Commit 5ee737bf2b2951378f2adeeffd7f1abc7b7670b5
1 parent
6060efac
Porownanie modelu MLP na danych polskich i angielskich z tree-lstmem. Pierwsza w…
…ersja i eksperymenty sieci z parametryzajcja krawedziami
Showing
24 changed files
with
5173 additions
and
479 deletions
Too many changes to show.
To preserve performance only 19 of 24 files are displayed.
main_for_experiments_on_polish_data_LSTM.py
0 → 100644
1 | +import numpy as np | |
2 | +import time | |
3 | +import sys | |
4 | +import subprocess | |
5 | +import os | |
6 | +import random | |
7 | + | |
8 | +#from modules.data import load | |
9 | +from modules.rnn.LSTM_models import * | |
10 | +#from modules.metrics.accuracy import conlleval | |
11 | +from modules.utils.tools import load_stanford_data4, shuffle | |
12 | + | |
13 | +from theano import pp | |
14 | + | |
15 | +import theano.tensor as T | |
16 | +import theano | |
17 | +from theano.sandbox.rng_mrg import MRG_RandomStreams #as MRG_RandomStreams | |
18 | + | |
19 | +import itertools | |
20 | + | |
21 | +import os.path | |
22 | +import pickle | |
23 | + | |
24 | +from collections import Counter | |
25 | + | |
26 | + | |
27 | + | |
28 | +from theano import tensor as T, printing | |
29 | +from collections import OrderedDict | |
30 | +from theano.ifelse import ifelse | |
31 | + | |
32 | +from keras.preprocessing import sequence as seq | |
33 | + | |
34 | +dataType = 'int64' | |
35 | + | |
36 | + | |
37 | +if __name__ == '__main__': | |
38 | + | |
39 | + | |
40 | + | |
41 | + w2v_DIM = "300" | |
42 | + | |
43 | + | |
44 | + | |
45 | + file_with_filtered_embeddings = "embeddings/embedding_and_words2ids_dim"+w2v_DIM+"_polish.pkl" | |
46 | + if not os.path.exists(file_with_filtered_embeddings): | |
47 | + print("Cannot find file with only needed embeddings. We use 'filter_embeddings' in order to create it.") | |
48 | + filter_embeddings(["data/dane_polskie/train/train_labels.txt", "data/dane_polskie/train/train_parents.txt","data/dane_polskie/train/train_sentence.txt", | |
49 | + "data/dane_polskie/dev/dev_labels.txt", "data/dane_polskie/dev/dev_parents.txt","data/dane_polskie/dev/dev_sentence.txt", | |
50 | + "data/dane_polskie/test/test_labels.txt", "data/dane_polskie/test/test_parents.txt","data/dane_polskie/test/test_sentence.txt"], | |
51 | + | |
52 | + "/home/norbert/Doktorat/clarin2sent/deeptagger/embeddings/w2v_allwiki_nkjpfull_"+w2v_DIM+".txt", | |
53 | + file_with_filtered_embeddings) | |
54 | + | |
55 | + | |
56 | + s = {'lr':0.002, | |
57 | + 'nepochs':40, | |
58 | + 'seed':345, | |
59 | + 'nc':3 # number of y classes | |
60 | + } | |
61 | + batch_size = 1 | |
62 | + | |
63 | + | |
64 | + for h_dim in [100, 150]: | |
65 | + | |
66 | + np.random.seed(s['seed']) | |
67 | + random.seed(s['seed']) | |
68 | + | |
69 | + | |
70 | + rnn = LSTM_1( h_dim, | |
71 | + nc = s['nc'], | |
72 | + w2v_model_path = file_with_filtered_embeddings, #sciezka do pliku z embeddingami | |
73 | + max_phrase_length = 60 ) | |
74 | + | |
75 | + | |
76 | + train_data = load_stanford_data4("data/dane_polskie/train/train_labels.txt", "data/dane_polskie/train/train_parents.txt","data/dane_polskie/train/train_sentence.txt",rnn.words2ids,True,batch_size,s['nc']) | |
77 | + train_data_check = train_data | |
78 | + dev_data = load_stanford_data4("data/dane_polskie/dev/dev_labels.txt", "data/dane_polskie/dev/dev_parents.txt","data/dane_polskie/dev/dev_sentence.txt",rnn.words2ids,False,0,s['nc']) | |
79 | + test_data = load_stanford_data4("data/dane_polskie/test/test_labels.txt", "data/dane_polskie/test/test_parents.txt","data/dane_polskie/test/test_sentence.txt",rnn.words2ids,False,0,s['nc']) | |
80 | + | |
81 | + n_train = len(train_data) | |
82 | + n_dev = len(dev_data) | |
83 | + n_test = len(test_data) | |
84 | + | |
85 | + print "" | |
86 | + #print "model 56 : h_dim = ", h_dim, "h2_dim = ", h2_dim, "h3_dim = ", h3_dim, " learning rate = ", s['lr']#, "dropout rate: ", dropout_rate | |
87 | + print "model LSTM_` : " , "h_dim = ", h_dim | |
88 | + print "" | |
89 | + | |
90 | + best_prediction_valid_all = 0 | |
91 | + best_prediction_test_all = 0 | |
92 | + best_prediction_test_root = 0 | |
93 | + early_stop = 0 | |
94 | + | |
95 | + | |
96 | + tic = time.time() | |
97 | + | |
98 | + for e in xrange(s['nepochs']): | |
99 | + | |
100 | + #if e >= 1: | |
101 | + # s['lr'] = 0.8 * s['lr'] | |
102 | + | |
103 | + if early_stop == 10: | |
104 | + break | |
105 | + | |
106 | + | |
107 | + # shuffle | |
108 | + shuffle([train_data], s['seed']) | |
109 | + | |
110 | + for i in range(n_train): | |
111 | + rnn.train(train_data[i][0],train_data[i][1], train_data[i][2], train_data[i][3], s['lr']) | |
112 | + | |
113 | + | |
114 | + | |
115 | + # Dev: | |
116 | + counts_dev = np.zeros((s['nc'],s['nc']),dtype='int') | |
117 | + counts_dev_root = np.zeros((s['nc'],s['nc']),dtype='int') | |
118 | + for ii in range(n_dev): | |
119 | + pred = rnn.classify(dev_data[ii][0],dev_data[ii][1], dev_data[ii][3]) | |
120 | + for j in range(len(pred)): | |
121 | + counts_dev[pred[j], dev_data[ii][2][j]] += 1 | |
122 | + counts_dev_root[pred[-1], dev_data[ii][2][-1]] += 1 | |
123 | + | |
124 | + | |
125 | + # Test: | |
126 | + counts_test = np.zeros((s['nc'],s['nc']),dtype='int') | |
127 | + counts_test_root = np.zeros((s['nc'],s['nc']),dtype='int') | |
128 | + for i in range(n_test): | |
129 | + pred = rnn.classify(test_data[i][0],test_data[i][1], test_data[i][3]) | |
130 | + for j in range(len(pred)): | |
131 | + counts_test[pred[j], test_data[i][2][j]] += 1 | |
132 | + counts_test_root[pred[-1], test_data[i][2][-1]] += 1 | |
133 | + | |
134 | + # Train | |
135 | + counts = np.zeros((s['nc'],s['nc']),dtype='int') | |
136 | + counts_root = np.zeros((s['nc'],s['nc']),dtype='int') | |
137 | + for i in range(len(train_data_check)): | |
138 | + | |
139 | + if i % 1 == 0: #sprawdzamy dopasowanie na 1/100 zbioru zeby oszczedzic czas | |
140 | + pred = rnn.classify(train_data_check[i][0],train_data_check[i][1], train_data_check[i][3]) | |
141 | + for j in range(len(pred)): | |
142 | + counts[pred[j], train_data_check[i][2][j]] += 1 | |
143 | + counts_root[pred[-1], train_data_check[i][2][-1]] += 1 | |
144 | + | |
145 | + print("epoch: ", e, | |
146 | + "V all: ", "%0.2f" % (100 * np.diag(counts_dev).sum()/float(counts_dev.sum())), | |
147 | + " Test all: ", "%0.2f" % (100 * np.diag(counts_test).sum()/float(counts_test.sum())), | |
148 | + "V root: ", "%0.2f" % (100 * np.diag(counts_dev_root).sum()/float(counts_dev_root.sum())), | |
149 | + " Test root: ", "%0.2f" % (100 * np.diag(counts_test_root).sum()/float(counts_test_root.sum())), | |
150 | + " Train: ", "%0.2f" % (100 * np.diag(counts).sum()/float(counts.sum())), | |
151 | + " Train root: ", "%0.2f" % (100 * np.diag(counts_root).sum()/float(counts_root.sum())) | |
152 | + ) | |
153 | + | |
154 | + | |
155 | + if np.diag(counts_dev).sum()/float(counts_dev.sum()) > best_prediction_valid_all: | |
156 | + best_prediction_valid_all = np.diag(counts_dev).sum()/float(counts_dev.sum()) | |
157 | + best_prediction_test_all = np.diag(counts_test).sum()/float(counts_test.sum()) | |
158 | + best_prediction_test_root = np.diag(counts_test_root).sum()/float(counts_test_root.sum()) | |
159 | + | |
160 | + early_stop = 0 | |
161 | + else: | |
162 | + early_stop = early_stop + 1 | |
163 | + | |
164 | + | |
165 | + print("Best valid: ", "%0.2f" % (100 * best_prediction_valid_all)," Test all: ","%0.2f" % (100 * best_prediction_test_all),"Test root: ","%0.2f" % (100 * best_prediction_test_root), " time: ", time.time()-tic) | |
166 | + | |
... | ... |
main_for_experiments_on_polish_data_MLP2.py
0 → 100644
1 | +import numpy as np | |
2 | +import time | |
3 | +import sys | |
4 | +import subprocess | |
5 | +import os | |
6 | +import random | |
7 | + | |
8 | +#from modules.data import load | |
9 | +from modules.rnn.models_with_relations import * | |
10 | +#from modules.metrics.accuracy import conlleval | |
11 | +from modules.utils.tools import load_stanford_data6, shuffle | |
12 | + | |
13 | +from theano import pp | |
14 | + | |
15 | +import theano.tensor as T | |
16 | +import theano | |
17 | +from theano.sandbox.rng_mrg import MRG_RandomStreams #as MRG_RandomStreams | |
18 | + | |
19 | +import itertools | |
20 | + | |
21 | +import os.path | |
22 | +import pickle | |
23 | + | |
24 | +from collections import Counter | |
25 | + | |
26 | + | |
27 | + | |
28 | +from theano import tensor as T, printing | |
29 | +from collections import OrderedDict | |
30 | +from theano.ifelse import ifelse | |
31 | + | |
32 | +from keras.preprocessing import sequence as seq | |
33 | + | |
34 | +dataType = 'int64' | |
35 | + | |
36 | + | |
37 | +if __name__ == '__main__': | |
38 | + | |
39 | + | |
40 | + | |
41 | + w2v_DIM = "300" | |
42 | + | |
43 | + | |
44 | + | |
45 | + file_with_filtered_embeddings = "embeddings/embedding_and_words2ids_dim"+w2v_DIM+"_polish.pkl" | |
46 | + if not os.path.exists(file_with_filtered_embeddings): | |
47 | + print("Cannot find file with only needed embeddings. We use 'filter_embeddings' in order to create it.") | |
48 | + filter_embeddings(["data/dane_polskie/train/train_labels.txt", "data/dane_polskie/train/train_parents.txt","data/dane_polskie/train/train_sentence.txt", | |
49 | + "data/dane_polskie/dev/dev_labels.txt", "data/dane_polskie/dev/dev_parents.txt","data/dane_polskie/dev/dev_sentence.txt", | |
50 | + "data/dane_polskie/test/test_labels.txt", "data/dane_polskie/test/test_parents.txt","data/dane_polskie/test/test_sentence.txt"], | |
51 | + | |
52 | + "/home/norbert/Doktorat/clarin2sent/deeptagger/embeddings/w2v_allwiki_nkjpfull_"+w2v_DIM+".txt", | |
53 | + file_with_filtered_embeddings) | |
54 | + | |
55 | + | |
56 | + s = {'lr':0.002, | |
57 | + 'nepochs':40, | |
58 | + 'seed':345, | |
59 | + 'nc':3 # number of y classes | |
60 | + } | |
61 | + batch_size = 1 | |
62 | + | |
63 | + | |
64 | + for h_dim in [50]: | |
65 | + | |
66 | + np.random.seed(s['seed']) | |
67 | + random.seed(s['seed']) | |
68 | + | |
69 | + | |
70 | + rnn = MLP_2_1( h_dim, h_dim, | |
71 | + nc = s['nc'], | |
72 | + w2v_model_path = file_with_filtered_embeddings, #sciezka do pliku z embeddingami | |
73 | + max_phrase_length = 60 ) | |
74 | + | |
75 | + | |
76 | + train_data = load_stanford_data4("data/dane_polskie/train/train_labels.txt", | |
77 | + "data/dane_polskie/train/train_parents.txt", | |
78 | + "data/dane_polskie/train/train_sentence.txt", | |
79 | + | |
80 | +############################################################################################# | |
81 | + | |
82 | + NIE MA RELACJI DLA POLSKICH DANYCH | |
83 | + | |
84 | +############################################################################################# | |
85 | + | |
86 | + rnn.words2ids,True,batch_size,s['nc']) | |
87 | + train_data_check = train_data | |
88 | + dev_data = load_stanford_data4("data/dane_polskie/dev/dev_labels.txt", "data/dane_polskie/dev/dev_parents.txt","data/dane_polskie/dev/dev_sentence.txt",rnn.words2ids,False,0,s['nc']) | |
89 | + test_data = load_stanford_data4("data/dane_polskie/test/test_labels.txt", "data/dane_polskie/test/test_parents.txt","data/dane_polskie/test/test_sentence.txt",rnn.words2ids,False,0,s['nc']) | |
90 | + | |
91 | + n_train = len(train_data) | |
92 | + n_dev = len(dev_data) | |
93 | + n_test = len(test_data) | |
94 | + | |
95 | + print "" | |
96 | + #print "model 56 : h_dim = ", h_dim, "h2_dim = ", h2_dim, "h3_dim = ", h3_dim, " learning rate = ", s['lr']#, "dropout rate: ", dropout_rate | |
97 | + print "model LSTM_` : " , "h_dim = ", h_dim | |
98 | + print "" | |
99 | + | |
100 | + best_prediction_valid_all = 0 | |
101 | + best_prediction_test_all = 0 | |
102 | + best_prediction_test_root = 0 | |
103 | + early_stop = 0 | |
104 | + | |
105 | + | |
106 | + tic = time.time() | |
107 | + | |
108 | + for e in xrange(s['nepochs']): | |
109 | + | |
110 | + #if e >= 1: | |
111 | + # s['lr'] = 0.8 * s['lr'] | |
112 | + | |
113 | + if early_stop == 10: | |
114 | + break | |
115 | + | |
116 | + | |
117 | + # shuffle | |
118 | + shuffle([train_data], s['seed']) | |
119 | + | |
120 | + for i in range(n_train): | |
121 | + rnn.train(train_data[i][0],train_data[i][1], train_data[i][2], train_data[i][3], s['lr']) | |
122 | + | |
123 | + | |
124 | + | |
125 | + # Dev: | |
126 | + counts_dev = np.zeros((s['nc'],s['nc']),dtype='int') | |
127 | + counts_dev_root = np.zeros((s['nc'],s['nc']),dtype='int') | |
128 | + for ii in range(n_dev): | |
129 | + pred = rnn.classify(dev_data[ii][0],dev_data[ii][1], dev_data[ii][3]) | |
130 | + for j in range(len(pred)): | |
131 | + counts_dev[pred[j], dev_data[ii][2][j]] += 1 | |
132 | + counts_dev_root[pred[-1], dev_data[ii][2][-1]] += 1 | |
133 | + | |
134 | + | |
135 | + # Test: | |
136 | + counts_test = np.zeros((s['nc'],s['nc']),dtype='int') | |
137 | + counts_test_root = np.zeros((s['nc'],s['nc']),dtype='int') | |
138 | + for i in range(n_test): | |
139 | + pred = rnn.classify(test_data[i][0],test_data[i][1], test_data[i][3]) | |
140 | + for j in range(len(pred)): | |
141 | + counts_test[pred[j], test_data[i][2][j]] += 1 | |
142 | + counts_test_root[pred[-1], test_data[i][2][-1]] += 1 | |
143 | + | |
144 | + # Train | |
145 | + counts = np.zeros((s['nc'],s['nc']),dtype='int') | |
146 | + counts_root = np.zeros((s['nc'],s['nc']),dtype='int') | |
147 | + for i in range(len(train_data_check)): | |
148 | + | |
149 | + if i % 1 == 0: #sprawdzamy dopasowanie na 1/100 zbioru zeby oszczedzic czas | |
150 | + pred = rnn.classify(train_data_check[i][0],train_data_check[i][1], train_data_check[i][3]) | |
151 | + for j in range(len(pred)): | |
152 | + counts[pred[j], train_data_check[i][2][j]] += 1 | |
153 | + counts_root[pred[-1], train_data_check[i][2][-1]] += 1 | |
154 | + | |
155 | + print("epoch: ", e, | |
156 | + "V all: ", "%0.2f" % (100 * np.diag(counts_dev).sum()/float(counts_dev.sum())), | |
157 | + " Test all: ", "%0.2f" % (100 * np.diag(counts_test).sum()/float(counts_test.sum())), | |
158 | + "V root: ", "%0.2f" % (100 * np.diag(counts_dev_root).sum()/float(counts_dev_root.sum())), | |
159 | + " Test root: ", "%0.2f" % (100 * np.diag(counts_test_root).sum()/float(counts_test_root.sum())), | |
160 | + " Train: ", "%0.2f" % (100 * np.diag(counts).sum()/float(counts.sum())), | |
161 | + " Train root: ", "%0.2f" % (100 * np.diag(counts_root).sum()/float(counts_root.sum())) | |
162 | + ) | |
163 | + | |
164 | + | |
165 | + if np.diag(counts_dev).sum()/float(counts_dev.sum()) > best_prediction_valid_all: | |
166 | + best_prediction_valid_all = np.diag(counts_dev).sum()/float(counts_dev.sum()) | |
167 | + best_prediction_test_all = np.diag(counts_test).sum()/float(counts_test.sum()) | |
168 | + best_prediction_test_root = np.diag(counts_test_root).sum()/float(counts_test_root.sum()) | |
169 | + | |
170 | + early_stop = 0 | |
171 | + else: | |
172 | + early_stop = early_stop + 1 | |
173 | + | |
174 | + | |
175 | + print("Best valid: ", "%0.2f" % (100 * best_prediction_valid_all)," Test all: ","%0.2f" % (100 * best_prediction_test_all),"Test root: ","%0.2f" % (100 * best_prediction_test_root), " time: ", time.time()-tic) | |
176 | + | |
... | ... |
main_for_experiments_on_sst_MLP2.py
0 → 100644
1 | +import numpy as np | |
2 | +import time | |
3 | +import sys | |
4 | +import subprocess | |
5 | +import os | |
6 | +import random | |
7 | + | |
8 | +#from modules.data import load | |
9 | +from modules.rnn.models_with_relations import * | |
10 | +#from modules.metrics.accuracy import conlleval | |
11 | +from modules.utils.tools import load_stanford_data6, shuffle | |
12 | + | |
13 | +from theano import pp | |
14 | + | |
15 | +import theano.tensor as T | |
16 | +import theano | |
17 | +from theano.sandbox.rng_mrg import MRG_RandomStreams #as MRG_RandomStreams | |
18 | + | |
19 | +import itertools | |
20 | + | |
21 | +import os.path | |
22 | +import pickle | |
23 | + | |
24 | +from collections import Counter | |
25 | + | |
26 | + | |
27 | + | |
28 | +from theano import tensor as T, printing | |
29 | +from collections import OrderedDict | |
30 | +from theano.ifelse import ifelse | |
31 | + | |
32 | +from keras.preprocessing import sequence as seq | |
33 | + | |
34 | +dataType = 'int64' | |
35 | + | |
36 | + | |
37 | + | |
38 | +if __name__ == '__main__': | |
39 | + | |
40 | + #theano.config.floatX = 'float64' | |
41 | + | |
42 | + file_with_filtered_embeddings = "embeddings/embedding_and_words2ids.pkl" | |
43 | + if not os.path.exists(file_with_filtered_embeddings): | |
44 | + print("Cannot find file with only needed embeddings. We use 'filter_embeddings' in order to create it.") | |
45 | + filter_embeddings(["data/sst/train/dlabels.txt", "data/sst/train/dparents.txt","data/sst/train/sents.toks", "data/sst/dev/dlabels.txt", "data/sst/dev/dparents.txt","data/sst/dev/sents.toks", "data/sst/test/dlabels.txt", "data/sst/test/dparents.txt","data/sst/test/sents.toks"], | |
46 | + | |
47 | + "/home/norbert/Doktorat/clarin2sent/treelstm/data/glove/glove.840B.300d.txt", | |
48 | + file_with_filtered_embeddings) | |
49 | + | |
50 | + | |
51 | + batch_size = 1 | |
52 | + | |
53 | + s = {'lr':0.002, | |
54 | + 'nepochs':30, | |
55 | + 'seed':345, | |
56 | + 'nc':5 # number of y classes | |
57 | + } | |
58 | + | |
59 | + | |
60 | + batch_size = 1 | |
61 | + | |
62 | + | |
63 | + | |
64 | + for ne_dim, nchd_dim, nh2_dim, number_of_relations in [(50,50, 50, 5),(100,100, 100, 5),(50,50, 50, 10),(100,100, 100, 10),(200,200, 100, 5)]: | |
65 | + | |
66 | + np.random.seed(s['seed']) | |
67 | + random.seed(s['seed']) | |
68 | + | |
69 | + | |
70 | + rnn = MLP_2_2( ne = ne_dim, nchd = nchd_dim, nh2 = nh2_dim, | |
71 | + nc = s['nc'], | |
72 | + w2v_model_path = file_with_filtered_embeddings, #sciezka do pliku z embeddingami | |
73 | + max_phrase_length = 60, | |
74 | + number_of_relations = number_of_relations ) | |
75 | + | |
76 | + train_data = load_stanford_data6("data/sst/train/dlabels.txt", "data/sst/train/dparents.txt","data/sst/train/sents.toks","data/sst/train/rels.txt",rnn.words2ids,True,batch_size,s['nc'], k_most_common_relations = number_of_relations) | |
77 | + | |
78 | + dev_data = load_stanford_data6("data/sst/dev/dlabels.txt", "data/sst/dev/dparents.txt","data/sst/dev/sents.toks","data/sst/dev/rels.txt",rnn.words2ids,False,0,s['nc'], k_most_common_relations = number_of_relations) | |
79 | + | |
80 | + test_data = load_stanford_data6("data/sst/test/dlabels.txt", "data/sst/test/dparents.txt","data/sst/test/sents.toks","data/sst/test/rels.txt",rnn.words2ids,False,0,s['nc'], k_most_common_relations = number_of_relations) | |
81 | + | |
82 | + n_train = len(train_data) | |
83 | + n_dev = len(dev_data) | |
84 | + n_test = len(test_data) | |
85 | + | |
86 | + print "" | |
87 | + print "lr = ", s['lr'], "number_of_relations = ", number_of_relations | |
88 | + print "model MLP_2_2 : ", "nchd_dim = ", nchd_dim ,"ne_dim = ", ne_dim , "nh2 =", nh2_dim | |
89 | + print "" | |
90 | + | |
91 | + best_prediction_valid_all = 0 | |
