #first_layer = False layer_counter += 1 # Set up the output layer n_in, n_out = self.weight_matrix_sizes[-1] dropout_output_layer = LogisticRegression( input=next_dropout_layer_input, n_in=n_in, n_out=n_out) self.dropout_layers.append(dropout_output_layer) # Again, reuse paramters in the dropout output. output_layer = LogisticRegression( input=next_layer_input, # scale the weight matrix W with (1-p) W=dropout_output_layer.W * (1 - dropout_rates[-1]), b=dropout_output_layer.b, n_in=n_in, n_out=n_out) self.layers.append(output_layer) # Use the negative log likelihood of the logistic regression layer as # the objective. self.dropout_negative_log_likelihood = self.dropout_layers[-1].negative_log_likelihood self.dropout_errors = self.dropout_layers[-1].errors self.negative_log_likelihood = self.layers[-1].negative_log_likelihood self.errors = self.layers[-1].errors # Grab all the parameters together. self.params = [ param for layer in self.dropout_layers for param in layer.params ] def predict(self, new_data): next_layer_input = new_data for i,layer in enumerate(self.layers): if i<len(self.layers)-1: next_layer_input = self.activations[i](T.dot(next_layer_input,layer.W) + layer.b) else: p_y_given_x = T.nnet.softmax(T.dot(next_layer_input, layer.W) + layer.b) y_pred = T.argmax(p_y_given_x, axis=1) return y_pred def predict_p(self, new_data): next_layer_input = new_data for i,layer in enumerate(self.layers): if i<len(self.layers)-1: next_layer_input = self.activations[i](T.dot(next_layer_input,layer.W) + layer.b) else: p_y_given_x = T.nnet.softmax(T.dot(next_layer_input, layer.W) + layer.b) return p_y_given_x class MLP(object): """Multi-Layer Perceptron Class A multilayer perceptron is a feedforward artificial neural network model that has one layer or more of hidden units and nonlinear activations. Intermediate layers usually have as activation function tanh or the sigmoid function (defined here by a ``HiddenLayer`` class) while the top layer is a softamx layer (defined here by a ``LogisticRegression`` class). """ def __init__(self, rng, input, n_in, n_hidden, n_out): """Initialize the parameters for the multilayer perceptron :type rng: numpy.random.RandomState :param rng: a random number generator used to initialize weights :type input: theano.tensor.TensorType :param input: symbolic variable that describes the input of the architecture (one minibatch) :type n_in: int :param n_in: number of input units, the dimension of the space in which the datapoints lie :type n_hidden: int :param n_hidden: number of hidden units :type n_out: int :param n_out: number of output units, the dimension of the space in which the labels lie """ # Since we are dealing with a one hidden layer MLP, this will translate # into a HiddenLayer with a tanh activation function connected to the # LogisticRegression layer; the activation function can be replaced by # sigmoid or any other nonlinear function self.hiddenLayer = HiddenLayer(rng=rng, input=input, n_in=n_in, n_out=n_hidden, activation=T.tanh) # The logistic regression layer gets as input the hidden units # of the hidden layer self.logRegressionLayer = LogisticRegression( input=self.hiddenLayer.output, n_in=n_hidden, n_out=n_out) # L1 norm ; one regularization option is to enforce L1 norm to # be small # negative log likelihood of the MLP is given by the negative # log likelihood of the output of the model, computed in the # logistic regression layer self.negative_log_likelihood = self.logRegressionLayer.negative_log_likelihood # same holds for the function computing the number of errors self.errors = self.logRegressionLayer.errors # the parameters of the model are the parameters of the two layer it is # made out of self.params = self.hiddenLayer.params + self.logRegressionLayer.params class LogisticRegression(object): """Multi-class Logistic Regression Class The logistic regression is fully described by a weight matrix :math:`W` and bias vector :math:`b`. Classification is done by projecting data points onto a set of hyperplanes, the distance to which is used to determine a class membership probability. """ def __init__(self, input, n_in, n_out, W=None, b=None): """ Initialize the parameters of the logistic regression :type input: theano.tensor.TensorType :param input: symbolic variable that describes the input of the architecture (one minibatch) :type n_in: int :param n_in: number of input units, the dimension of the space in which the datapoints lie :type n_out: int :param n_out: number of output units, the dimension of the space in which the labels lie """ # initialize with 0 the weights W as a matrix of shape (n_in, n_out) if W is None: self.W = theano.shared( value=numpy.zeros((n_in, n_out), dtype=theano.config.floatX), name='W') else: self.W = W # initialize the baises b as a vector of n_out 0s if b is None: self.b = theano.shared( value=numpy.zeros((n_out,), dtype=theano.config.floatX), name='b') else: self.b = b # compute vector of class-membership probabilities in symbolic form self.p_y_given_x = T.nnet.softmax(T.dot(input, self.W) + self.b) # compute prediction as class whose probability is maximal in # symbolic form self.y_pred = T.argmax(self.p_y_given_x, axis=1) # parameters of the model self.params = [self.W, self.b] def negative_log_likelihood(self, y): """Return