92 | + best_prediction_test_all = 0 | |
93 | + best_prediction_test_root = 0 | |
94 | + early_stop = 0 | |
95 | + | |
96 | + | |
97 | + tic = time.time() | |
98 | + | |
99 | + for e in xrange(s['nepochs']): | |
100 | + | |
101 | + #if e >= 1: | |
102 | + # s['lr'] = 0.8 * s['lr'] | |
103 | + | |
104 | + if early_stop == 5: | |
105 | + break | |
106 | + | |
107 | + | |
108 | + # shuffle | |
109 | + shuffle([train_data], s['seed']) | |
110 | + | |
111 | + for i in range(n_train): | |
112 | + rnn.train(train_data[i][0],train_data[i][1], train_data[i][2], train_data[i][3], train_data[i][4],s['lr']) | |
113 | + | |
114 | + | |
115 | + | |
116 | + # Dev: | |
117 | + counts_dev = np.zeros((s['nc'],s['nc']),dtype='int') | |
118 | + counts_dev_root = np.zeros((s['nc'],s['nc']),dtype='int') | |
119 | + for ii in range(n_dev): | |
120 | + pred = rnn.classify(dev_data[ii][0],dev_data[ii][1], dev_data[ii][3], dev_data[ii][4]) | |
121 | + for j in range(len(pred)): | |
122 | + counts_dev[pred[j], dev_data[ii][2][j]] += 1 | |
123 | + counts_dev_root[pred[-1], dev_data[ii][2][-1]] += 1 | |
124 | + | |
125 | + | |
126 | + # Test: | |
127 | + counts_test = np.zeros((s['nc'],s['nc']),dtype='int') | |
128 | + counts_test_root = np.zeros((s['nc'],s['nc']),dtype='int') | |
129 | + for i in range(n_test): | |
130 | + pred = rnn.classify(test_data[i][0],test_data[i][1], test_data[i][3], test_data[i][4]) | |
131 | + for j in range(len(pred)): | |
132 | + counts_test[pred[j], test_data[i][2][j]] += 1 | |
133 | + counts_test_root[pred[-1], test_data[i][2][-1]] += 1 | |
134 | + | |
135 | + # Train | |
136 | + counts = np.zeros((s['nc'],s['nc']),dtype='int') | |
137 | + counts_root = np.zeros((s['nc'],s['nc']),dtype='int') | |
138 | + for i in range(len(train_data)): | |
139 | + | |
140 | + if i % 10 == 0: #sprawdzamy dopasowanie na 1/10 zbioru zeby oszczedzic czas | |
141 | + pred = rnn.classify(train_data[i][0],train_data[i][1], train_data[i][3], train_data[i][4]) | |
142 | + for j in range(len(pred)): | |
143 | + counts[pred[j], train_data[i][2][j]] += 1 | |
144 | + counts_root[pred[-1], train_data[i][2][-1]] += 1 | |
145 | + | |
146 | + | |
147 | + | |
148 | + print("Valid: ", "%0.2f" % (100 * np.diag(counts_dev).sum()/float(counts_dev.sum())), | |
149 | + "Valid root: ","%0.2f" % (100 * np.diag(counts_dev_root).sum()/float(counts_dev_root.sum())), | |
150 | + " Test all: ","%0.2f" % (100 * np.diag(counts_test).sum()/float(counts_test.sum())), | |
151 | + "Test root: ","%0.2f" % (100 * np.diag(counts_test_root).sum()/float(counts_test_root.sum())), | |
152 | + " Train all: ","%0.2f" % (100 * np.diag(counts).sum()/float(counts.sum())), | |
153 | + "Train root: ","%0.2f" % (100 * np.diag(counts_root).sum()/float(counts_root.sum()))," time: ", time.time()-tic) | |
154 | + | |
155 | + if np.diag(counts_dev).sum()/float(counts_dev.sum()) > best_prediction_valid_all: | |
156 | + best_prediction_valid_all = np.diag(counts_dev).sum()/float(counts_dev.sum()) | |
157 | + best_prediction_test_all = np.diag(counts_test).sum()/float(counts_test.sum()) | |
158 | + best_prediction_test_root = np.diag(counts_test_root).sum()/float(counts_test_root.sum()) | |
159 | + | |
160 | + early_stop = 0 | |
161 | + else: | |
162 | + early_stop = early_stop + 1 | |
163 | + | |
164 | + | |
165 | + print("Best valid: ", "%0.2f" % (100 * best_prediction_valid_all)," Test all: ","%0.2f" % (100 * best_prediction_test_all),"Test root: ","%0.2f" % (100 * best_prediction_test_root), " time: ", time.time()-tic) | |
166 | + | |
167 | + | |
168 | + | |
169 | + | |
170 | + | |
... | ... |
main_for_experiments_on_stanford_data.py
0 → 100644
1 | +import numpy as np | |
2 | +import time | |
3 | +import sys | |
4 | +import subprocess | |
5 | +import os | |
6 | +import random | |
7 | + | |
8 | +#from modules.data import load | |
9 | +from modules.rnn.models import * | |
10 | +#from modules.metrics.accuracy import conlleval | |
11 | +from modules.utils.tools import load_stanford_data4, shuffle | |
12 | + | |
13 | +from theano import pp | |
14 | + | |
15 | +import theano.tensor as T | |
16 | +import theano | |
17 | +from theano.sandbox.rng_mrg import MRG_RandomStreams #as MRG_RandomStreams | |
18 | + | |
19 | +import itertools | |
20 | + | |
21 | +import os.path | |
22 | +import pickle | |
23 | + | |
24 | +from collections import Counter | |
25 | + | |
26 | + | |
27 | + | |
28 | +from theano import tensor as T, printing | |
29 | +from collections import OrderedDict | |
30 | +from theano.ifelse import ifelse | |
31 | + | |
32 | +from keras.preprocessing import sequence as seq | |
33 | + | |
34 | +dataType = 'int64' | |
35 | + | |
36 | + | |
37 | + | |
38 | +if __name__ == '__main__': | |
39 | + | |
40 | + #theano.config.floatX = 'float64' | |
41 | + | |
42 | + file_with_filtered_embeddings = "embeddings/embedding_and_words2ids.pkl" | |
43 | + if not os.path.exists(file_with_filtered_embeddings): | |
44 | + print("Cannot find file with only needed embeddings. We use 'filter_embeddings' in order to create it.") | |
45 | + filter_embeddings(["data/sst/train/dlabels.txt", "data/sst/train/dparents.txt","data/sst/train/sents.toks", "data/sst/dev/dlabels.txt", "data/sst/dev/dparents.txt","data/sst/dev/sents.toks", "data/sst/test/dlabels.txt", "data/sst/test/dparents.txt","data/sst/test/sents.toks"], | |
46 | + | |
47 | + "/home/norbert/Doktorat/clarin2sent/treelstm/data/glove/glove.840B.300d.txt", | |
48 | + file_with_filtered_embeddings) | |
49 | + | |
50 | + | |
51 | + batch_size = 1 | |
52 | + | |
53 | + | |
54 | + # ZDABAC MODELE 7,8,9 , 1, 2, 3 , 10, 11, 5,6 | |
55 | + | |
56 | + | |
57 | + | |
58 | + s = {'lr':0.002, | |
59 | + 'nepochs':30, | |
60 | + 'seed':345, | |
61 | + 'nc':5 # number of y classes | |
62 | + } | |
63 | + | |
64 | + | |
65 | + batch_size = 1 | |
66 | + | |
67 | + | |
68 | + | |
69 | + | |
70 | + | |
71 | + for ne_dim, nchd_dim in [(100,100)]:#,(200,200, 200,100),(200,200, 300,100),(100,100, 200,100)]: | |
72 | + | |
73 | + np.random.seed(s['seed']) | |
74 | + random.seed(s['seed']) | |
75 | + | |
76 | + | |
77 | + rnn = model55_pf1( ne = ne_dim, nchd = nchd_dim,# nh2 = nh2_dim, | |
78 | + nc = s['nc'], | |
79 | + w2v_model_path = file_with_filtered_embeddings, #sciezka do pliku z embeddingami | |
80 | + max_phrase_length = 60 ) | |
81 | + | |
82 | + train_data = load_stanford_data4("data/sst/train/dlabels.txt", "data/sst/train/dparents.txt","data/sst/train/sents.toks",rnn.words2ids,True,batch_size,s['nc']) | |
83 | + | |
84 | + dev_data = load_stanford_data4("data/sst/dev/dlabels.txt", "data/sst/dev/dparents.txt","data/sst/dev/sents.toks",rnn.words2ids,False,0,s['nc']) | |
85 | + | |
86 | + test_data = load_stanford_data4("data/sst/test/dlabels.txt", "data/sst/test/dparents.txt","data/sst/test/sents.toks",rnn.words2ids,False,0,s['nc']) | |
87 | + | |
88 | + n_train = len(train_data) | |
89 | + n_dev = len(dev_data) | |
90 | + n_test = len(test_data) | |
91 | + | |
92 | + print "" | |
93 | + print "lr = ", s['lr'] | |
94 | + print "model 55_pf1 : ", "nchd_dim = ", nchd_dim ,"ne_dim = ", ne_dim #, "nh2_dim = ", nh2_dim | |
95 | + print "" | |
96 | + | |
97 | + best_prediction_valid_all = 0 | |
98 | + best_prediction_test_all = 0 | |
99 | + best_prediction_test_root = 0 | |
100 | + early_stop = 0 | |
101 | + | |
102 | + | |
103 | + tic = time.time() | |
104 | + | |
105 | + for e in xrange(s['nepochs']): | |
106 | + | |
107 | + #if e >= 1: | |
108 | + # s['lr'] = 0.8 * s['lr'] | |
109 | + | |
110 | + if early_stop == 5: | |
111 | + break | |
112 | + | |
113 | + | |
114 | + # shuffle | |
115 | + shuffle([train_data], s['seed']) | |
116 | + | |
117 | + for i in range(n_train): | |
118 | + rnn.train(train_data[i][0],train_data[i][1], train_data[i][2], train_data[i][3], s['lr']) | |
119 | + | |
120 | + | |
121 | + | |
122 | + # Dev: | |
123 | + counts_dev = np.zeros((s['nc'],s['nc']),dtype='int') | |
124 | + counts_dev_root = np.zeros((s['nc'],s['nc']),dtype='int') | |
125 | + for ii in range(n_dev): | |
126 | + pred = rnn.classify(dev_data[ii][0],dev_data[ii][1], dev_data[ii][3]) | |
127 | + for j in range(len(pred)): | |
128 | + counts_dev[pred[j], dev_data[ii][2][j]] += 1 | |
129 | + counts_dev_root[pred[-1], dev_data[ii][2][-1]] += 1 | |
130 | + | |
131 | + | |
132 | + # Test: | |
133 | + counts_test = np.zeros((s['nc'],s['nc']),dtype='int') | |
134 | + counts_test_root = np.zeros((s['nc'],s['nc']),dtype='int') | |
135 | + for i in range(n_test): | |
136 | + pred = rnn.classify(test_data[i][0],test_data[i][1], test_data[i][3]) | |
137 | + for j in range(len(pred)): | |
138 | + counts_test[pred[j], test_data[i][2][j]] += 1 | |
139 | + counts_test_root[pred[-1], test_data[i][2][-1]] += 1 | |
140 | + | |
141 | + # Train | |
142 | + counts = np.zeros((s['nc'],s['nc']),dtype='int') | |
143 | + counts_root = np.zeros((s['nc'],s['nc']),dtype='int') | |
144 | + for i in range(len(train_data)): | |
145 | + | |
146 | + if i % 10 == 0: #sprawdzamy dopasowanie na 1/10 zbioru zeby oszczedzic czas | |
147 | + pred = rnn.classify(train_data[i][0],train_data[i][1], train_data[i][3]) | |
148 | + for j in range(len(pred)): | |
149 | + counts[pred[j], train_data[i][2][j]] += 1 | |
150 | + counts_root[pred[-1], train_data[i][2][-1]] += 1 | |
151 | + | |
152 | + | |
153 | + | |
154 | + print("Valid: ", "%0.2f" % (100 * np.diag(counts_dev).sum()/float(counts_dev.sum())), | |
155 | + "Valid root: ","%0.2f" % (100 * np.diag(counts_dev_root).sum()/float(counts_dev_root.sum())), | |
156 | + " Test all: ","%0.2f" % (100 * np.diag(counts_test).sum()/float(counts_test.sum())), | |
157 | + "Test root: ","%0.2f" % (100 * np.diag(counts_test_root).sum()/float(counts_test_root.sum())), | |
158 | + " Train all: ","%0.2f" % (100 * np.diag(counts).sum()/float(counts.sum())), | |
159 | + "Train root: ","%0.2f" % (100 * np.diag(counts_root).sum()/float(counts_root.sum()))," time: ", time.time()-tic) | |
160 | + | |
161 | + if np.diag(counts_dev).sum()/float(counts_dev.sum()) > best_prediction_valid_all: | |
162 | + best_prediction_valid_all = np.diag(counts_dev).sum()/float(counts_dev.sum()) | |
163 | + best_prediction_test_all = np.diag(counts_test).sum()/float(counts_test.sum()) | |
164 | + best_prediction_test_root = np.diag(counts_test_root).sum()/float(counts_test_root.sum()) | |
165 | + | |
166 | + early_stop = 0 | |
167 | + else: | |
168 | + early_stop = early_stop + 1 | |
169 | + | |
170 | + | |
171 | + print("Best valid: ", "%0.2f" % (100 * best_prediction_valid_all)," Test all: ","%0.2f" % (100 * best_prediction_test_all),"Test root: ","%0.2f" % (100 * best_prediction_test_root), " time: ", time.time()-tic) | |
172 | + | |
173 | + | |
174 | + | |
175 | + | |
176 | +### 9 , 2,3 , 5,6 | |
177 | + | |
178 | + | |
179 | + | |
... | ... |
main_for_sst_LSTM.py
0 → 100644
1 | +import numpy as np | |
2 | +import time | |
3 | +import sys | |
4 | +import subprocess | |
5 | +import os | |
6 | +import random | |
7 | + | |
8 | +#from modules.data import load | |
9 | +from modules.rnn.LSTM_models import * | |
10 | +#from modules.metrics.accuracy import conlleval | |
11 | +from modules.utils.tools import load_stanford_data4, shuffle | |
12 | + | |
13 | +from theano import pp | |
14 | + | |
15 | +import theano.tensor as T | |
16 | +import theano | |
17 | +from theano.sandbox.rng_mrg import MRG_RandomStreams #as MRG_RandomStreams | |
18 | + | |
19 | +import itertools | |
20 | + | |
21 | +import os.path | |
22 | +import pickle | |
23 | + | |
24 | +from collections import Counter | |
25 | + | |
26 | + | |
27 | + | |
28 | +from theano import tensor as T, printing | |
29 | +from collections import OrderedDict | |
30 | +from theano.ifelse import ifelse | |
31 | + | |
32 | +from keras.preprocessing import sequence as seq | |
33 | + | |
34 | +dataType = 'int64' | |
35 | + | |
36 | + | |
37 | +if __name__ == '__main__': | |
38 | + | |
39 | + | |
40 | + sys.setrecursionlimit(10000) | |
41 | + | |
42 | + #w2v_DIM = "300" | |
43 | + | |
44 | + | |
45 | + | |
46 | + file_with_filtered_embeddings = "embeddings/embedding_and_words2ids.pkl" | |
47 | + #if not os.path.exists(file_with_filtered_embeddings): | |
48 | + # print("Cannot find file with only needed embeddings. We use 'filter_embeddings' in order to create it.") | |
49 | + # filter_embeddings(["data/dane_polskie/train/train_labels.txt", "data/dane_polskie/train/train_parents.txt","data/dane_polskie/train/train_sentence.txt", | |
50 | + # "data/dane_polskie/dev/dev_labels.txt", "data/dane_polskie/dev/dev_parents.txt","data/dane_polskie/dev/dev_sentence.txt", | |
51 | + # "data/dane_polskie/test/test_labels.txt", "data/dane_polskie/test/test_parents.txt","data/dane_polskie/test/test_sentence.txt"], | |
52 | + | |
53 | + #"/home/norbert/Doktorat/clarin2sent/deeptagger/embeddings/w2v_allwiki_nkjpfull_"+w2v_DIM+".txt", | |
54 | + #file_with_filtered_embeddings) | |
55 | + | |
56 | + | |
57 | + s = {'lr':0.002, | |
58 | + 'nepochs':40, | |
59 | + 'seed':345, | |
60 | + 'nc':5 # number of y classes | |
61 | + } | |
62 | + batch_size = 1 | |
63 | + | |
64 | + | |
65 | + for h_dim in [100]: #100, 150, 200 | |
66 | + | |
67 | + np.random.seed(s['seed']) | |
68 | + random.seed(s['seed']) | |
69 | + | |
70 | + | |
71 | + rnn = LSTM_1( h_dim, | |
72 | + nc = s['nc'], | |
73 | + w2v_model_path = file_with_filtered_embeddings, #sciezka do pliku z embeddingami | |
74 | + max_phrase_length = 60 ) | |
75 | + | |
76 | + | |
77 | + train_data = load_stanford_data4("data/sst/train/dlabels.txt", "data/sst/train/dparents.txt","data/sst/train/sents.toks",rnn.words2ids,True,batch_size,s['nc']) | |
78 | + dev_data = load_stanford_data4("data/sst/dev/dlabels.txt", "data/sst/dev/dparents.txt","data/sst/dev/sents.toks",rnn.words2ids,False,0,s['nc']) | |
79 | + test_data = load_stanford_data4("data/sst/test/dlabels.txt", "data/sst/test/dparents.txt","data/sst/test/sents.toks",rnn.words2ids,False,0,s['nc']) | |
80 | + | |
81 | + n_train = len(train_data) | |
82 | + n_dev = len(dev_data) | |
83 | + n_test = len(test_data) | |
84 | + | |
85 | + print "" | |
86 | + print "learning rate: ", s['lr'] | |
87 | + print "model LSTM_1 : " , "h_dim = ", h_dim | |
88 | + print "" | |
89 | + | |
90 | + best_prediction_valid_all = 0 | |
91 | + best_prediction_test_all = 0 | |
92 | + best_prediction_test_root = 0 | |
93 | + early_stop = 0 | |
94 | + | |
95 | + | |
96 | + tic = time.time() | |
97 | + | |
98 | + for e in xrange(s['nepochs']): | |
99 | + | |
100 | + #if e >= 1: | |
101 | + # s['lr'] = 0.8 * s['lr'] | |
102 | + | |
103 | + if early_stop == 5: | |
104 | + break | |
105 | + | |
106 | + | |
107 | + # shuffle | |
108 | + shuffle([train_data], s['seed']) | |
109 | + | |
110 | + for i in range(n_train): | |
111 | + rnn.train(train_data[i][0],train_data[i][1], train_data[i][2], train_data[i][3], s['lr']) | |
112 | + | |
113 | + pickle.dump(rnn, open("model" + str(e) + ".pkl",'wb')) | |
114 | + | |
115 | + # Dev: | |
116 | + counts_dev = np.zeros((s['nc'],s['nc']),dtype='int') | |
117 | + counts_dev_root = np.zeros((s['nc'],s['nc']),dtype='int') | |
118 | + for ii in range(n_dev): | |
119 | + pred = rnn.classify(dev_data[ii][0],dev_data[ii][1], dev_data[ii][3]) | |
120 | + for j in range(len(pred)): | |
121 | + counts_dev[pred[j], dev_data[ii][2][j]] += 1 | |
122 | + counts_dev_root[pred[-1], dev_data[ii][2][-1]] += 1 | |
123 | + | |
124 | + | |
125 | + # Test: | |
126 | + counts_test = np.zeros((s['nc'],s['nc']),dtype='int') | |
127 | + counts_test_root = np.zeros((s['nc'],s['nc']),dtype='int') | |
128 | + for i in range(n_test): | |
129 | + pred = rnn.classify(test_data[i][0],test_data[i][1], test_data[i][3]) | |
130 | + for j in range(len(pred)): | |
131 | + counts_test[pred[j], test_data[i][2][j]] += 1 | |
132 | + counts_test_root[pred[-1], test_data[i][2][-1]] += 1 | |
133 | + | |
134 | + # Train | |
135 | + counts = np.zeros((s['nc'],s['nc']),dtype='int') | |
136 | + counts_root = np.zeros((s['nc'],s['nc']),dtype='int') | |
137 | + for i in range(len(train_data)): | |
138 | + | |
139 | + if i % 5 == 0: #sprawdzamy dopasowanie na 1/100 zbioru zeby oszczedzic czas | |
140 | + pred = rnn.classify(train_data[i][0],train_data[i][1], train_data[i][3]) | |
141 | + for j in range(len(pred)): | |
142 | + counts[pred[j], train_data[i][2][j]] += 1 | |
143 | + counts_root[pred[-1], train_data[i][2][-1]] += 1 | |
144 | + | |
145 | + print("epoch: ", e, | |
146 | + "V all: ", "%0.2f" % (100 * np.diag(counts_dev).sum()/float(counts_dev.sum())), | |
147 | + " Test all: ", "%0.2f" % (100 * np.diag(counts_test).sum()/float(counts_test.sum())), | |
148 | + "V root: ", "%0.2f" % (100 * np.diag(counts_dev_root).sum()/float(counts_dev_root.sum())), | |
149 | + " Test root: ", "%0.2f" % (100 * np.diag(counts_test_root).sum()/float(counts_test_root.sum())), | |
150 | + " Train: ", "%0.2f" % (100 * np.diag(counts).sum()/float(counts.sum())), | |
151 | + " Train root: ", "%0.2f" % (100 * np.diag(counts_root).sum()/float(counts_root.sum())) | |
152 | + ) | |
153 | + | |
154 | + | |
155 | + if np.diag(counts_dev).sum()/float(counts_dev.sum()) > best_prediction_valid_all: | |
156 | + best_prediction_valid_all = np.diag(counts_dev).sum()/float(counts_dev.sum()) | |
157 | + best_prediction_test_all = np.diag(counts_test).sum()/float(counts_test.sum()) | |
158 | + best_prediction_test_root = np.diag(counts_test_root).sum()/float(counts_test_root.sum()) | |
159 | + | |
160 | + early_stop = 0 | |
161 | + else: | |
162 | + early_stop = early_stop + 1 | |
163 | + | |
164 | + | |
165 | + print("Best valid: ", "%0.2f" % (100 * best_prediction_valid_all)," Test all: ","%0.2f" % (100 * best_prediction_test_all),"Test root: ","%0.2f" % (100 * best_prediction_test_root), " time: ", time.time()-tic) | |
166 | + | |
... | ... |
modules/rnn/LSTM_models.py
0 → 100644
1 | +import numpy as np | |
2 | +import time | |
3 | +import sys | |
4 | +import subprocess | |
5 | +import os | |
6 | +import random | |
7 | + | |
8 | +#from modules.data import load | |
9 | +#from modules.rnn.many_models import * | |
10 | +#from modules.metrics.accuracy import conlleval | |
11 | +from modules.utils.tools import load_stanford_data4 | |
12 | + | |
13 | +from theano import pp | |
14 | + | |
15 | +import theano.tensor as T | |
16 | +import theano | |
17 | +from theano.sandbox.rng_mrg import MRG_RandomStreams #as MRG_RandomStreams | |
18 | + | |
19 | +import itertools | |
20 | + | |
21 | +import os.path | |
22 | +import pickle | |
23 | + | |
24 | +from collections import Counter | |
25 | + | |
26 | + | |
27 | + | |
28 | +from theano import tensor as T, printing | |
29 | +from collections import OrderedDict | |
30 | +from theano.ifelse import ifelse | |
31 | + | |
32 | +from keras.preprocessing import sequence as seq | |
33 | + | |
34 | +dataType = 'int64' | |
35 | + | |
36 | + | |
37 | + | |
38 | + | |
39 | +class LSTM_1(object): | |
40 | + def __init__(self, h_dim, nc, w2v_model_path, max_phrase_length): | |
41 | + | |