the mean of the negative log-likelihood of the prediction of this model under a given target distribution. .. math:: \frac{1}{|\mathcal{D}|} \mathcal{L} (\theta=\{W,b\}, \mathcal{D}) = \frac{1}{|\mathcal{D}|} \sum_{i=0}^{|\mathcal{D}|} \log(P(Y=y^{(i)}|x^{(i)}, W,b)) \\ \ell (\theta=\{W,b\}, \mathcal{D}) :type y: theano.tensor.TensorType :param y: corresponds to a vector that gives for each example the correct label Note: we use the mean instead of the sum so that the learning rate is less dependent on the batch size """ # y.shape[0] is (symbolically) the number of rows in y, i.e., # number of examples (call it n) in the minibatch # T.arange(y.shape[0]) is a symbolic vector which will contain # [0,1,2,... n-1] T.log(self.p_y_given_x) is a matrix of # Log-Probabilities (call it LP) with one row per example and # one column per class LP[T.arange(y.shape[0]),y] is a vector # v containing [LP[0,y[0]], LP[1,y[1]], LP[2,y[2]], ..., # LP[n-1,y[n-1]]] and T.mean(LP[T.arange(y.shape[0]),y]) is # the mean (across minibatch examples) of the elements in v, # i.e., the mean log-likelihood across the minibatch. return -T.mean(T.log(self.p_y_given_x)[T.arange(y.shape[0]), y]) def errors(self, y): """Return a float representing the number of errors in the minibatch ; zero one loss over the size of the minibatch :type y: theano.tensor.TensorType :param y: corresponds to a vector that gives for each example the correct label """ # check if y has same dimension of y_pred if y.ndim != self.y_pred.ndim: raise TypeError('y should have the same shape as self.y_pred', ('y', target.type, 'y_pred', self.y_pred.type)) # check if y is of the correct datatype if y.dtype.startswith('int'): # the T.neq operator returns a vector of 0s and 1s, where 1 # represents a mistake in prediction return T.mean(T.neq(self.y_pred, y)) else: raise NotImplementedError() class LeNetConvPoolLayer(object): """Pool Layer of a convolutional network """ def __init__(self, rng, filter_shape, image_shape, poolsize=(2, 2), non_linear="tanh"): """ Allocate a LeNetConvPoolLayer with shared variable internal parameters. :type rng: numpy.random.RandomState :param rng: a random number generator used to initialize weights :type input: theano.tensor.dtensor4 :param input: symbolic image tensor, of shape image_shape :type filter_shape: tuple or list of length 4 :param filter_shape: (number of filters, num input feature maps, filter height,filter width) :type image_shape: tuple or list of length 4 :param image_shape: (batch size, num input feature maps, image height, image width) :type poolsize: tuple or list of length 2 :param poolsize: the downsampling (pooling) factor (#rows,#cols) """ # assert image_shape[1] == filter_shape[1] # self.input = input self.filter_shape = filter_shape self.image_shape = image_shape self.poolsize = poolsize self.non_linear = non_linear # there are "num input feature maps * filter height * filter width" # inputs to each hidden unit fan_in = numpy.prod(filter_shape[1:]) # each unit in the lower layer receives a gradient from: # "num output feature maps * filter height * filter width" / # pooling size fan_out = (filter_shape[0] * numpy.prod(filter_shape[2:]) /numpy.prod(poolsize)) # initialize weights with random weights if self.non_linear=="none" or self.non_linear=="relu": self.W = theano.shared(numpy.asarray(rng.uniform(low=-0.01,high=0.01,size=filter_shape), dtype=theano.config.floatX),borrow=True,name="W_conv") else: W_bound = numpy.sqrt(6. / (fan_in + fan_out)) self.W = theano.shared(numpy.asarray(rng.uniform(low=-W_bound, high=W_bound, size=filter_shape), dtype=theano.config.floatX),borrow=True,name="W_conv") b_values = numpy.zeros((filter_shape[0],), dtype=theano.config.floatX) self.b = theano.shared(value=b_values, borrow=True, name="b_conv") self.params = [self.W, self.b] def set_input(self, input): # convolve input feature maps with filters conv_out = conv(input=input, filters=self.W,filter_shape=self.filter_shape, image_shape=self.image_shape) if self.non_linear=="tanh": conv_out_tanh = T.tanh(conv_out + self.b.dimshuffle('x', 0, 'x', 'x')) output = pool.pool_2d(input=conv_out_tanh, ds=self.poolsize, ignore_border=True) elif self.non_linear=="relu": conv_out_tanh = ReLU(conv_out + self.b.dimshuffle('x', 0, 'x', 'x')) output = pool.pool_2d(input=conv_out_tanh, ds=self.poolsize, ignore_border=True) else: pooled_out = pool.pool_2d(input=conv_out, ds=self.poolsize, ignore_border=True) output = pooled_out + self.b.dimshuffle('x', 0, 'x', 'x') return output def predict(self, new_data, batch_size): """ predict for new data """ img_shape = None#(batch_size, 1, self.image_shape[2], self.image_shape[3]) conv_out = conv.conv2d(input=new_data, filters=self.W, filter_shape=self.filter_shape, image_shape=img_shape) if self.non_linear=="tanh": conv_out_tanh = T.tanh(conv_out + self.b.dimshuffle('x', 0, 'x', 'x')) output = pool.pool_2d(input=conv_out_tanh, ds=self.poolsize, ignore_border=True) if self.non_linear=="relu": conv_out_tanh = ReLU(conv_out + self.b.dimshuffle('x', 0, 'x', 'x')) output = pool.pool_2d(input=conv_out_tanh, ds=self.poolsize, ignore_border=True) else: pooled_out = pool.pool_2d(input=conv_out, ds=self.poolsize, ignore_border=True) output = pooled_out + self.b.dimshuffle('x', 0, 'x', 'x') return output
Partager