42 | + ''' | |
43 | + nh :: dimension of hidden state | |
44 | + nc :: number of classes | |
45 | + ''' | |
46 | + | |
47 | + self.max_phrase_length = max_phrase_length | |
48 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
49 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
50 | + self.words2ids = w2vecs["words2ids"] | |
51 | + | |
52 | + emb_dim = w2vecs["vectors"].shape[1] | |
53 | + del w2vecs | |
54 | + | |
55 | + r = 0.05 | |
56 | + | |
57 | + self.W_i = theano.shared(r * np.random.uniform(-1.0, 1.0, (emb_dim, h_dim) ).astype(theano.config.floatX)) | |
58 | + self.U_i = theano.shared(r * np.random.uniform(-1.0, 1.0, (h_dim, h_dim) ).astype(theano.config.floatX)) | |
59 | + self.b_i = theano.shared(r * np.random.uniform(-1.0, 1.0, h_dim ).astype(theano.config.floatX)) | |
60 | + | |
61 | + self.W_f = theano.shared(r * np.random.uniform(-1.0, 1.0, (emb_dim, h_dim) ).astype(theano.config.floatX)) | |
62 | + self.U_f = theano.shared(r * np.random.uniform(-1.0, 1.0, (h_dim, h_dim) ).astype(theano.config.floatX)) | |
63 | + self.b_f = theano.shared(r * np.random.uniform(-1.0, 1.0, h_dim ).astype(theano.config.floatX)) | |
64 | + | |
65 | + self.W_o = theano.shared(r * np.random.uniform(-1.0, 1.0, (emb_dim, h_dim) ).astype(theano.config.floatX)) | |
66 | + self.U_o = theano.shared(r * np.random.uniform(-1.0, 1.0, (h_dim, h_dim) ).astype(theano.config.floatX)) | |
67 | + self.b_o = theano.shared(r * np.random.uniform(-1.0, 1.0, h_dim ).astype(theano.config.floatX)) | |
68 | + | |
69 | + self.W_u = theano.shared(r * np.random.uniform(-1.0, 1.0, (emb_dim, h_dim) ).astype(theano.config.floatX)) | |
70 | + self.U_u = theano.shared(r * np.random.uniform(-1.0, 1.0, (h_dim, h_dim) ).astype(theano.config.floatX)) | |
71 | + self.b_u = theano.shared(r * np.random.uniform(-1.0, 1.0, h_dim ).astype(theano.config.floatX)) | |
72 | + | |
73 | + self.W_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (h_dim, nc)).astype(theano.config.floatX)) | |
74 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
75 | + | |
76 | + | |
77 | + def one_step(word_id, word_children_positions, y_true, k, hidden_states, cell_states, learning_rate): | |
78 | + | |
79 | + x = self.emb[word_id] | |
80 | + # czyli wektor zerowy # sprawdzic + 0.5 | |
81 | + tmp = word_children_positions>=0.0 | |
82 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
83 | + idx_tmp = tmp.nonzero() # indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
84 | + | |
85 | + h_aggregated = ifelse(T.gt(number_of_children, 0.0), hidden_states[word_children_positions[idx_tmp]].sum(axis=0), hidden_states[-1]) | |
86 | + | |
87 | + #number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
88 | + #h_aggregated = h_aggregated/number_of_children # Usrednianie stanow ukrytych dzieci - | |
89 | + | |
90 | + | |
91 | + i = T.nnet.sigmoid( T.dot(x, self.W_i) + T.dot(h_aggregated, self.U_i) + self.b_i) | |
92 | + | |
93 | + o = T.nnet.sigmoid( T.dot(x, self.W_o) + T.dot(h_aggregated, self.U_o) + self.b_o) | |
94 | + | |
95 | + u = T.tanh( T.dot(x, self.W_u) + T.dot(h_aggregated, self.U_u) + self.b_u) | |
96 | + | |
97 | + f_c = ifelse(T.gt(number_of_children, 0.0), | |
98 | + (T.nnet.sigmoid( T.dot(x, self.W_f ) + T.dot(hidden_states[word_children_positions[idx_tmp]], self.U_f) + self.b_f )*cell_states[word_children_positions[idx_tmp]]).sum(axis=0), | |
99 | + T.nnet.sigmoid( T.dot(x, self.W_f ) + T.dot(hidden_states[-1], self.U_f) + self.b_f ) * cell_states[-1] | |
100 | + ) | |
101 | + | |
102 | + c = i*u + f_c | |
103 | + | |
104 | + h = o * T.tanh(c) | |
105 | + | |
106 | + current_cell_state = cell_states[k] | |
107 | + cell_states_new = T.set_subtensor(current_cell_state, c) | |
108 | + | |
109 | + current_hidden_state = hidden_states[k] | |
110 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
111 | + | |
112 | + | |
113 | + y_prob = T.nnet.softmax(T.dot(h,self.W_y) + self.b_y)[0] | |
114 | + | |
115 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
116 | + | |
117 | + return cross_entropy, hidden_states_new, cell_states_new | |
118 | + | |
119 | + | |
120 | + y = T.vector('y',dtype=dataType) | |
121 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
122 | + words = T.vector(dtype=dataType) | |
123 | + children_positions = T.matrix(dtype=dataType) | |
124 | + words_indexes = T.vector(dtype=dataType) | |
125 | + | |
126 | + [cross_entropy_vector, _, _] , _ = theano.scan(fn=one_step, \ | |
127 | + sequences = [words, children_positions,y,words_indexes], | |
128 | + outputs_info = [None, | |
129 | + theano.shared(np.zeros((self.max_phrase_length+1,h_dim), dtype = theano.config.floatX)), | |
130 | + theano.shared(np.zeros((self.max_phrase_length+1,h_dim), dtype = theano.config.floatX))], | |
131 | + non_sequences = learning_rate, | |
132 | + n_steps = words.shape[0]) | |
133 | + cost = T.sum(cross_entropy_vector) | |
134 | + | |
135 | + updates = OrderedDict([ | |
136 | + (self.W_i, self.W_i-learning_rate*T.grad(cost, self.W_i)), | |
137 | + (self.W_f, self.W_f-learning_rate*T.grad(cost, self.W_f)), | |
138 | + (self.W_o, self.W_o-learning_rate*T.grad(cost, self.W_o)), | |
139 | + (self.W_u, self.W_u-learning_rate*T.grad(cost, self.W_u)), | |
140 | + (self.W_y, self.W_y-learning_rate*T.grad(cost, self.W_y)), | |
141 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
142 | + (self.b_i, self.b_i-learning_rate*T.grad(cost,self.b_i)), | |
143 | + (self.b_f, self.b_f-learning_rate*T.grad(cost,self.b_f)), | |
144 | + (self.b_o, self.b_o-learning_rate*T.grad(cost,self.b_o)), | |
145 | + (self.b_u, self.b_u-learning_rate*T.grad(cost,self.b_u)), | |
146 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
147 | + ]) | |
148 | + | |
149 | + self.train = theano.function( inputs = [words, children_positions, y, words_indexes, learning_rate], | |
150 | + outputs = [], | |
151 | + updates = updates, | |
152 | + allow_input_downcast=True, | |
153 | + mode='FAST_RUN' | |
154 | + ) | |
155 | + | |
156 | + | |
157 | + def one_step_classify(word_id, word_children_positions, k, hidden_states, cell_states): | |
158 | + | |
159 | + x = self.emb[word_id] | |
160 | + # czyli wektor zerowy # sprawdzic + 0.5 | |
161 | + tmp = word_children_positions>=0.0 | |
162 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
163 | + idx_tmp = tmp.nonzero() # indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
164 | + | |
165 | + h_aggregated = ifelse(T.gt(number_of_children, 0.0), hidden_states[word_children_positions[idx_tmp]].sum(axis=0), hidden_states[-1]) | |
166 | + | |
167 | + #number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
168 | + #h_aggregated = h_aggregated/number_of_children # Usrednianie stanow ukrytych dzieci - | |
169 | + | |
170 | + | |
171 | + i = T.nnet.sigmoid( T.dot(x, self.W_i) + T.dot(h_aggregated, self.U_i) + self.b_i) | |
172 | + | |
173 | + o = T.nnet.sigmoid( T.dot(x, self.W_o) + T.dot(h_aggregated, self.U_o) + self.b_o) | |
174 | + | |
175 | + u = T.tanh( T.dot(x, self.W_u) + T.dot(h_aggregated, self.U_u) + self.b_u) | |
176 | + | |
177 | + f_c = ifelse(T.gt(number_of_children, 0.0), | |
178 | + (T.nnet.sigmoid( T.dot(x, self.W_f ) + T.dot(hidden_states[word_children_positions[idx_tmp]], self.U_f) + self.b_f )*cell_states[word_children_positions[idx_tmp]]).sum(axis=0), | |
179 | + T.nnet.sigmoid( T.dot(x, self.W_f ) + T.dot(hidden_states[-1], self.U_f) + self.b_f ) * cell_states[-1] | |
180 | + ) | |
181 | + | |
182 | + c = i*u + f_c | |
183 | + | |
184 | + h = o * T.tanh(c) | |
185 | + | |
186 | + current_cell_state = cell_states[k] | |
187 | + cell_states_new = T.set_subtensor(current_cell_state, c) | |
188 | + | |
189 | + current_hidden_state = hidden_states[k] | |
190 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
191 | + | |
192 | + | |
193 | + y_prob = T.nnet.softmax(T.dot(h,self.W_y) + self.b_y)[0] | |
194 | + | |
195 | + return y_prob, hidden_states_new, cell_states_new | |
196 | + | |
197 | + | |
198 | + [y_probs_classify, _, _ ], _ = theano.scan( | |
199 | + fn=one_step_classify, | |
200 | + sequences = [words, children_positions, words_indexes], | |
201 | + outputs_info = [None, | |
202 | + theano.shared(np.zeros((self.max_phrase_length+1,h_dim), dtype = theano.config.floatX)), | |
203 | + theano.shared(np.zeros((self.max_phrase_length+1,h_dim), dtype = theano.config.floatX))]) | |
204 | + | |
205 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
206 | + sequences = [words_indexes]) | |
207 | + | |
208 | + self.classify = theano.function(inputs=[words,children_positions,words_indexes], | |
209 | + outputs=predictions, | |
210 | + allow_input_downcast=True, | |
211 | + mode='FAST_RUN' | |
212 | + ) | |
213 | + | |
214 | + | |
... | ... |
modules/rnn/LSTM_models.pyc
0 → 100644
No preview for this file type
modules/rnn/models.py
0 → 100644
1 | +import numpy as np | |
2 | +import time | |
3 | +import sys | |
4 | +import subprocess | |
5 | +import os | |
6 | +import random | |
7 | + | |
8 | +#from modules.data import load | |
9 | +#from modules.rnn.many_models import * | |
10 | +#from modules.metrics.accuracy import conlleval | |
11 | +from modules.utils.tools import load_stanford_data4 | |
12 | + | |
13 | +from theano import pp | |
14 | + | |
15 | +import theano.tensor as T | |
16 | +import theano | |
17 | +from theano.sandbox.rng_mrg import MRG_RandomStreams #as MRG_RandomStreams | |
18 | + | |
19 | +import itertools | |
20 | + | |
21 | +import os.path | |
22 | +import pickle | |
23 | + | |
24 | +from collections import Counter | |
25 | + | |
26 | + | |
27 | + | |
28 | +from theano import tensor as T, printing | |
29 | +from collections import OrderedDict | |
30 | +from theano.ifelse import ifelse | |
31 | + | |
32 | +from keras.preprocessing import sequence as seq | |
33 | + | |
34 | +dataType = 'int64' | |
35 | + | |
36 | + | |
37 | + | |
38 | +# UWAGA: "ne" to NIE JEST to co jest napisane - to jest wymiar warstwy bezposrednio nad embeddingiem | |
39 | + | |
40 | + | |
41 | +class model55_pf1(object): | |
42 | + def __init__(self, ne, nchd, nc, w2v_model_path, max_phrase_length): | |
43 | + ''' | |
44 | + nh :: dimension of hidden state | |
45 | + nc :: number of classes | |
46 | + ne :: number of word embeddings in the vocabulary | |
47 | + de :: dimension of the word embeddings | |
48 | + ds :: dimension of the sentiment state | |
49 | + ''' | |
50 | + self.max_phrase_length = max_phrase_length | |
51 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
52 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
53 | + self.words2ids = w2vecs["words2ids"] | |
54 | + | |
55 | + #ne = len(w2vecs["words2ids"]) | |
56 | + de = w2vecs["vectors"].shape[1] | |
57 | + del w2vecs | |
58 | + | |
59 | + r = 0.05 | |
60 | + self.W_e_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (de, ne)).astype(theano.config.floatX)) | |
61 | + self.W_h_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nc)).astype(theano.config.floatX)) | |
62 | + | |
63 | + self.W_sh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nchd)).astype(theano.config.floatX)) | |
64 | + | |
65 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
66 | + | |
67 | + | |
68 | + def one_step(word_id, word_children_positions, y_true, i, hidden_states, learning_rate): | |
69 | + | |
70 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
71 | + | |
72 | + tmp = word_children_positions>=0.0 | |
73 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
74 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
75 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
76 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
77 | + schh = schh/number_of_children | |
78 | + | |
79 | + h = T.tanh(T.concatenate([T.dot(self.emb[word_id],self.W_e_h), T.dot(schh,self.W_sh_h)])) | |
80 | + | |
81 | + current_hidden_state = hidden_states[i] | |
82 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
83 | + | |
84 | + y_prob = T.nnet.softmax(T.dot(h,self.W_h_y) + self.b_y)[0] | |
85 | + | |
86 | + #l2_norm = T.sum(self.W_h_h2**2) + T.sum(self.W_h2_y**2) + T.sum(self.W_e_h**2) + T.sum(self.W_sh_h**2) + T.sum(self.emb**2) + T.sum(self.b_h**2) + T.sum(self.b_y**2) | |
87 | + | |
88 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
89 | + | |
90 | + return cross_entropy, hidden_states_new | |
91 | + | |
92 | + | |
93 | + y = T.vector('y',dtype=dataType) | |
94 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
95 | + words = T.vector(dtype=dataType) | |
96 | + children_positions = T.matrix(dtype=dataType) | |
97 | + words_indexes = T.vector(dtype=dataType) | |
98 | + | |
99 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
100 | + sequences = [words, children_positions,y,words_indexes], | |
101 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))], | |
102 | + non_sequences = learning_rate, | |
103 | + n_steps = words.shape[0]) | |
104 | + cost = T.sum(cross_entropy_vector[0]) | |
105 | + | |
106 | + updates = OrderedDict([ | |
107 | + (self.W_h_y, self.W_h_y-learning_rate*T.grad(cost, self.W_h_y)), | |
108 | + (self.W_e_h, self.W_e_h-learning_rate*T.grad(cost, self.W_e_h)), | |
109 | + (self.W_sh_h, self.W_sh_h-learning_rate*T.grad(cost, self.W_sh_h)), | |
110 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
111 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
112 | + ]) | |
113 | + | |
114 | + self.train = theano.function( inputs = [words, children_positions, y, words_indexes, learning_rate], | |
115 | + outputs = [], | |
116 | + updates = updates, | |
117 | + allow_input_downcast=True, | |
118 | + mode='FAST_RUN' | |
119 | + ) | |
120 | + | |
121 | + | |
122 | + def one_step_classify(word_id, word_children_positions, i, hidden_states): | |
123 | + | |
124 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
125 | + | |
126 | + tmp = word_children_positions>=0.0 | |
127 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
128 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
129 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
130 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
131 | + schh = schh/number_of_children | |
132 | + | |
133 | + h = T.tanh(T.concatenate([T.dot(self.emb[word_id],self.W_e_h), T.dot(schh,self.W_sh_h)]) ) | |
134 | + | |
135 | + current_hidden_state = hidden_states[i] | |
136 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
137 | + | |
138 | + y_prob = T.nnet.softmax(T.dot(h,self.W_h_y) + self.b_y)[0] | |
139 | + | |
140 | + return y_prob, hidden_states_new | |
141 | + | |
142 | + | |
143 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
144 | + fn=one_step_classify, | |
145 | + sequences = [words, children_positions,words_indexes], | |
146 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))]) | |
147 | + | |
148 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
149 | + sequences = [words_indexes]) | |
150 | + | |
151 | + self.classify = theano.function(inputs=[words,children_positions,words_indexes], | |
152 | + outputs=predictions, | |
153 | + allow_input_downcast=True, | |
154 | + mode='FAST_RUN' | |
155 | + ) | |
156 | + | |
157 | + | |
158 | + | |
159 | +class model55_pf2(object): | |
160 | + def __init__(self, ne, nchd, nh2, nc, w2v_model_path, max_phrase_length): | |
161 | + ''' | |
162 | + nh :: dimension of hidden state | |
163 | + nc :: number of classes | |
164 | + ne :: number of word embeddings in the vocabulary | |
165 | + de :: dimension of the word embeddings | |
166 | + ds :: dimension of the sentiment state | |
167 | + ''' | |
168 | + self.max_phrase_length = max_phrase_length | |
169 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
170 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
171 | + self.words2ids = w2vecs["words2ids"] | |
172 | + | |
173 | + #ne = len(w2vecs["words2ids"]) | |
174 | + de = w2vecs["vectors"].shape[1] | |
175 | + del w2vecs | |
176 | + | |
177 | + r = 0.05 | |
178 | + | |
179 | + self.W_e_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (de, ne)).astype(theano.config.floatX)) | |
180 | + | |
181 | + self.W_sh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nchd)).astype(theano.config.floatX)) | |
182 | + | |
183 | + self.W_h_h2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nh2)).astype(theano.config.floatX)) | |
184 | + self.W_h2_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh2, nc)).astype(theano.config.floatX)) | |
185 | + | |
186 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
187 | + | |
188 | + | |
189 | + def one_step(word_id, word_children_positions, y_true, i, hidden_states, learning_rate): | |
190 | + | |
191 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
192 | + | |
193 | + tmp = word_children_positions>=0.0 | |
194 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
195 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
196 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
197 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
198 | + schh = schh/number_of_children | |
199 | + | |
200 | + h = T.tanh(T.concatenate([T.dot(self.emb[word_id],self.W_e_h), T.dot(schh,self.W_sh_h)])) | |
201 | + #h = T.nnet.sigmoid(T.dot(self.emb[word_id],self.W_eh) + T.dot(schh,self.W_shsh) + self.bh) | |
202 | + | |
203 | + current_hidden_state = hidden_states[i] | |
204 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
205 | + | |
206 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
207 | + | |
208 | + y_prob = T.nnet.softmax(T.dot(h2,self.W_h2_y) + self.b_y)[0] | |
209 | + | |
210 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
211 | + | |
212 | + return cross_entropy, hidden_states_new | |
213 | + | |
214 | + | |
215 | + y = T.vector('y',dtype=dataType) | |
216 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
217 | + words = T.vector(dtype=dataType) | |
218 | + children_positions = T.matrix(dtype=dataType) | |
219 | + words_indexes = T.vector(dtype=dataType) | |
220 | + | |
221 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
222 | + sequences = [words, children_positions,y,words_indexes], | |
223 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))], | |
224 | + non_sequences = learning_rate, | |
225 | + n_steps = words.shape[0]) | |
226 | + cost = T.sum(cross_entropy_vector[0]) | |
227 | + | |
228 | + updates = OrderedDict([ | |
229 | + (self.W_h_h2, self.W_h_h2-learning_rate*T.grad(cost, self.W_h_h2)), | |
230 | + (self.W_h2_y, self.W_h2_y-learning_rate*T.grad(cost, self.W_h2_y)), | |
231 | + (self.W_e_h, self.W_e_h-learning_rate*T.grad(cost, self.W_e_h)), | |
232 | + (self.W_sh_h, self.W_sh_h-learning_rate*T.grad(cost, self.W_sh_h)), | |
233 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
234 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
235 | + ]) | |
236 | + | |
237 | + self.train = theano.function( inputs = [words, children_positions, y, words_indexes, learning_rate], | |
238 | + outputs = [], | |
239 | + updates = updates, | |
240 | + allow_input_downcast=True, | |
241 | + mode='FAST_RUN' | |
242 | + ) | |
243 | + | |
244 | + | |
245 | + def one_step_classify(word_id, word_children_positions, i, hidden_states): | |
246 | + | |
247 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
248 | + | |
249 | + tmp = word_children_positions>=0.0 | |
250 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
251 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
252 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
253 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
254 | + schh = schh/number_of_children | |
255 | + | |
256 | + h = T.tanh(T.concatenate([T.dot(self.emb[word_id],self.W_e_h), T.dot(schh,self.W_sh_h)])) | |
257 | + | |
258 | + current_hidden_state = hidden_states[i] | |
259 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
260 | + | |
261 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
262 | + | |
263 | + y_prob = T.nnet.softmax(T.dot(h2,self.W_h2_y) + self.b_y)[0] | |
264 | + | |
265 | + return y_prob, hidden_states_new | |
266 | + | |
267 | + | |
268 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
269 | + fn=one_step_classify, | |
270 | + sequences = [words, children_positions,words_indexes], | |
271 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))]) | |
272 | + | |
273 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
274 | + sequences = [words_indexes]) | |
275 | + | |
276 | + self.classify = theano.function(inputs=[words,children_positions,words_indexes], | |
277 | + outputs=predictions, | |
278 | + allow_input_downcast=True, | |
279 | + mode='FAST_RUN' | |
280 | + ) | |
281 | + | |
282 | + | |
283 | + | |
284 | + | |
285 | +class model55_pf3(object): | |
286 | + def __init__(self, ne, nchd, nh2, nh3, nc, w2v_model_path, max_phrase_length): | |
287 | + ''' | |
288 | + nh :: dimension of hidden state | |
289 | + nc :: number of classes | |
290 | + ne :: number of word embeddings in the vocabulary | |
291 | + de :: dimension of the word embeddings | |
292 | + ds :: dimension of the sentiment state | |
293 | + ''' | |
294 | + self.max_phrase_length = max_phrase_length | |
295 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
296 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
297 | + self.words2ids = w2vecs["words2ids"] | |
298 | + | |
299 | + #ne = len(w2vecs["words2ids"]) | |
300 | + de = w2vecs["vectors"].shape[1] | |
301 | + del w2vecs | |
302 | + | |
303 | + r = 0.05 | |
304 | + self.W_e_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (de, ne)).astype(theano.config.floatX)) | |
305 | + self.W_sh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nchd)).astype(theano.config.floatX)) | |
306 | + | |
307 | + self.W_h_h2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nh2)).astype(theano.config.floatX)) | |
308 | + self.W_h2_h3 = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh2, nh3)).astype(theano.config.floatX)) | |
309 | + self.W_h3_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh3, nc)).astype(theano.config.floatX)) | |
310 | + | |
311 | + | |
312 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
313 | + | |
314 | + | |
315 | + def one_step(word_id, word_children_positions, y_true, i, hidden_states, learning_rate): | |
316 | + | |
317 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
318 | + | |
319 | + tmp = word_children_positions>=0.0 | |
320 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
321 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
322 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
323 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
324 | + schh = schh/number_of_children | |
325 | + | |
326 | + h = T.tanh(T.concatenate([T.dot(self.emb[word_id],self.W_e_h), T.dot(schh,self.W_sh_h)])) | |
327 | + #h = T.nnet.sigmoid(T.dot(self.emb[word_id],self.W_eh) + T.dot(schh,self.W_shsh) + self.bh) | |
328 | + | |
329 | + current_hidden_state = hidden_states[i] | |
330 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
331 | + | |
332 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
333 | + | |
334 | + h3 = T.tanh(T.dot(h2, self.W_h2_h3)) | |
335 | + | |
336 | + y_prob = T.nnet.softmax(T.dot(h3,self.W_h3_y) + self.b_y)[0] | |
337 | + | |
338 | + #l2_norm = T.sum(self.W_h_h2**2) + T.sum(self.W_h2_y**2) + T.sum(self.W_e_h**2) + T.sum(self.W_sh_h**2) + T.sum(self.emb**2) + T.sum(self.b_h**2) + T.sum(self.b_y**2) | |
339 | + | |
340 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
341 | + | |
342 | + return cross_entropy, hidden_states_new | |
343 | + | |
344 | + | |
345 | + y = T.vector('y',dtype=dataType) | |
346 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
347 | + words = T.vector(dtype=dataType) | |
348 | + children_positions = T.matrix(dtype=dataType) | |
349 | + words_indexes = T.vector(dtype=dataType) | |
350 | + | |
351 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
352 | + sequences = [words, children_positions,y,words_indexes], | |
353 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))], | |
354 | + non_sequences = learning_rate, | |
355 | + n_steps = words.shape[0]) | |
356 | + cost = T.sum(cross_entropy_vector[0]) | |
357 | + | |
358 | + updates = OrderedDict([ | |
359 | + (self.W_h_h2, self.W_h_h2-learning_rate*T.grad(cost, self.W_h_h2)), | |
360 | + (self.W_h2_h3, self.W_h2_h3-learning_rate*T.grad(cost, self.W_h2_h3)), | |
361 | + (self.W_h3_y, self.W_h3_y-learning_rate*T.grad(cost, self.W_h3_y)), | |
362 | + (self.W_e_h, self.W_e_h-learning_rate*T.grad(cost, self.W_e_h)), | |
363 | + (self.W_sh_h, self.W_sh_h-learning_rate*T.grad(cost, self.W_sh_h)), | |
364 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
365 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
366 | + ]) | |
367 | + | |
368 | + self.train = theano.function( inputs = [words, children_positions, y, words_indexes, learning_rate], | |
369 | + outputs = [], | |
370 | + updates = updates, | |
371 | + allow_input_downcast=True, | |
372 | + mode='FAST_RUN' | |
373 | + ) | |
374 | + | |
375 | + | |
376 | + def one_step_classify(word_id, word_children_positions, i, hidden_states): | |
377 | + | |
378 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
379 | + | |
380 | + tmp = word_children_positions>=0.0 | |
381 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
382 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
383 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
384 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
385 | + schh = schh/number_of_children | |
386 | + | |
387 | + h = T.tanh(T.concatenate([T.dot(self.emb[word_id],self.W_e_h), T.dot(schh,self.W_sh_h)]) ) | |
388 | + | |
389 | + current_hidden_state = hidden_states[i] | |
390 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
391 | + | |
392 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
393 | + | |
394 | + h3 = T.tanh(T.dot(h2, self.W_h2_h3)) | |
395 | + | |
396 | + y_prob = T.nnet.softmax(T.dot(h3,self.W_h3_y) + self.b_y)[0] | |
397 | + | |
398 | + return y_prob, hidden_states_new | |
399 | + | |
400 | + | |
401 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
402 | + fn=one_step_classify, | |
403 | + sequences = [words, children_positions,words_indexes], | |
404 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))]) | |
405 | + | |
406 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
407 | + sequences = [words_indexes]) | |
408 | + | |
409 | + self.classify = theano.function(inputs=[words,children_positions,words_indexes], | |
410 | + outputs=predictions, | |
411 | + allow_input_downcast=True, | |
412 | + mode='FAST_RUN' | |
413 | + ) | |
414 | + | |
415 | + | |
416 | + | |
417 | + | |
418 | + | |
419 | + | |
420 | +class model55_pf4(object): | |
421 | + def __init__(self, neh, ne, nchd, nc, w2v_model_path, max_phrase_length): | |
422 | + ''' | |
423 | + nh :: dimension of hidden state | |
424 | + nc :: number of classes | |
425 | + ne :: number of word embeddings in the vocabulary | |
426 | + de :: dimension of the word embeddings | |
427 | + ds :: dimension of the sentiment state | |
428 | + ''' | |
429 | + self.max_phrase_length = max_phrase_length | |
430 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
431 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
432 | + self.words2ids = w2vecs["words2ids"] | |
433 | + | |
434 | + de = w2vecs["vectors"].shape[1] | |
435 | + del w2vecs | |
436 | + | |
437 | + r = 0.05 | |
438 | + self.W_e_eh = theano.shared(r * np.random.uniform(-1.0, 1.0, (de, neh)).astype(theano.config.floatX)) | |
439 | + | |
440 | + self.W_eh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (neh, ne)).astype(theano.config.floatX)) | |
441 | + | |
442 | + self.W_h_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nc)).astype(theano.config.floatX)) | |
443 | + | |
444 | + self.W_sh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nchd)).astype(theano.config.floatX)) | |
445 | + | |
446 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
447 | + | |
448 | + | |
449 | + def one_step(word_id, word_children_positions, y_true, i, hidden_states, learning_rate): | |
450 | + | |
451 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
452 | + | |
453 | + tmp = word_children_positions>=0.0 | |
454 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
455 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
456 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
457 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
458 | + schh = schh/number_of_children | |
459 | + | |
460 | + eh = T.tanh(T.dot(self.emb[word_id],self.W_e_eh)) | |
461 | + | |
462 | + h = T.tanh(T.concatenate([T.dot(eh, self.W_eh_h), T.dot(schh,self.W_sh_h)])) | |
463 | + | |
464 | + current_hidden_state = hidden_states[i] | |
465 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
466 | + | |
467 | + y_prob = T.nnet.softmax(T.dot(h,self.W_h_y) + self.b_y)[0] | |
468 | + | |
469 | + #l2_norm = T.sum(self.W_h_h2**2) + T.sum(self.W_h2_y**2) + T.sum(self.W_e_h**2) + T.sum(self.W_sh_h**2) + T.sum(self.emb**2) + T.sum(self.b_h**2) + T.sum(self.b_y**2) | |
470 | + | |
471 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
472 | + | |
473 | + return cross_entropy, hidden_states_new | |
474 | + | |
475 | + | |
476 | + y = T.vector('y',dtype=dataType) | |
477 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
478 | + words = T.vector(dtype=dataType) | |
479 | + children_positions = T.matrix(dtype=dataType) | |
480 | + words_indexes = T.vector(dtype=dataType) | |
481 | + | |
482 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
483 | + sequences = [words, children_positions,y,words_indexes], | |
484 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))], | |
485 | + non_sequences = learning_rate, | |
486 | + n_steps = words.shape[0]) | |
487 | + cost = T.sum(cross_entropy_vector[0]) | |
488 | + | |
489 | + updates = OrderedDict([ | |
490 | + (self.W_h_y, self.W_h_y-learning_rate*T.grad(cost, self.W_h_y)), | |
491 | + (self.W_e_eh, self.W_e_eh-learning_rate*T.grad(cost, self.W_e_eh)), | |
492 | + (self.W_eh_h, self.W_eh_h-learning_rate*T.grad(cost, self.W_eh_h)), | |
493 | + (self.W_sh_h, self.W_sh_h-learning_rate*T.grad(cost, self.W_sh_h)), | |
494 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
495 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
496 | + ]) | |
497 | + | |
498 | + self.train = theano.function( inputs = [words, children_positions, y, words_indexes, learning_rate], | |
499 | + outputs = [], | |
500 | + updates = updates, | |
501 | + allow_input_downcast=True, | |
502 | + mode='FAST_RUN' | |
503 | + ) | |
504 | + | |
505 | + | |
506 | + def one_step_classify(word_id, word_children_positions, i, hidden_states): | |
507 | + | |
508 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
509 | + | |
510 | + tmp = word_children_positions>=0.0 | |
511 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
512 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
513 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
514 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
515 | + schh = schh/number_of_children | |
516 | + | |
517 | + eh = T.tanh(T.dot(self.emb[word_id],self.W_e_eh)) | |
518 | + | |
519 | + h = T.tanh(T.concatenate([T.dot(eh, self.W_eh_h), T.dot(schh,self.W_sh_h)])) | |
520 | + | |
521 | + current_hidden_state = hidden_states[i] | |
522 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
523 | + | |
524 | + y_prob = T.nnet.softmax(T.dot(h,self.W_h_y) + self.b_y)[0] | |
525 | + | |
526 | + return y_prob, hidden_states_new | |
527 | + | |
528 | + | |
529 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
530 | + fn=one_step_classify, | |
531 | + sequences = [words, children_positions,words_indexes], | |
532 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))]) | |
533 | + | |
534 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
535 | + sequences = [words_indexes]) | |
536 | + | |
537 | + self.classify = theano.function(inputs=[words,children_positions,words_indexes], | |
538 | + outputs=predictions, | |
539 | + allow_input_downcast=True, | |
540 | + mode='FAST_RUN' | |
541 | + ) | |
542 | + | |
543 | + | |
544 | + | |
545 | + | |
546 | + | |
547 | + | |
548 | +class model55_pf5(object): | |
549 | + def __init__(self, ne, nshh, nchd, nc, w2v_model_path, max_phrase_length): | |
550 | + ''' | |
551 | + nh :: dimension of hidden state | |
552 | + nc :: number of classes | |
553 | + ne :: number of word embeddings in the vocabulary | |
554 | + de :: dimension of the word embeddings | |
555 | + ds :: dimension of the sentiment state | |
556 | + ''' | |
557 | + self.max_phrase_length = max_phrase_length | |
558 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
559 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
560 | + self.words2ids = w2vecs["words2ids"] | |
561 | + | |
562 | + #ne = len(w2vecs["words2ids"]) | |
563 | + de = w2vecs["vectors"].shape[1] | |
564 | + del w2vecs | |
565 | + | |
566 | + r = 0.05 | |
567 | + self.W_e_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (de, ne)).astype(theano.config.floatX)) | |
568 | + | |
569 | + | |
570 | + self.W_h_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nc)).astype(theano.config.floatX)) | |
571 | + | |
572 | + self.W_sh_shh = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nshh)).astype(theano.config.floatX)) | |
573 | + | |
574 | + self.W_shh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (nshh, nchd)).astype(theano.config.floatX)) | |
575 | + | |
576 | + | |
577 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
578 | + | |
579 | + | |
580 | + def one_step(word_id, word_children_positions, y_true, i, hidden_states, learning_rate): | |
581 | + | |
582 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
583 | + | |
584 | + tmp = word_children_positions>=0.0 | |
585 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
586 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
587 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
588 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
589 | + schh = schh/number_of_children | |
590 | + | |
591 | + shh = T.tanh(T.dot(schh,self.W_sh_shh)) | |
592 | + | |
593 | + h = T.tanh(T.concatenate([T.dot(self.emb[word_id], self.W_e_h), T.dot(shh,self.W_shh_h)])) | |
594 | + | |
595 | + current_hidden_state = hidden_states[i] | |
596 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
597 | + | |
598 | + y_prob = T.nnet.softmax(T.dot(h,self.W_h_y) + self.b_y)[0] | |
599 | + | |
600 | + #l2_norm = T.sum(self.W_h_h2**2) + T.sum(self.W_h2_y**2) + T.sum(self.W_e_h**2) + T.sum(self.W_sh_h**2) + T.sum(self.emb**2) + T.sum(self.b_h**2) + T.sum(self.b_y**2) | |
601 | + | |
602 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
603 | + | |
604 | + return cross_entropy, hidden_states_new | |
605 | + | |
606 | + | |
607 | + y = T.vector('y',dtype=dataType) | |
608 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
609 | + words = T.vector(dtype=dataType) | |
610 | + children_positions = T.matrix(dtype=dataType) | |
611 | + words_indexes = T.vector(dtype=dataType) | |
612 | + | |
613 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
614 | + sequences = [words, children_positions,y,words_indexes], | |
615 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))], | |
616 | + non_sequences = learning_rate, | |
617 | + n_steps = words.shape[0]) | |
618 | + cost = T.sum(cross_entropy_vector[0]) | |
619 | + | |
620 | + updates = OrderedDict([ | |
621 | + (self.W_h_y, self.W_h_y-learning_rate*T.grad(cost, self.W_h_y)), | |
622 | + (self.W_e_h, self.W_e_h-learning_rate*T.grad(cost, self.W_e_h)), | |
623 | + (self.W_shh_h, self.W_shh_h-learning_rate*T.grad(cost, self.W_shh_h)), | |
624 | + (self.W_sh_shh, self.W_sh_shh-learning_rate*T.grad(cost, self.W_sh_shh)), | |
625 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
626 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
627 | + ]) | |
628 | + | |
629 | + self.train = theano.function( inputs = [words, children_positions, y, words_indexes, learning_rate], | |
630 | + outputs = [], | |
631 | + updates = updates, | |
632 | + allow_input_downcast=True, | |
633 | + mode='FAST_RUN' | |
634 | + ) | |
635 | + | |
636 | + | |
637 | + def one_step_classify(word_id, word_children_positions, i, hidden_states): | |
638 | + | |
639 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
640 | + | |
641 | + tmp = word_children_positions>=0.0 | |
642 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
643 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
644 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
645 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
646 | + schh = schh/number_of_children | |
647 | + | |
648 | + shh = T.tanh(T.dot(schh,self.W_sh_shh)) | |
649 | + | |
650 | + h = T.tanh(T.concatenate([T.dot(self.emb[word_id], self.W_e_h), T.dot(shh,self.W_shh_h)])) | |
651 | + | |
652 | + | |
653 | + current_hidden_state = hidden_states[i] | |
654 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
655 | + | |
656 | + y_prob = T.nnet.softmax(T.dot(h,self.W_h_y) + self.b_y)[0] | |
657 | + | |
658 | + return y_prob, hidden_states_new | |
659 | + | |
660 | + | |
661 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
662 | + fn=one_step_classify, | |
663 | + sequences = [words, children_positions,words_indexes], | |
664 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))]) | |
665 | + | |
666 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
667 | + sequences = [words_indexes]) | |
668 | + | |
669 | + self.classify = theano.function(inputs=[words,children_positions,words_indexes], | |
670 | + outputs=predictions, | |
671 | + allow_input_downcast=True, | |
672 | + mode='FAST_RUN' | |
673 | + ) | |
674 | + | |
675 | + | |
676 | + | |
677 | + | |
678 | +class model55_pf6(object): | |
679 | + def __init__(self, neh, ne, nshh, nchd, nc, w2v_model_path, max_phrase_length): | |
680 | + ''' | |
681 | + nh :: dimension of hidden state | |
682 | + nc :: number of classes | |
683 | + ne :: number of word embeddings in the vocabulary | |
684 | + de :: dimension of the word embeddings | |
685 | + ds :: dimension of the sentiment state | |
686 | + ''' | |
687 | + self.max_phrase_length = max_phrase_length | |
688 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
689 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
690 | + self.words2ids = w2vecs["words2ids"] | |
691 | + | |
692 | + #ne = len(w2vecs["words2ids"]) | |
693 | + de = w2vecs["vectors"].shape[1] | |
694 | + del w2vecs | |
695 | + | |
696 | + r = 0.05 | |
697 | + self.W_e_eh = theano.shared(r * np.random.uniform(-1.0, 1.0, (de, neh)).astype(theano.config.floatX)) | |
698 | + | |
699 | + self.W_eh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (neh, ne)).astype(theano.config.floatX)) | |
700 | + | |
701 | + self.W_h_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nc)).astype(theano.config.floatX)) | |
702 | + | |
703 | + self.W_sh_shh = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nshh)).astype(theano.config.floatX)) | |
704 | + | |
705 | + self.W_shh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (nshh, nchd)).astype(theano.config.floatX)) | |
706 | + | |
707 | + | |
708 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
709 | + | |
710 | + | |
711 | + def one_step(word_id, word_children_positions, y_true, i, hidden_states, learning_rate): | |
712 | + | |
713 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
714 | + | |
715 | + tmp = word_children_positions>=0.0 | |
716 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
717 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
718 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
719 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
720 | + schh = schh/number_of_children | |
721 | + | |
722 | + eh = T.tanh(T.dot(self.emb[word_id],self.W_e_eh)) | |
723 | + shh = T.tanh(T.dot(schh,self.W_sh_shh)) | |
724 | + | |
725 | + h = T.tanh(T.concatenate([T.dot(eh, self.W_eh_h), T.dot(shh,self.W_shh_h)])) | |
726 | + | |
727 | + current_hidden_state = hidden_states[i] | |
728 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
729 | + | |
730 | + y_prob = T.nnet.softmax(T.dot(h,self.W_h_y) + self.b_y)[0] | |
731 | + | |
732 | + #l2_norm = T.sum(self.W_h_h2**2) + T.sum(self.W_h2_y**2) + T.sum(self.W_e_h**2) + T.sum(self.W_sh_h**2) + T.sum(self.emb**2) + T.sum(self.b_h**2) + T.sum(self.b_y**2) | |
733 | + | |
734 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
735 | + | |
736 | + return cross_entropy, hidden_states_new | |
737 | + | |
738 | + | |
739 | + y = T.vector('y',dtype=dataType) | |
740 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
741 | + words = T.vector(dtype=dataType) | |
742 | + children_positions = T.matrix(dtype=dataType) | |
743 | + words_indexes = T.vector(dtype=dataType) | |
744 | + | |
745 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
746 | + sequences = [words, children_positions,y,words_indexes], | |
747 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))], | |
748 | + non_sequences = learning_rate, | |
749 | + n_steps = words.shape[0]) | |
750 | + cost = T.sum(cross_entropy_vector[0]) | |
751 | + | |
752 | + updates = OrderedDict([ | |
753 | + (self.W_h_y, self.W_h_y-learning_rate*T.grad(cost, self.W_h_y)), | |
754 | + (self.W_e_eh, self.W_e_eh-learning_rate*T.grad(cost, self.W_e_eh)), | |
755 | + (self.W_eh_h, self.W_eh_h-learning_rate*T.grad(cost, self.W_eh_h)), | |
756 | + (self.W_shh_h, self.W_shh_h-learning_rate*T.grad(cost, self.W_shh_h)), | |
757 | + (self.W_sh_shh, self.W_sh_shh-learning_rate*T.grad(cost, self.W_sh_shh)), | |
758 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
759 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
760 | + ]) | |
761 | + | |
762 | + self.train = theano.function( inputs = [words, children_positions, y, words_indexes, learning_rate], | |
763 | + outputs = [], | |
764 | + updates = updates, | |
765 | + allow_input_downcast=True, | |
766 | + mode='FAST_RUN' | |
767 | + ) | |
768 | + | |
769 | + | |
770 | + def one_step_classify(word_id, word_children_positions, i, hidden_states): | |
771 | + | |
772 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
773 | + | |
774 | + tmp = word_children_positions>=0.0 | |
775 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
776 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
777 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
778 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
779 | + schh = schh/number_of_children | |
780 | + | |
781 | + eh = T.tanh(T.dot(self.emb[word_id],self.W_e_eh)) | |
782 | + shh = T.tanh(T.dot(schh,self.W_sh_shh)) | |
783 | + | |
784 | + h = T.tanh(T.concatenate([T.dot(eh, self.W_eh_h), T.dot(shh,self.W_shh_h)])) | |
785 | + | |
786 | + | |
787 | + current_hidden_state = hidden_states[i] | |
788 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
789 | + | |
790 | + y_prob = T.nnet.softmax(T.dot(h,self.W_h_y) + self.b_y)[0] | |
791 | + | |
792 | + return y_prob, hidden_states_new | |
793 | + | |
794 | + | |
795 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
796 | + fn=one_step_classify, | |
797 | + sequences = [words, children_positions,words_indexes], | |
798 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))]) | |
799 | + | |
800 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
801 | + sequences = [words_indexes]) | |
802 | + | |
803 | + self.classify = theano.function(inputs=[words,children_positions,words_indexes], | |
804 | + outputs=predictions, | |
805 | + allow_input_downcast=True, | |
806 | + mode='FAST_RUN' | |
807 | + ) | |
808 | + | |
809 | + | |
810 | + | |
811 | + | |
812 | + | |
813 | +class model55_pf7(object): | |
814 | + def __init__(self, neh, ne, nshh, nchd, nh2, nc, w2v_model_path, max_phrase_length): | |
815 | + ''' | |
816 | + nh :: dimension of hidden state | |
817 | + nc :: number of classes | |
818 | + ne :: number of word embeddings in the vocabulary | |
819 | + de :: dimension of the word embeddings | |
820 | + ds :: dimension of the sentiment state | |
821 | + ''' | |
822 | + self.max_phrase_length = max_phrase_length | |
823 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
824 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
825 | + self.words2ids = w2vecs["words2ids"] | |
826 | + | |
827 | + #ne = len(w2vecs["words2ids"]) | |
828 | + de = w2vecs["vectors"].shape[1] | |
829 | + del w2vecs | |
830 | + | |
831 | + r = 0.05 | |
832 | + self.W_e_eh = theano.shared(r * np.random.uniform(-1.0, 1.0, (de, neh)).astype(theano.config.floatX)) | |
833 | + | |
834 | + self.W_eh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (neh, ne)).astype(theano.config.floatX)) | |
835 | + | |
836 | + self.W_h_h2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nh2)).astype(theano.config.floatX)) | |
837 | + self.W_h2_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh2, nc)).astype(theano.config.floatX)) | |
838 | + | |
839 | + self.W_sh_shh = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nshh)).astype(theano.config.floatX)) | |
840 | + | |
841 | + self.W_shh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (nshh, nchd)).astype(theano.config.floatX)) | |
842 | + | |
843 | + | |
844 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
845 | + | |
846 | + | |
847 | + def one_step(word_id, word_children_positions, y_true, i, hidden_states, learning_rate): | |
848 | + | |
849 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
850 | + | |
851 | + tmp = word_children_positions>=0.0 | |
852 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
853 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
854 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
855 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
856 | + schh = schh/number_of_children | |
857 | + | |
858 | + eh = T.tanh(T.dot(self.emb[word_id],self.W_e_eh)) | |
859 | + shh = T.tanh(T.dot(schh,self.W_sh_shh)) | |
860 | + | |
861 | + h = T.tanh(T.concatenate([T.dot(eh, self.W_eh_h), T.dot(shh,self.W_shh_h)])) | |
862 | + | |
863 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
864 | + | |
865 | + current_hidden_state = hidden_states[i] | |
866 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
867 | + | |
868 | + y_prob = T.nnet.softmax(T.dot(h2,self.W_h2_y) + self.b_y)[0] | |
869 | + | |
870 | + #l2_norm = T.sum(self.W_h_h2**2) + T.sum(self.W_h2_y**2) + T.sum(self.W_e_h**2) + T.sum(self.W_sh_h**2) + T.sum(self.emb**2) + T.sum(self.b_h**2) + T.sum(self.b_y**2) | |
871 | + | |
872 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
873 | + | |
874 | + return cross_entropy, hidden_states_new | |
875 | + | |
876 | + | |
877 | + y = T.vector('y',dtype=dataType) | |
878 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
879 | + words = T.vector(dtype=dataType) | |
880 | + children_positions = T.matrix(dtype=dataType) | |
881 | + words_indexes = T.vector(dtype=dataType) | |
882 | + | |
883 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
884 | + sequences = [words, children_positions,y,words_indexes], | |
885 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))], | |
886 | + non_sequences = learning_rate, | |
887 | + n_steps = words.shape[0]) | |
888 | + | |
889 | + cost = T.sum(cross_entropy_vector[0]) | |
890 | + | |
891 | + updates = OrderedDict([ | |
892 | + (self.W_h_h2, self.W_h_h2-learning_rate*T.grad(cost, self.W_h_h2)), | |
893 | + (self.W_h2_y, self.W_h2_y-learning_rate*T.grad(cost, self.W_h2_y)), | |
894 | + (self.W_e_eh, self.W_e_eh-learning_rate*T.grad(cost, self.W_e_eh)), | |
895 | + (self.W_eh_h, self.W_eh_h-learning_rate*T.grad(cost, self.W_eh_h)), | |
896 | + (self.W_shh_h, self.W_shh_h-learning_rate*T.grad(cost, self.W_shh_h)), | |
897 | + (self.W_sh_shh, self.W_sh_shh-learning_rate*T.grad(cost, self.W_sh_shh)), | |
898 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
899 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
900 | + ]) | |
901 | + | |
902 | + self.train = theano.function( inputs = [words, children_positions, y, words_indexes, learning_rate], | |
903 | + outputs = [], | |
904 | + updates = updates, | |
905 | + allow_input_downcast=True, | |
906 | + mode='FAST_RUN' | |
907 | + ) | |
908 | + | |
909 | + | |
910 | + def one_step_classify(word_id, word_children_positions, i, hidden_states): | |
911 | + | |
912 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
913 | + | |
914 | + tmp = word_children_positions>=0.0 | |
915 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
916 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
917 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
918 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
919 | + schh = schh/number_of_children | |
920 | + | |
921 | + eh = T.tanh(T.dot(self.emb[word_id],self.W_e_eh)) | |
922 | + shh = T.tanh(T.dot(schh,self.W_sh_shh)) | |
923 | + | |
924 | + h = T.tanh(T.concatenate([T.dot(eh, self.W_eh_h), T.dot(shh,self.W_shh_h)])) | |
925 | + | |
926 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
927 | + | |
928 | + current_hidden_state = hidden_states[i] | |
929 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
930 | + | |
931 | + y_prob = T.nnet.softmax(T.dot(h2,self.W_h2_y) + self.b_y)[0] | |
932 | + | |
933 | + return y_prob, hidden_states_new | |
934 | + | |
935 | + | |
936 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
937 | + fn=one_step_classify, | |
938 | + sequences = [words, children_positions,words_indexes], | |
939 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))]) | |
940 | + | |
941 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
942 | + sequences = [words_indexes]) | |
943 | + | |
944 | + self.classify = theano.function(inputs=[words,children_positions,words_indexes], | |
945 | + outputs=predictions, | |
946 | + allow_input_downcast=True, | |
947 | + mode='FAST_RUN' | |
948 | + ) | |
949 | + | |
950 | + | |
951 | + | |
952 | + | |
953 | + | |
954 | +class model55_pf8(object): | |
955 | + def __init__(self, neh, ne, nshh, nchd, nh2, nh3, nc, w2v_model_path, max_phrase_length): | |
956 | + ''' | |
957 | + nh :: dimension of hidden state | |
958 | + nc :: number of classes | |
959 | + ne :: number of word embeddings in the vocabulary | |
960 | + de :: dimension of the word embeddings | |
961 | + ds :: dimension of the sentiment state | |
962 | + ''' | |
963 | + self.max_phrase_length = max_phrase_length | |
964 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
965 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
966 | + self.words2ids = w2vecs["words2ids"] | |
967 | + | |
968 | + #ne = len(w2vecs["words2ids"]) | |
969 | + de = w2vecs["vectors"].shape[1] | |
970 | + del w2vecs | |
971 | + | |
972 | + r = 0.05 | |
973 | + self.W_e_eh = theano.shared(r * np.random.uniform(-1.0, 1.0, (de, neh)).astype(theano.config.floatX)) | |
974 | + | |
975 | + self.W_eh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (neh, ne)).astype(theano.config.floatX)) | |
976 | + | |
977 | + self.W_h_h2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nh2)).astype(theano.config.floatX)) | |
978 | + self.W_h2_h3 = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh2, nh3)).astype(theano.config.floatX)) | |
979 | + self.W_h3_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh3, nc)).astype(theano.config.floatX)) | |
980 | + | |
981 | + self.W_sh_shh = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nshh)).astype(theano.config.floatX)) | |
982 | + | |
983 | + self.W_shh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (nshh, nchd)).astype(theano.config.floatX)) | |
984 | + | |
985 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
986 | + | |
987 | + | |
988 | + def one_step(word_id, word_children_positions, y_true, i, hidden_states, learning_rate): | |
989 | + | |
990 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
991 | + | |
992 | + tmp = word_children_positions>=0.0 | |
993 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
994 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
995 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
996 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
997 | + schh = schh/number_of_children | |
998 | + | |
999 | + eh = T.tanh(T.dot(self.emb[word_id],self.W_e_eh)) | |
1000 | + shh = T.tanh(T.dot(schh,self.W_sh_shh)) | |
1001 | + | |
1002 | + h = T.tanh(T.concatenate([T.dot(eh, self.W_eh_h), T.dot(shh,self.W_shh_h)])) | |
1003 | + | |
1004 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
1005 | + | |
1006 | + | |
1007 | + h3 = T.tanh(T.dot(h2, self.W_h2_h3)) | |
1008 | + | |
1009 | + current_hidden_state = hidden_states[i] | |
1010 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
1011 | + | |
1012 | + y_prob = T.nnet.softmax(T.dot(h3,self.W_h3_y) + self.b_y)[0] | |
1013 | + | |
1014 | + #l2_norm = T.sum(self.W_h_h2**2) + T.sum(self.W_h2_y**2) + T.sum(self.W_e_h**2) + T.sum(self.W_sh_h**2) + T.sum(self.emb**2) + T.sum(self.b_h**2) + T.sum(self.b_y**2) | |
1015 | + | |
1016 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
1017 | + | |
1018 | + return cross_entropy, hidden_states_new | |
1019 | + | |
1020 | + | |
1021 | + y = T.vector('y',dtype=dataType) | |
1022 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
1023 | + words = T.vector(dtype=dataType) | |
1024 | + children_positions = T.matrix(dtype=dataType) | |
1025 | + words_indexes = T.vector(dtype=dataType) | |
1026 | + | |
1027 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
1028 | + sequences = [words, children_positions,y,words_indexes], | |
1029 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))], | |
1030 | + non_sequences = learning_rate, | |
1031 | + n_steps = words.shape[0]) | |
1032 | + | |
1033 | + cost = T.sum(cross_entropy_vector[0]) | |
1034 | + | |
1035 | + updates = OrderedDict([ | |
1036 | + (self.W_h_h2, self.W_h_h2-learning_rate*T.grad(cost, self.W_h_h2)), | |
1037 | + (self.W_h2_h3, self.W_h2_h3-learning_rate*T.grad(cost, self.W_h2_h3)), | |
1038 | + (self.W_h3_y, self.W_h3_y-learning_rate*T.grad(cost, self.W_h3_y)), | |
1039 | + (self.W_e_eh, self.W_e_eh-learning_rate*T.grad(cost, self.W_e_eh)), | |
1040 | + (self.W_eh_h, self.W_eh_h-learning_rate*T.grad(cost, self.W_eh_h)), | |
1041 | + (self.W_shh_h, self.W_shh_h-learning_rate*T.grad(cost, self.W_shh_h)), | |
1042 | + (self.W_sh_shh, self.W_sh_shh-learning_rate*T.grad(cost, self.W_sh_shh)), | |
1043 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
1044 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
1045 | + ]) | |
1046 | + | |
1047 | + self.train = theano.function( inputs = [words, children_positions, y, words_indexes, learning_rate], | |
1048 | + outputs = [], | |
1049 | + updates = updates, | |
1050 | + allow_input_downcast=True, | |
1051 | + mode='FAST_RUN' | |
1052 | + ) | |
1053 | + | |
1054 | + | |
1055 | + def one_step_classify(word_id, word_children_positions, i, hidden_states): | |
1056 | + | |
1057 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
1058 | + | |
1059 | + tmp = word_children_positions>=0.0 | |
1060 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
1061 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
1062 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
1063 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
1064 | + schh = schh/number_of_children | |
1065 | + | |
1066 | + eh = T.tanh(T.dot(self.emb[word_id],self.W_e_eh)) | |
1067 | + shh = T.tanh(T.dot(schh,self.W_sh_shh)) | |
1068 | + | |
1069 | + h = T.tanh(T.concatenate([T.dot(eh, self.W_eh_h), T.dot(shh,self.W_shh_h)])) | |
1070 | + | |
1071 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
1072 | + | |
1073 | + h3 = T.tanh(T.dot(h2, self.W_h2_h3)) | |
1074 | + | |
1075 | + current_hidden_state = hidden_states[i] | |
1076 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
1077 | + | |
1078 | + y_prob = T.nnet.softmax(T.dot(h3,self.W_h3_y) + self.b_y)[0] | |
1079 | + | |
1080 | + return y_prob, hidden_states_new | |
1081 | + | |
1082 | + | |
1083 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
1084 | + fn=one_step_classify, | |
1085 | + sequences = [words, children_positions,words_indexes], | |
1086 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))]) | |
1087 | + | |
1088 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
1089 | + sequences = [words_indexes]) | |
1090 | + | |
1091 | + self.classify = theano.function(inputs=[words,children_positions,words_indexes], | |
1092 | + outputs=predictions, | |
1093 | + allow_input_downcast=True, | |
1094 | + mode='FAST_RUN' | |
1095 | + ) | |
1096 | + | |
1097 | + | |
1098 | +class model55_pf9(object): | |
1099 | + def __init__(self, ne, nchd, nh2, nh3, nc, w2v_model_path, max_phrase_length): | |
1100 | + ''' | |
1101 | + nh :: dimension of hidden state | |
1102 | + nc :: number of classes | |
1103 | + ne :: number of word embeddings in the vocabulary | |
1104 | + de :: dimension of the word embeddings | |
1105 | + ds :: dimension of the sentiment state | |
1106 | + ''' | |
1107 | + self.max_phrase_length = max_phrase_length | |
1108 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
1109 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
1110 | + self.words2ids = w2vecs["words2ids"] | |
1111 | + | |
1112 | + #ne = len(w2vecs["words2ids"]) | |
1113 | + de = w2vecs["vectors"].shape[1] | |
1114 | + del w2vecs | |
1115 | + | |
1116 | + r = 0.05 | |
1117 | + self.W_e_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (de, ne)).astype(theano.config.floatX)) | |
1118 | + | |
1119 | + | |
1120 | + self.W_h_h2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nh2)).astype(theano.config.floatX)) | |
1121 | + self.W_h2_h3 = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh2, nh3)).astype(theano.config.floatX)) | |
1122 | + self.W_h3_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh3, nc)).astype(theano.config.floatX)) | |
1123 | + | |
1124 | + self.W_sh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nchd)).astype(theano.config.floatX)) | |
1125 | + | |
1126 | + | |
1127 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
1128 | + | |
1129 | + | |
1130 | + def one_step(word_id, word_children_positions, y_true, i, hidden_states, learning_rate): | |
1131 | + | |
1132 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
1133 | + | |
1134 | + tmp = word_children_positions>=0.0 | |
1135 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
1136 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
1137 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
1138 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
1139 | + schh = schh/number_of_children | |
1140 | + | |
1141 | + | |
1142 | + h = T.tanh(T.concatenate([T.dot(self.emb[word_id],self.W_e_h), T.dot(schh,self.W_sh_h)])) | |
1143 | + | |
1144 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
1145 | + | |
1146 | + | |
1147 | + h3 = T.tanh(T.dot(h2, self.W_h2_h3)) | |
1148 | + | |
1149 | + current_hidden_state = hidden_states[i] | |
1150 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
1151 | + | |
1152 | + y_prob = T.nnet.softmax(T.dot(h3,self.W_h3_y) + self.b_y)[0] | |
1153 | + | |
1154 | + #l2_norm = T.sum(self.W_h_h2**2) + T.sum(self.W_h2_y**2) + T.sum(self.W_e_h**2) + T.sum(self.W_sh_h**2) + T.sum(self.emb**2) + T.sum(self.b_h**2) + T.sum(self.b_y**2) | |
1155 | + | |
1156 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
1157 | + | |
1158 | + return cross_entropy, hidden_states_new | |
1159 | + | |
1160 | + | |
1161 | + y = T.vector('y',dtype=dataType) | |
1162 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
1163 | + words = T.vector(dtype=dataType) | |
1164 | + children_positions = T.matrix(dtype=dataType) | |
1165 | + words_indexes = T.vector(dtype=dataType) | |
1166 | + | |
1167 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
1168 | + sequences = [words, children_positions,y,words_indexes], | |
1169 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))], | |
1170 | + non_sequences = learning_rate, | |
1171 | + n_steps = words.shape[0]) | |
1172 | + | |
1173 | + cost = T.sum(cross_entropy_vector[0]) | |
1174 | + | |
1175 | + updates = OrderedDict([ | |
1176 | + (self.W_h_h2, self.W_h_h2-learning_rate*T.grad(cost, self.W_h_h2)), | |
1177 | + (self.W_h2_h3, self.W_h2_h3-learning_rate*T.grad(cost, self.W_h2_h3)), | |
1178 | + (self.W_h3_y, self.W_h3_y-learning_rate*T.grad(cost, self.W_h3_y)), | |
1179 | + (self.W_e_h, self.W_e_h-learning_rate*T.grad(cost, self.W_e_h)), | |
1180 | + (self.W_sh_h, self.W_sh_h-learning_rate*T.grad(cost, self.W_sh_h)), | |
1181 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
1182 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
1183 | + ]) | |
1184 | + | |
1185 | + self.train = theano.function( inputs = [words, children_positions, y, words_indexes, learning_rate], | |
1186 | + outputs = [], | |
1187 | + updates = updates, | |
1188 | + allow_input_downcast=True, | |
1189 | + mode='FAST_RUN' | |
1190 | + ) | |
1191 | + | |
1192 | + | |
1193 | + def one_step_classify(word_id, word_children_positions, i, hidden_states): | |
1194 | + | |
1195 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
1196 | + | |
1197 | + tmp = word_children_positions>=0.0 | |
1198 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
1199 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
1200 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
1201 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
1202 | + schh = schh/number_of_children | |
1203 | + | |
1204 | + | |
1205 | + h = T.tanh(T.concatenate([T.dot(self.emb[word_id],self.W_e_h), T.dot(schh,self.W_sh_h)])) | |
1206 | + | |
1207 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
1208 | + | |
1209 | + h3 = T.tanh(T.dot(h2, self.W_h2_h3)) | |
1210 | + | |
1211 | + current_hidden_state = hidden_states[i] | |
1212 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
1213 | + | |
1214 | + y_prob = T.nnet.softmax(T.dot(h3,self.W_h3_y) + self.b_y)[0] | |
1215 | + | |
1216 | + return y_prob, hidden_states_new | |
1217 | + | |
1218 | + | |
1219 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
1220 | + fn=one_step_classify, | |
1221 | + sequences = [words, children_positions,words_indexes], | |
1222 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))]) | |
1223 | + | |
1224 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
1225 | + sequences = [words_indexes]) | |
1226 | + | |
1227 | + self.classify = theano.function(inputs=[words,children_positions,words_indexes], | |
1228 | + outputs=predictions, | |
1229 | + allow_input_downcast=True, | |
1230 | + mode='FAST_RUN' | |
1231 | + ) | |
1232 | + | |
1233 | + | |
1234 | + | |
1235 | +class model55_pf10(object): | |
1236 | + def __init__(self, nchd, nh2, nc, w2v_model_path, max_phrase_length): | |
1237 | + ''' | |
1238 | + nh :: dimension of hidden state | |
1239 | + nc :: number of classes | |
1240 | + ne :: number of word embeddings in the vocabulary | |
1241 | + de :: dimension of the word embeddings | |
1242 | + ds :: dimension of the sentiment state | |
1243 | + ''' | |
1244 | + self.max_phrase_length = max_phrase_length | |
1245 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
1246 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
1247 | + self.words2ids = w2vecs["words2ids"] | |
1248 | + | |
1249 | + #ne = len(w2vecs["words2ids"]) | |
1250 | + de = w2vecs["vectors"].shape[1] | |
1251 | + del w2vecs | |
1252 | + | |
1253 | + r = 0.05 | |
1254 | + | |
1255 | + self.W_h_h2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (de+nchd, nh2)).astype(theano.config.floatX)) | |
1256 | + self.W_h2_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh2, nc)).astype(theano.config.floatX)) | |
1257 | + | |
1258 | + self.W_sh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (de+nchd, nchd)).astype(theano.config.floatX)) | |
1259 | + | |
1260 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
1261 | + | |
1262 | + | |
1263 | + def one_step(word_id, word_children_positions, y_true, i, hidden_states, learning_rate): | |
1264 | + | |
1265 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
1266 | + | |
1267 | + tmp = word_children_positions>=0.0 | |
1268 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
1269 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
1270 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
1271 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
1272 | + schh = schh/number_of_children | |
1273 | + | |
1274 | + | |
1275 | + h = T.tanh(T.concatenate([self.emb[word_id], T.dot(schh,self.W_sh_h)])) | |
1276 | + | |
1277 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
1278 | + | |
1279 | + | |
1280 | + current_hidden_state = hidden_states[i] | |
1281 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
1282 | + | |
1283 | + y_prob = T.nnet.softmax(T.dot(h2,self.W_h2_y) + self.b_y)[0] | |
1284 | + | |
1285 | + #l2_norm = T.sum(self.W_h_h2**2) + T.sum(self.W_h2_y**2) + T.sum(self.W_e_h**2) + T.sum(self.W_sh_h**2) + T.sum(self.emb**2) + T.sum(self.b_h**2) + T.sum(self.b_y**2) | |
1286 | + | |
1287 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
1288 | + | |
1289 | + return cross_entropy, hidden_states_new | |
1290 | + | |
1291 | + | |
1292 | + y = T.vector('y',dtype=dataType) | |
1293 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
1294 | + words = T.vector(dtype=dataType) | |
1295 | + children_positions = T.matrix(dtype=dataType) | |
1296 | + words_indexes = T.vector(dtype=dataType) | |
1297 | + | |
1298 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
1299 | + sequences = [words, children_positions,y,words_indexes], | |
1300 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,de+nchd), dtype = theano.config.floatX))], | |
1301 | + non_sequences = learning_rate, | |
1302 | + n_steps = words.shape[0]) | |
1303 | + | |
1304 | + cost = T.sum(cross_entropy_vector[0]) | |
1305 | + | |
1306 | + updates = OrderedDict([ | |
1307 | + (self.W_h_h2, self.W_h_h2-learning_rate*T.grad(cost, self.W_h_h2)), | |
1308 | + (self.W_h2_y, self.W_h2_y-learning_rate*T.grad(cost, self.W_h2_y)), | |
1309 | + (self.W_sh_h, self.W_sh_h-learning_rate*T.grad(cost, self.W_sh_h)), | |
1310 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
1311 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
1312 | + ]) | |
1313 | + | |
1314 | + self.train = theano.function( inputs = [words, children_positions, y, words_indexes, learning_rate], | |
1315 | + outputs = [], | |
1316 | + updates = updates, | |
1317 | + allow_input_downcast=True, | |
1318 | + mode='FAST_RUN' | |
1319 | + ) | |
1320 | + | |
1321 | + | |
1322 | + def one_step_classify(word_id, word_children_positions, i, hidden_states): | |
1323 | + | |
1324 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
1325 | + | |
1326 | + tmp = word_children_positions>=0.0 | |
1327 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
1328 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
1329 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
1330 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
1331 | + schh = schh/number_of_children | |
1332 | + | |
1333 | + h = T.tanh(T.concatenate([self.emb[word_id], T.dot(schh,self.W_sh_h)])) | |
1334 | + | |
1335 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
1336 | + | |
1337 | + | |
1338 | + current_hidden_state = hidden_states[i] | |
1339 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
1340 | + | |
1341 | + y_prob = T.nnet.softmax(T.dot(h2,self.W_h2_y) + self.b_y)[0] | |
1342 | + | |
1343 | + return y_prob, hidden_states_new | |
1344 | + | |
1345 | + | |
1346 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
1347 | + fn=one_step_classify, | |
1348 | + sequences = [words, children_positions,words_indexes], | |
1349 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,de+nchd), dtype = theano.config.floatX))]) | |
1350 | + | |
1351 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
1352 | + sequences = [words_indexes]) | |
1353 | + | |
1354 | + self.classify = theano.function(inputs=[words,children_positions,words_indexes], | |
1355 | + outputs=predictions, | |
1356 | + allow_input_downcast=True, | |
1357 | + mode='FAST_RUN' | |
1358 | + ) | |
1359 | + | |
1360 | + | |
1361 | + | |
1362 | + | |
1363 | +class model55_pf11(object): | |
1364 | + def __init__(self, nchd, nh2, nh3, nc, w2v_model_path, max_phrase_length): | |
1365 | + ''' | |
1366 | + nh :: dimension of hidden state | |
1367 | + nc :: number of classes | |
1368 | + ne :: number of word embeddings in the vocabulary | |
1369 | + de :: dimension of the word embeddings | |
1370 | + ds :: dimension of the sentiment state | |
1371 | + ''' | |
1372 | + self.max_phrase_length = max_phrase_length | |
1373 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
1374 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
1375 | + self.words2ids = w2vecs["words2ids"] | |
1376 | + | |
1377 | + #ne = len(w2vecs["words2ids"]) | |
1378 | + de = w2vecs["vectors"].shape[1] | |
1379 | + del w2vecs | |
1380 | + | |
1381 | + r = 0.05 | |
1382 | + | |
1383 | + | |
1384 | + self.W_h_h2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (de+nchd, nh2)).astype(theano.config.floatX)) | |
1385 | + self.W_h2_h3 = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh2, nh3)).astype(theano.config.floatX)) | |
1386 | + self.W_h3_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh3, nc)).astype(theano.config.floatX)) | |
1387 | + | |
1388 | + self.W_sh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (de+nchd, nchd)).astype(theano.config.floatX)) | |
1389 | + | |
1390 | + | |
1391 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
1392 | + | |
1393 | + | |
1394 | + def one_step(word_id, word_children_positions, y_true, i, hidden_states, learning_rate): | |
1395 | + | |
1396 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
1397 | + | |
1398 | + tmp = word_children_positions>=0.0 | |
1399 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
1400 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
1401 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
1402 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
1403 | + schh = schh/number_of_children | |
1404 | + | |
1405 | + | |
1406 | + h = T.tanh(T.concatenate([self.emb[word_id], T.dot(schh,self.W_sh_h)])) | |
1407 | + | |
1408 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
1409 | + | |
1410 | + | |
1411 | + h3 = T.tanh(T.dot(h2, self.W_h2_h3)) | |
1412 | + | |
1413 | + current_hidden_state = hidden_states[i] | |
1414 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
1415 | + | |
1416 | + y_prob = T.nnet.softmax(T.dot(h3,self.W_h3_y) + self.b_y)[0] | |
1417 | + | |
1418 | + #l2_norm = T.sum(self.W_h_h2**2) + T.sum(self.W_h2_y**2) + T.sum(self.W_e_h**2) + T.sum(self.W_sh_h**2) + T.sum(self.emb**2) + T.sum(self.b_h**2) + T.sum(self.b_y**2) | |
1419 | + | |
1420 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
1421 | + | |
1422 | + return cross_entropy, hidden_states_new | |
1423 | + | |
1424 | + | |
1425 | + y = T.vector('y',dtype=dataType) | |
1426 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
1427 | + words = T.vector(dtype=dataType) | |
1428 | + children_positions = T.matrix(dtype=dataType) | |
1429 | + words_indexes = T.vector(dtype=dataType) | |
1430 | + | |
1431 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
1432 | + sequences = [words, children_positions,y,words_indexes], | |
1433 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,de+nchd), dtype = theano.config.floatX))], | |
1434 | + non_sequences = learning_rate, | |
1435 | + n_steps = words.shape[0]) | |
1436 | + | |
1437 | + cost = T.sum(cross_entropy_vector[0]) | |
1438 | + | |
1439 | + updates = OrderedDict([ | |
1440 | + (self.W_h_h2, self.W_h_h2-learning_rate*T.grad(cost, self.W_h_h2)), | |
1441 | + (self.W_h2_h3, self.W_h2_h3-learning_rate*T.grad(cost, self.W_h2_h3)), | |
1442 | + (self.W_h3_y, self.W_h3_y-learning_rate*T.grad(cost, self.W_h3_y)), | |
1443 | + (self.W_sh_h, self.W_sh_h-learning_rate*T.grad(cost, self.W_sh_h)), | |
1444 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
1445 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
1446 | + ]) | |
1447 | + | |
1448 | + self.train = theano.function( inputs = [words, children_positions, y, words_indexes, learning_rate], | |
1449 | + outputs = [], | |
1450 | + updates = updates, | |
1451 | + allow_input_downcast=True, | |
1452 | + mode='FAST_RUN' | |
1453 | + ) | |
1454 | + | |
1455 | + | |
1456 | + def one_step_classify(word_id, word_children_positions, i, hidden_states): | |
1457 | + | |
1458 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
1459 | + | |
1460 | + tmp = word_children_positions>=0.0 | |
1461 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
1462 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
1463 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
1464 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
1465 | + schh = schh/number_of_children | |
1466 | + | |
1467 | + | |
1468 | + h = T.tanh(T.concatenate([self.emb[word_id], T.dot(schh,self.W_sh_h)])) | |
1469 | + | |
1470 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
1471 | + | |
1472 | + h3 = T.tanh(T.dot(h2, self.W_h2_h3)) | |
1473 | + | |
1474 | + current_hidden_state = hidden_states[i] | |
1475 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
1476 | + | |
1477 | + y_prob = T.nnet.softmax(T.dot(h3,self.W_h3_y) + self.b_y)[0] | |
1478 | + | |
1479 | + return y_prob, hidden_states_new | |
1480 | + | |
1481 | + | |
1482 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
1483 | + fn=one_step_classify, | |
1484 | + sequences = [words, children_positions,words_indexes], | |
1485 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,de+nchd), dtype = theano.config.floatX))]) | |
1486 | + | |
1487 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
1488 | + sequences = [words_indexes]) | |
1489 | + | |
1490 | + self.classify = theano.function(inputs=[words,children_positions,words_indexes], | |
1491 | + outputs=predictions, | |
1492 | + allow_input_downcast=True, | |
1493 | + mode='FAST_RUN' | |
1494 | + ) | |
1495 | + | |
1496 | + | |
1497 | + | |
1498 | + | |
1499 | + | |
1500 | +class model55_pf12(object): | |
1501 | + def __init__(self, neh, ne, nshh, nchd, nh2, nh3, nh4, nc, w2v_model_path, max_phrase_length): | |
1502 | + ''' | |
1503 | + nh :: dimension of hidden state | |
1504 | + nc :: number of classes | |
1505 | + ne :: number of word embeddings in the vocabulary | |
1506 | + de :: dimension of the word embeddings | |
1507 | + ds :: dimension of the sentiment state | |
1508 | + ''' | |
1509 | + self.max_phrase_length = max_phrase_length | |
1510 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
1511 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
1512 | + self.words2ids = w2vecs["words2ids"] | |
1513 | + | |
1514 | + #ne = len(w2vecs["words2ids"]) | |
1515 | + de = w2vecs["vectors"].shape[1] | |
1516 | + del w2vecs | |
1517 | + | |
1518 | + r = 0.05 | |
1519 | + self.W_e_eh = theano.shared(r * np.random.uniform(-1.0, 1.0, (de, neh)).astype(theano.config.floatX)) | |
1520 | + | |
1521 | + self.W_eh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (neh, ne)).astype(theano.config.floatX)) | |
1522 | + | |
1523 | + self.W_h_h2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nh2)).astype(theano.config.floatX)) | |
1524 | + self.W_h2_h3 = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh2, nh3)).astype(theano.config.floatX)) | |
1525 | + self.W_h3_h4 = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh3, nh4)).astype(theano.config.floatX)) | |
1526 | + self.W_h4_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh4, nc)).astype(theano.config.floatX)) | |
1527 | + | |
1528 | + self.W_sh_shh = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nshh)).astype(theano.config.floatX)) | |
1529 | + | |
1530 | + self.W_shh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (nshh, nchd)).astype(theano.config.floatX)) | |
1531 | + | |
1532 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
1533 | + | |
1534 | + | |
1535 | + def one_step(word_id, word_children_positions, y_true, i, hidden_states, learning_rate): | |
1536 | + | |
1537 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
1538 | + | |
1539 | + tmp = word_children_positions>=0.0 | |
1540 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
1541 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
1542 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
1543 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
1544 | + schh = schh/number_of_children | |
1545 | + | |
1546 | + eh = T.tanh(T.dot(self.emb[word_id],self.W_e_eh)) | |
1547 | + shh = T.tanh(T.dot(schh,self.W_sh_shh)) | |
1548 | + | |
1549 | + h = T.tanh(T.concatenate([T.dot(eh, self.W_eh_h), T.dot(shh,self.W_shh_h)])) | |
1550 | + | |
1551 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
1552 | + | |
1553 | + | |
1554 | + h3 = T.tanh(T.dot(h2, self.W_h2_h3)) | |
1555 | + h4 = T.tanh(T.dot(h3, self.W_h3_h4)) | |
1556 | + | |
1557 | + current_hidden_state = hidden_states[i] | |
1558 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
1559 | + | |
1560 | + y_prob = T.nnet.softmax(T.dot(h4,self.W_h4_y) + self.b_y)[0] | |
1561 | + | |
1562 | + #l2_norm = T.sum(self.W_h_h2**2) + T.sum(self.W_h2_y**2) + T.sum(self.W_e_h**2) + T.sum(self.W_sh_h**2) + T.sum(self.emb**2) + T.sum(self.b_h**2) + T.sum(self.b_y**2) | |
1563 | + | |
1564 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
1565 | + | |
1566 | + return cross_entropy, hidden_states_new | |
1567 | + | |
1568 | + | |
1569 | + y = T.vector('y',dtype=dataType) | |
1570 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
1571 | + words = T.vector(dtype=dataType) | |
1572 | + children_positions = T.matrix(dtype=dataType) | |
1573 | + words_indexes = T.vector(dtype=dataType) | |
1574 | + | |
1575 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
1576 | + sequences = [words, children_positions,y,words_indexes], | |
1577 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))], | |
1578 | + non_sequences = learning_rate, | |
1579 | + n_steps = words.shape[0]) | |
1580 | + | |
1581 | + cost = T.sum(cross_entropy_vector[0]) | |
1582 | + | |
1583 | + updates = OrderedDict([ | |
1584 | + (self.W_h_h2, self.W_h_h2-learning_rate*T.grad(cost, self.W_h_h2)), | |
1585 | + (self.W_h2_h3, self.W_h2_h3-learning_rate*T.grad(cost, self.W_h2_h3)), | |
1586 | + (self.W_h3_h4, self.W_h3_h4-learning_rate*T.grad(cost, self.W_h3_h4)), | |
1587 | + (self.W_h4_y, self.W_h4_y-learning_rate*T.grad(cost, self.W_h4_y)), | |
1588 | + (self.W_e_eh, self.W_e_eh-learning_rate*T.grad(cost, self.W_e_eh)), | |
1589 | + (self.W_eh_h, self.W_eh_h-learning_rate*T.grad(cost, self.W_eh_h)), | |
1590 | + (self.W_shh_h, self.W_shh_h-learning_rate*T.grad(cost, self.W_shh_h)), | |
1591 | + (self.W_sh_shh, self.W_sh_shh-learning_rate*T.grad(cost, self.W_sh_shh)), | |
1592 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
1593 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
1594 | + ]) | |
1595 | + | |
1596 | + self.train = theano.function( inputs = [words, children_positions, y, words_indexes, learning_rate], | |
1597 | + outputs = [], | |
1598 | + updates = updates, | |
1599 | + allow_input_downcast=True, | |
1600 | + mode='FAST_RUN' | |
1601 | + ) | |
1602 | + | |
1603 | + | |
1604 | + def one_step_classify(word_id, word_children_positions, i, hidden_states): | |
1605 | + | |
1606 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
1607 | + | |
1608 | + tmp = word_children_positions>=0.0 | |
1609 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
1610 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
1611 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
1612 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
1613 | + schh = schh/number_of_children | |
1614 | + | |
1615 | + eh = T.tanh(T.dot(self.emb[word_id],self.W_e_eh)) | |
1616 | + shh = T.tanh(T.dot(schh,self.W_sh_shh)) | |
1617 | + | |
1618 | + h = T.tanh(T.concatenate([T.dot(eh, self.W_eh_h), T.dot(shh,self.W_shh_h)])) | |
1619 | + | |
1620 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
1621 | + | |
1622 | + h3 = T.tanh(T.dot(h2, self.W_h2_h3)) | |
1623 | + h4 = T.tanh(T.dot(h3, self.W_h3_h4)) | |
1624 | + | |
1625 | + current_hidden_state = hidden_states[i] | |
1626 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
1627 | + | |
1628 | + y_prob = T.nnet.softmax(T.dot(h4,self.W_h4_y) + self.b_y)[0] | |
1629 | + | |
1630 | + return y_prob, hidden_states_new | |
1631 | + | |
1632 | + | |
1633 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
1634 | + fn=one_step_classify, | |
1635 | + sequences = [words, children_positions,words_indexes], | |
1636 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))]) | |
1637 | + | |
1638 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
1639 | + sequences = [words_indexes]) | |
1640 | + | |
1641 | + self.classify = theano.function(inputs=[words,children_positions,words_indexes], | |
1642 | + outputs=predictions, | |
1643 | + allow_input_downcast=True, | |
1644 | + mode='FAST_RUN' | |
1645 | + ) | |
1646 | + | |
1647 | + | |
1648 | + | |
1649 | +class model55_pf13(object): | |
1650 | + def __init__(self, neh, neh2, ne, nshh, nshh2, nchd, nc, w2v_model_path, max_phrase_length): | |
1651 | + ''' | |
1652 | + nh :: dimension of hidden state | |
1653 | + nc :: number of classes | |
1654 | + ne :: number of word embeddings in the vocabulary | |
1655 | + de :: dimension of the word embeddings | |
1656 | + ds :: dimension of the sentiment state | |
1657 | + ''' | |
1658 | + self.max_phrase_length = max_phrase_length | |
1659 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
1660 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
1661 | + self.words2ids = w2vecs["words2ids"] | |
1662 | + | |
1663 | + #ne = len(w2vecs["words2ids"]) | |
1664 | + de = w2vecs["vectors"].shape[1] | |
1665 | + del w2vecs | |
1666 | + | |
1667 | + r = 0.05 | |
1668 | + self.W_e_eh = theano.shared(r * np.random.uniform(-1.0, 1.0, (de, neh)).astype(theano.config.floatX)) | |
1669 | + self.W_eh_eh2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (neh, neh2)).astype(theano.config.floatX)) | |
1670 | + self.W_eh2_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (neh2, ne)).astype(theano.config.floatX)) | |
1671 | + | |
1672 | + self.W_sh_shh = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nshh)).astype(theano.config.floatX)) | |
1673 | + self.W_shh_shh2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (nshh, nshh2)).astype(theano.config.floatX)) | |
1674 | + self.W_shh2_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (nshh2, nchd)).astype(theano.config.floatX)) | |
1675 | + | |
1676 | + self.W_h_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nc)).astype(theano.config.floatX)) | |
1677 | + | |
1678 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
1679 | + | |
1680 | + | |
1681 | + def one_step(word_id, word_children_positions, y_true, i, hidden_states, learning_rate): | |
1682 | + | |
1683 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
1684 | + | |
1685 | + tmp = word_children_positions>=0.0 | |
1686 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
1687 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
1688 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
1689 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
1690 | + schh = schh/number_of_children | |
1691 | + | |
1692 | + eh = T.tanh(T.dot(self.emb[word_id],self.W_e_eh)) | |
1693 | + eh2 = T.tanh(T.dot(eh,self.W_eh_eh2)) | |
1694 | + | |
1695 | + shh = T.tanh(T.dot(schh,self.W_sh_shh)) | |
1696 | + shh2 = T.tanh(T.dot(shh,self.W_shh_shh2)) | |
1697 | + | |
1698 | + | |
1699 | + h = T.tanh(T.concatenate([T.dot(eh2, self.W_eh2_h), T.dot(shh2,self.W_shh2_h)])) | |
1700 | + | |
1701 | + | |
1702 | + current_hidden_state = hidden_states[i] | |
1703 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
1704 | + | |
1705 | + y_prob = T.nnet.softmax(T.dot(h,self.W_h_y) + self.b_y)[0] | |
1706 | + | |
1707 | + #l2_norm = T.sum(self.W_h_h2**2) + T.sum(self.W_h2_y**2) + T.sum(self.W_e_h**2) + T.sum(self.W_sh_h**2) + T.sum(self.emb**2) + T.sum(self.b_h**2) + T.sum(self.b_y**2) | |
1708 | + | |
1709 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
1710 | + | |
1711 | + return cross_entropy, hidden_states_new | |
1712 | + | |
1713 | + | |
1714 | + y = T.vector('y',dtype=dataType) | |
1715 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
1716 | + words = T.vector(dtype=dataType) | |
1717 | + children_positions = T.matrix(dtype=dataType) | |
1718 | + words_indexes = T.vector(dtype=dataType) | |
1719 | + | |
1720 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
1721 | + sequences = [words, children_positions,y,words_indexes], | |
1722 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))], | |
1723 | + non_sequences = learning_rate, | |
1724 | + n_steps = words.shape[0]) | |
1725 | + | |
1726 | + cost = T.sum(cross_entropy_vector[0]) | |
1727 | + | |
1728 | + updates = OrderedDict([ | |
1729 | + (self.W_h_y, self.W_h_y-learning_rate*T.grad(cost, self.W_h_y)), | |
1730 | + (self.W_e_eh, self.W_e_eh-learning_rate*T.grad(cost, self.W_e_eh)), | |
1731 | + (self.W_eh_eh2, self.W_eh_eh2-learning_rate*T.grad(cost, self.W_eh_eh2)), | |
1732 | + (self.W_eh2_h, self.W_eh2_h-learning_rate*T.grad(cost, self.W_eh2_h)), | |
1733 | + (self.W_shh2_h, self.W_shh2_h-learning_rate*T.grad(cost, self.W_shh2_h)), | |
1734 | + (self.W_sh_shh, self.W_sh_shh-learning_rate*T.grad(cost, self.W_sh_shh)), | |
1735 | + (self.W_shh_shh2, self.W_shh_shh2-learning_rate*T.grad(cost, self.W_shh_shh2)), | |
1736 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
1737 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
1738 | + ]) | |
1739 | + | |
1740 | + self.train = theano.function( inputs = [words, children_positions, y, words_indexes, learning_rate], | |
1741 | + outputs = [], | |
1742 | + updates = updates, | |
1743 | + allow_input_downcast=True, | |
1744 | + mode='FAST_RUN' | |
1745 | + ) | |
1746 | + | |
1747 | + | |
1748 | + def one_step_classify(word_id, word_children_positions, i, hidden_states): | |
1749 | + | |
1750 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
1751 | + | |
1752 | + tmp = word_children_positions>=0.0 | |
1753 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
1754 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
1755 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
1756 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
1757 | + schh = schh/number_of_children | |
1758 | + | |
1759 | + eh = T.tanh(T.dot(self.emb[word_id],self.W_e_eh)) | |
1760 | + eh2 = T.tanh(T.dot(eh,self.W_eh_eh2)) | |
1761 | + | |
1762 | + shh = T.tanh(T.dot(schh,self.W_sh_shh)) | |
1763 | + shh2 = T.tanh(T.dot(shh,self.W_shh_shh2)) | |
1764 | + | |
1765 | + | |
1766 | + h = T.tanh(T.concatenate([T.dot(eh2, self.W_eh2_h), T.dot(shh2,self.W_shh2_h)])) | |
1767 | + | |
1768 | + | |
1769 | + current_hidden_state = hidden_states[i] | |
1770 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
1771 | + | |
1772 | + y_prob = T.nnet.softmax(T.dot(h,self.W_h_y) + self.b_y)[0] | |
1773 | + | |
1774 | + | |
1775 | + return y_prob, hidden_states_new | |
1776 | + | |
1777 | + | |
1778 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
1779 | + fn=one_step_classify, | |
1780 | + sequences = [words, children_positions,words_indexes], | |
1781 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))]) | |
1782 | + | |
1783 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
1784 | + sequences = [words_indexes]) | |
1785 | + | |
1786 | + self.classify = theano.function(inputs=[words,children_positions,words_indexes], | |
1787 | + outputs=predictions, | |
1788 | + allow_input_downcast=True, | |
1789 | + mode='FAST_RUN' | |
1790 | + ) | |
1791 | + | |
1792 | + | |
1793 | + | |
1794 | +class model55_pf14(object): | |
1795 | + def __init__(self, neh, neh2, ne, nshh, nshh2, nchd, nh2, nc, w2v_model_path, max_phrase_length): | |
1796 | + ''' | |
1797 | + nh :: dimension of hidden state | |
1798 | + nc :: number of classes | |
1799 | + ne :: number of word embeddings in the vocabulary | |
1800 | + de :: dimension of the word embeddings | |
1801 | + ds :: dimension of the sentiment state | |
1802 | + ''' | |
1803 | + self.max_phrase_length = max_phrase_length | |
1804 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
1805 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
1806 | + self.words2ids = w2vecs["words2ids"] | |
1807 | + | |
1808 | + #ne = len(w2vecs["words2ids"]) | |
1809 | + de = w2vecs["vectors"].shape[1] | |
1810 | + del w2vecs | |
1811 | + | |
1812 | + r = 0.05 | |
1813 | + self.W_e_eh = theano.shared(r * np.random.uniform(-1.0, 1.0, (de, neh)).astype(theano.config.floatX)) | |
1814 | + self.W_eh_eh2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (neh, neh2)).astype(theano.config.floatX)) | |
1815 | + self.W_eh2_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (neh2, ne)).astype(theano.config.floatX)) | |
1816 | + | |
1817 | + self.W_sh_shh = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nshh)).astype(theano.config.floatX)) | |
1818 | + self.W_shh_shh2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (nshh, nshh2)).astype(theano.config.floatX)) | |
1819 | + self.W_shh2_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (nshh2, nchd)).astype(theano.config.floatX)) | |
1820 | + | |
1821 | + self.W_h_h2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nh2)).astype(theano.config.floatX)) | |
1822 | + self.W_h2_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh2, nc)).astype(theano.config.floatX)) | |
1823 | + | |
1824 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
1825 | + | |
1826 | + | |
1827 | + def one_step(word_id, word_children_positions, y_true, i, hidden_states, learning_rate): | |
1828 | + | |
1829 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
1830 | + | |
1831 | + tmp = word_children_positions>=0.0 | |
1832 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
1833 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
1834 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
1835 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
1836 | + schh = schh/number_of_children | |
1837 | + | |
1838 | + eh = T.tanh(T.dot(self.emb[word_id],self.W_e_eh)) | |
1839 | + eh2 = T.tanh(T.dot(eh,self.W_eh_eh2)) | |
1840 | + | |
1841 | + shh = T.tanh(T.dot(schh,self.W_sh_shh)) | |
1842 | + shh2 = T.tanh(T.dot(shh,self.W_shh_shh2)) | |
1843 | + | |
1844 | + | |
1845 | + h = T.tanh(T.concatenate([T.dot(eh2, self.W_eh2_h), T.dot(shh2,self.W_shh2_h)])) | |
1846 | + | |
1847 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
1848 | + | |
1849 | + current_hidden_state = hidden_states[i] | |
1850 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
1851 | + | |
1852 | + y_prob = T.nnet.softmax(T.dot(h2,self.W_h2_y) + self.b_y)[0] | |
1853 | + | |
1854 | + #l2_norm = T.sum(self.W_h_h2**2) + T.sum(self.W_h2_y**2) + T.sum(self.W_e_h**2) + T.sum(self.W_sh_h**2) + T.sum(self.emb**2) + T.sum(self.b_h**2) + T.sum(self.b_y**2) | |
1855 | + | |
1856 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
1857 | + | |
1858 | + return cross_entropy, hidden_states_new | |
1859 | + | |
1860 | + | |
1861 | + y = T.vector('y',dtype=dataType) | |
1862 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
1863 | + words = T.vector(dtype=dataType) | |
1864 | + children_positions = T.matrix(dtype=dataType) | |
1865 | + words_indexes = T.vector(dtype=dataType) | |
1866 | + | |
1867 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
1868 | + sequences = [words, children_positions,y,words_indexes], | |
1869 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))], | |
1870 | + non_sequences = learning_rate, | |
1871 | + n_steps = words.shape[0]) | |
1872 | + | |
1873 | + cost = T.sum(cross_entropy_vector[0]) | |
1874 | + | |
1875 | + updates = OrderedDict([ | |
1876 | + (self.W_h_h2, self.W_h_h2-learning_rate*T.grad(cost, self.W_h_h2)), | |
1877 | + (self.W_h2_y, self.W_h2_y-learning_rate*T.grad(cost, self.W_h2_y)), | |
1878 | + (self.W_e_eh, self.W_e_eh-learning_rate*T.grad(cost, self.W_e_eh)), | |
1879 | + (self.W_eh_eh2, self.W_eh_eh2-learning_rate*T.grad(cost, self.W_eh_eh2)), | |
1880 | + (self.W_eh2_h, self.W_eh2_h-learning_rate*T.grad(cost, self.W_eh2_h)), | |
1881 | + (self.W_shh2_h, self.W_shh2_h-learning_rate*T.grad(cost, self.W_shh2_h)), | |
1882 | + (self.W_sh_shh, self.W_sh_shh-learning_rate*T.grad(cost, self.W_sh_shh)), | |
1883 | + (self.W_shh_shh2, self.W_shh_shh2-learning_rate*T.grad(cost, self.W_shh_shh2)), | |
1884 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
1885 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
1886 | + ]) | |
1887 | + | |
1888 | + self.train = theano.function( inputs = [words, children_positions, y, words_indexes, learning_rate], | |
1889 | + outputs = [], | |
1890 | + updates = updates, | |
1891 | + allow_input_downcast=True, | |
1892 | + mode='FAST_RUN' | |
1893 | + ) | |
1894 | + | |
1895 | + | |
1896 | + def one_step_classify(word_id, word_children_positions, i, hidden_states): | |
1897 | + | |
1898 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
1899 | + | |
1900 | + tmp = word_children_positions>=0.0 | |
1901 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
1902 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
1903 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
1904 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
1905 | + schh = schh/number_of_children | |
1906 | + | |
1907 | + eh = T.tanh(T.dot(self.emb[word_id],self.W_e_eh)) | |
1908 | + eh2 = T.tanh(T.dot(eh,self.W_eh_eh2)) | |
1909 | + | |
1910 | + shh = T.tanh(T.dot(schh,self.W_sh_shh)) | |
1911 | + shh2 = T.tanh(T.dot(shh,self.W_shh_shh2)) | |
1912 | + | |
1913 | + | |
1914 | + h = T.tanh(T.concatenate([T.dot(eh2, self.W_eh2_h), T.dot(shh2,self.W_shh2_h)])) | |
1915 | + | |
1916 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
1917 | + | |
1918 | + current_hidden_state = hidden_states[i] | |
1919 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
1920 | + | |
1921 | + y_prob = T.nnet.softmax(T.dot(h2,self.W_h2_y) + self.b_y)[0] | |
1922 | + | |
1923 | + | |
1924 | + return y_prob, hidden_states_new | |
1925 | + | |
1926 | + | |
1927 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
1928 | + fn=one_step_classify, | |
1929 | + sequences = [words, children_positions,words_indexes], | |
1930 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))]) | |
1931 | + | |
1932 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
1933 | + sequences = [words_indexes]) | |
1934 | + | |
1935 | + self.classify = theano.function(inputs=[words,children_positions,words_indexes], | |
1936 | + outputs=predictions, | |
1937 | + allow_input_downcast=True, | |
1938 | + mode='FAST_RUN' | |
1939 | + ) | |
1940 | + | |
1941 | + | |
1942 | + | |
1943 | + | |
1944 | +class model55_pf15(object): | |
1945 | + def __init__(self, neh, neh2, ne, nshh, nshh2, nchd, nh2, nh3, nc, w2v_model_path, max_phrase_length): | |
1946 | + ''' | |
1947 | + nh :: dimension of hidden state | |
1948 | + nc :: number of classes | |
1949 | + ne :: number of word embeddings in the vocabulary | |
1950 | + de :: dimension of the word embeddings | |
1951 | + ds :: dimension of the sentiment state | |
1952 | + ''' | |
1953 | + self.max_phrase_length = max_phrase_length | |
1954 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
1955 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
1956 | + self.words2ids = w2vecs["words2ids"] | |
1957 | + | |
1958 | + #ne = len(w2vecs["words2ids"]) | |
1959 | + de = w2vecs["vectors"].shape[1] | |
1960 | + del w2vecs | |
1961 | + | |
1962 | + r = 0.05 | |
1963 | + self.W_e_eh = theano.shared(r * np.random.uniform(-1.0, 1.0, (de, neh)).astype(theano.config.floatX)) | |
1964 | + self.W_eh_eh2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (neh, neh2)).astype(theano.config.floatX)) | |
1965 | + self.W_eh2_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (neh2, ne)).astype(theano.config.floatX)) | |
1966 | + | |
1967 | + self.W_sh_shh = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nshh)).astype(theano.config.floatX)) | |
1968 | + self.W_shh_shh2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (nshh, nshh2)).astype(theano.config.floatX)) | |
1969 | + self.W_shh2_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (nshh2, nchd)).astype(theano.config.floatX)) | |
1970 | + | |
1971 | + self.W_h_h2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nh2)).astype(theano.config.floatX)) | |
1972 | + self.W_h2_h3 = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh2, nh3)).astype(theano.config.floatX)) | |
1973 | + self.W_h3_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh3, nc)).astype(theano.config.floatX)) | |
1974 | + | |
1975 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
1976 | + | |
1977 | + | |
1978 | + def one_step(word_id, word_children_positions, y_true, i, hidden_states, learning_rate): | |
1979 | + | |
1980 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
1981 | + | |
1982 | + tmp = word_children_positions>=0.0 | |
1983 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
1984 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
1985 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
1986 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
1987 | + schh = schh/number_of_children | |
1988 | + | |
1989 | + eh = T.tanh(T.dot(self.emb[word_id],self.W_e_eh)) | |
1990 | + eh2 = T.tanh(T.dot(eh,self.W_eh_eh2)) | |
1991 | + | |
1992 | + shh = T.tanh(T.dot(schh,self.W_sh_shh)) | |
1993 | + shh2 = T.tanh(T.dot(shh,self.W_shh_shh2)) | |
1994 | + | |
1995 | + | |
1996 | + h = T.tanh(T.concatenate([T.dot(eh2, self.W_eh2_h), T.dot(shh2,self.W_shh2_h)])) | |
1997 | + | |
1998 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
1999 | + h3 = T.tanh(T.dot(h2, self.W_h2_h3)) | |
2000 | + | |
2001 | + current_hidden_state = hidden_states[i] | |
2002 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
2003 | + | |
2004 | + y_prob = T.nnet.softmax(T.dot(h3,self.W_h3_y) + self.b_y)[0] | |
2005 | + | |
2006 | + #l2_norm = T.sum(self.W_h_h2**2) + T.sum(self.W_h2_y**2) + T.sum(self.W_e_h**2) + T.sum(self.W_sh_h**2) + T.sum(self.emb**2) + T.sum(self.b_h**2) + T.sum(self.b_y**2) | |
2007 | + | |
2008 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
2009 | + | |
2010 | + return cross_entropy, hidden_states_new | |
2011 | + | |
2012 | + | |
2013 | + y = T.vector('y',dtype=dataType) | |
2014 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
2015 | + words = T.vector(dtype=dataType) | |
2016 | + children_positions = T.matrix(dtype=dataType) | |
2017 | + words_indexes = T.vector(dtype=dataType) | |
2018 | + | |
2019 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
2020 | + sequences = [words, children_positions,y,words_indexes], | |
2021 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))], | |
2022 | + non_sequences = learning_rate, | |
2023 | + n_steps = words.shape[0]) | |
2024 | + | |
2025 | + cost = T.sum(cross_entropy_vector[0]) | |
2026 | + | |
2027 | + updates = OrderedDict([ | |
2028 | + (self.W_h_h2, self.W_h_h2-learning_rate*T.grad(cost, self.W_h_h2)), | |
2029 | + (self.W_h2_h3, self.W_h2_h3-learning_rate*T.grad(cost, self.W_h2_h3)), | |
2030 | + (self.W_h3_y, self.W_h3_y-learning_rate*T.grad(cost, self.W_h3_y)), | |
2031 | + (self.W_e_eh, self.W_e_eh-learning_rate*T.grad(cost, self.W_e_eh)), | |
2032 | + (self.W_eh_eh2, self.W_eh_eh2-learning_rate*T.grad(cost, self.W_eh_eh2)), | |
2033 | + (self.W_eh2_h, self.W_eh2_h-learning_rate*T.grad(cost, self.W_eh2_h)), | |
2034 | + (self.W_shh2_h, self.W_shh2_h-learning_rate*T.grad(cost, self.W_shh2_h)), | |
2035 | + (self.W_sh_shh, self.W_sh_shh-learning_rate*T.grad(cost, self.W_sh_shh)), | |
2036 | + (self.W_shh_shh2, self.W_shh_shh2-learning_rate*T.grad(cost, self.W_shh_shh2)), | |
2037 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
2038 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
2039 | + ]) | |
2040 | + | |
2041 | + self.train = theano.function( inputs = [words, children_positions, y, words_indexes, learning_rate], | |
2042 | + outputs = [], | |
2043 | + updates = updates, | |
2044 | + allow_input_downcast=True, | |
2045 | + mode='FAST_RUN' | |
2046 | + ) | |
2047 | + | |
2048 | + | |
2049 | + def one_step_classify(word_id, word_children_positions, i, hidden_states): | |
2050 | + | |
2051 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
2052 | + | |
2053 | + tmp = word_children_positions>=0.0 | |
2054 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
2055 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
2056 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
2057 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
2058 | + schh = schh/number_of_children | |
2059 | + | |
2060 | + eh = T.tanh(T.dot(self.emb[word_id],self.W_e_eh)) | |
2061 | + eh2 = T.tanh(T.dot(eh,self.W_eh_eh2)) | |
2062 | + | |
2063 | + shh = T.tanh(T.dot(schh,self.W_sh_shh)) | |
2064 | + shh2 = T.tanh(T.dot(shh,self.W_shh_shh2)) | |
2065 | + | |
2066 | + | |
2067 | + h = T.tanh(T.concatenate([T.dot(eh2, self.W_eh2_h), T.dot(shh2,self.W_shh2_h)])) | |
2068 | + | |
2069 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
2070 | + h3 = T.tanh(T.dot(h2, self.W_h2_h3)) | |
2071 | + | |
2072 | + current_hidden_state = hidden_states[i] | |
2073 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
2074 | + | |
2075 | + y_prob = T.nnet.softmax(T.dot(h3,self.W_h3_y) + self.b_y)[0] | |
2076 | + | |
2077 | + | |
2078 | + return y_prob, hidden_states_new | |
2079 | + | |
2080 | + | |
2081 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
2082 | + fn=one_step_classify, | |
2083 | + sequences = [words, children_positions,words_indexes], | |
2084 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))]) | |
2085 | + | |
2086 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
2087 | + sequences = [words_indexes]) | |
2088 | + | |
2089 | + self.classify = theano.function(inputs=[words,children_positions,words_indexes], | |
2090 | + outputs=predictions, | |
2091 | + allow_input_downcast=True, | |
2092 | + mode='FAST_RUN' | |
2093 | + ) | |
2094 | + | |
2095 | + | |
2096 | + | |
2097 | + | |
2098 | +class model55_pf16(object): | |
2099 | + def __init__(self, nchd, nc, w2v_model_path, max_phrase_length): | |
2100 | + ''' | |
2101 | + nh :: dimension of hidden state | |
2102 | + nc :: number of classes | |
2103 | + ne :: number of word embeddings in the vocabulary | |
2104 | + de :: dimension of the word embeddings | |
2105 | + ds :: dimension of the sentiment state | |
2106 | + ''' | |
2107 | + self.max_phrase_length = max_phrase_length | |
2108 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
2109 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
2110 | + self.words2ids = w2vecs["words2ids"] | |
2111 | + | |
2112 | + #ne = len(w2vecs["words2ids"]) | |
2113 | + de = w2vecs["vectors"].shape[1] | |
2114 | + del w2vecs | |
2115 | + | |
2116 | + r = 0.05 | |
2117 | + | |
2118 | + self.W_h_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (de+nchd, nc)).astype(theano.config.floatX)) | |
2119 | + | |
2120 | + self.W_sh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (de+nchd, nchd)).astype(theano.config.floatX)) | |
2121 | + | |
2122 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
2123 | + | |
2124 | + | |
2125 | + def one_step(word_id, word_children_positions, y_true, i, hidden_states, learning_rate): | |
2126 | + | |
2127 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
2128 | + | |
2129 | + tmp = word_children_positions>=0.0 | |
2130 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
2131 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
2132 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
2133 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
2134 | + schh = schh/number_of_children | |
2135 | + | |
2136 | + | |
2137 | + h = T.tanh(T.concatenate([self.emb[word_id], T.dot(schh,self.W_sh_h)])) | |
2138 | + | |
2139 | + | |
2140 | + | |
2141 | + current_hidden_state = hidden_states[i] | |
2142 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
2143 | + | |
2144 | + y_prob = T.nnet.softmax(T.dot(h,self.W_h_y) + self.b_y)[0] | |
2145 | + | |
2146 | + #l2_norm = T.sum(self.W_h_h2**2) + T.sum(self.W_h2_y**2) + T.sum(self.W_e_h**2) + T.sum(self.W_sh_h**2) + T.sum(self.emb**2) + T.sum(self.b_h**2) + T.sum(self.b_y**2) | |
2147 | + | |
2148 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
2149 | + | |
2150 | + return cross_entropy, hidden_states_new | |
2151 | + | |
2152 | + | |
2153 | + y = T.vector('y',dtype=dataType) | |
2154 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
2155 | + words = T.vector(dtype=dataType) | |
2156 | + children_positions = T.matrix(dtype=dataType) | |
2157 | + words_indexes = T.vector(dtype=dataType) | |
2158 | + | |
2159 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
2160 | + sequences = [words, children_positions,y,words_indexes], | |
2161 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,de+nchd), dtype = theano.config.floatX))], | |
2162 | + non_sequences = learning_rate, | |
2163 | + n_steps = words.shape[0]) | |
2164 | + | |
2165 | + cost = T.sum(cross_entropy_vector[0]) | |
2166 | + | |
2167 | + updates = OrderedDict([ | |
2168 | + (self.W_h_y, self.W_h_y-learning_rate*T.grad(cost, self.W_h_y)), | |
2169 | + (self.W_sh_h, self.W_sh_h-learning_rate*T.grad(cost, self.W_sh_h)), | |
2170 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
2171 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
2172 | + ]) | |
2173 | + | |
2174 | + self.train = theano.function( inputs = [words, children_positions, y, words_indexes, learning_rate], | |
2175 | + outputs = [], | |
2176 | + updates = updates, | |
2177 | + allow_input_downcast=True, | |
2178 | + mode='FAST_RUN' | |
2179 | + ) | |
2180 | + | |
2181 | + | |
2182 | + def one_step_classify(word_id, word_children_positions, i, hidden_states): | |
2183 | + | |
2184 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
2185 | + | |
2186 | + tmp = word_children_positions>=0.0 | |
2187 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
2188 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
2189 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
2190 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
2191 | + schh = schh/number_of_children | |
2192 | + | |
2193 | + h = T.tanh(T.concatenate([self.emb[word_id], T.dot(schh,self.W_sh_h)])) | |
2194 | + | |
2195 | + current_hidden_state = hidden_states[i] | |
2196 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
2197 | + | |
2198 | + y_prob = T.nnet.softmax(T.dot(h,self.W_h_y) + self.b_y)[0] | |
2199 | + | |
2200 | + return y_prob, hidden_states_new | |
2201 | + | |
2202 | + | |
2203 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
2204 | + fn=one_step_classify, | |
2205 | + sequences = [words, children_positions,words_indexes], | |
2206 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,de+nchd), dtype = theano.config.floatX))]) | |
2207 | + | |
2208 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
2209 | + sequences = [words_indexes]) | |
2210 | + | |
2211 | + self.classify = theano.function(inputs=[words,children_positions,words_indexes], | |
2212 | + outputs=predictions, | |
2213 | + allow_input_downcast=True, | |
2214 | + mode='FAST_RUN' | |
2215 | + ) | |
2216 | + | |
2217 | + | |
2218 | + | |
2219 | +class model55_pf17(object): | |
2220 | + def __init__(self, neh, neh2, ne, nshh, nshh2, nchd, nh2, nh3, nh4, nc, w2v_model_path, max_phrase_length): | |
2221 | + ''' | |
2222 | + nh :: dimension of hidden state | |
2223 | + nc :: number of classes | |
2224 | + ne :: number of word embeddings in the vocabulary | |
2225 | + de :: dimension of the word embeddings | |
2226 | + ds :: dimension of the sentiment state | |
2227 | + ''' | |
2228 | + self.max_phrase_length = max_phrase_length | |
2229 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
2230 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
2231 | + self.words2ids = w2vecs["words2ids"] | |
2232 | + | |
2233 | + #ne = len(w2vecs["words2ids"]) | |
2234 | + de = w2vecs["vectors"].shape[1] | |
2235 | + del w2vecs | |
2236 | + | |
2237 | + r = 0.05 | |
2238 | + self.W_e_eh = theano.shared(r * np.random.uniform(-1.0, 1.0, (de, neh)).astype(theano.config.floatX)) | |
2239 | + self.W_eh_eh2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (neh, neh2)).astype(theano.config.floatX)) | |
2240 | + self.W_eh2_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (neh2, ne)).astype(theano.config.floatX)) | |
2241 | + | |
2242 | + self.W_sh_shh = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nshh)).astype(theano.config.floatX)) | |
2243 | + self.W_shh_shh2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (nshh, nshh2)).astype(theano.config.floatX)) | |
2244 | + self.W_shh2_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (nshh2, nchd)).astype(theano.config.floatX)) | |
2245 | + | |
2246 | + self.W_h_h2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nh2)).astype(theano.config.floatX)) | |
2247 | + self.W_h2_h3 = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh2, nh3)).astype(theano.config.floatX)) | |
2248 | + self.W_h3_h4 = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh3, nh4)).astype(theano.config.floatX)) | |
2249 | + self.W_h4_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh4, nc)).astype(theano.config.floatX)) | |
2250 | + | |
2251 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
2252 | + | |
2253 | + | |
2254 | + def one_step(word_id, word_children_positions, y_true, i, hidden_states, learning_rate): | |
2255 | + | |
2256 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
2257 | + | |
2258 | + tmp = word_children_positions>=0.0 | |
2259 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
2260 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
2261 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
2262 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
2263 | + schh = schh/number_of_children | |
2264 | + | |
2265 | + eh = T.tanh(T.dot(self.emb[word_id],self.W_e_eh)) | |
2266 | + eh2 = T.tanh(T.dot(eh,self.W_eh_eh2)) | |
2267 | + | |
2268 | + shh = T.tanh(T.dot(schh,self.W_sh_shh)) | |
2269 | + shh2 = T.tanh(T.dot(shh,self.W_shh_shh2)) | |
2270 | + | |
2271 | + | |
2272 | + h = T.tanh(T.concatenate([T.dot(eh2, self.W_eh2_h), T.dot(shh2,self.W_shh2_h)])) | |
2273 | + | |
2274 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
2275 | + h3 = T.tanh(T.dot(h2, self.W_h2_h3)) | |
2276 | + h4 = T.tanh(T.dot(h3, self.W_h3_h4)) | |
2277 | + | |
2278 | + current_hidden_state = hidden_states[i] | |
2279 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
2280 | + | |
2281 | + y_prob = T.nnet.softmax(T.dot(h4,self.W_h4_y) + self.b_y)[0] | |
2282 | + | |
2283 | + #l2_norm = T.sum(self.W_h_h2**2) + T.sum(self.W_h2_y**2) + T.sum(self.W_e_h**2) + T.sum(self.W_sh_h**2) + T.sum(self.emb**2) + T.sum(self.b_h**2) + T.sum(self.b_y**2) | |
2284 | + | |
2285 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
2286 | + | |
2287 | + return cross_entropy, hidden_states_new | |
2288 | + | |
2289 | + | |
2290 | + y = T.vector('y',dtype=dataType) | |
2291 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
2292 | + words = T.vector(dtype=dataType) | |
2293 | + children_positions = T.matrix(dtype=dataType) | |
2294 | + words_indexes = T.vector(dtype=dataType) | |
2295 | + | |
2296 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
2297 | + sequences = [words, children_positions,y,words_indexes], | |
2298 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))], | |
2299 | + non_sequences = learning_rate, | |
2300 | + n_steps = words.shape[0]) | |
2301 | + | |
2302 | + cost = T.sum(cross_entropy_vector[0]) | |
2303 | + | |
2304 | + updates = OrderedDict([ | |
2305 | + (self.W_h_h2, self.W_h_h2-learning_rate*T.grad(cost, self.W_h_h2)), | |
2306 | + (self.W_h2_h3, self.W_h2_h3-learning_rate*T.grad(cost, self.W_h2_h3)), | |
2307 | + (self.W_h3_h4, self.W_h3_h4-learning_rate*T.grad(cost, self.W_h3_h4)), | |
2308 | + (self.W_h4_y, self.W_h4_y-learning_rate*T.grad(cost, self.W_h4_y)), | |
2309 | + (self.W_e_eh, self.W_e_eh-learning_rate*T.grad(cost, self.W_e_eh)), | |
2310 | + (self.W_eh_eh2, self.W_eh_eh2-learning_rate*T.grad(cost, self.W_eh_eh2)), | |
2311 | + (self.W_eh2_h, self.W_eh2_h-learning_rate*T.grad(cost, self.W_eh2_h)), | |
2312 | + (self.W_shh2_h, self.W_shh2_h-learning_rate*T.grad(cost, self.W_shh2_h)), | |
2313 | + (self.W_sh_shh, self.W_sh_shh-learning_rate*T.grad(cost, self.W_sh_shh)), | |
2314 | + (self.W_shh_shh2, self.W_shh_shh2-learning_rate*T.grad(cost, self.W_shh_shh2)), | |
2315 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
2316 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
2317 | + ]) | |
2318 | + | |
2319 | + self.train = theano.function( inputs = [words, children_positions, y, words_indexes, learning_rate], | |
2320 | + outputs = [], | |
2321 | + updates = updates, | |
2322 | + allow_input_downcast=True, | |
2323 | + mode='FAST_RUN' | |
2324 | + ) | |
2325 | + | |
2326 | + | |
2327 | + def one_step_classify(word_id, word_children_positions, i, hidden_states): | |
2328 | + | |
2329 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
2330 | + | |
2331 | + tmp = word_children_positions>=0.0 | |
2332 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
2333 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
2334 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
2335 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
2336 | + schh = schh/number_of_children | |
2337 | + | |
2338 | + eh = T.tanh(T.dot(self.emb[word_id],self.W_e_eh)) | |
2339 | + eh2 = T.tanh(T.dot(eh,self.W_eh_eh2)) | |
2340 | + | |
2341 | + shh = T.tanh(T.dot(schh,self.W_sh_shh)) | |
2342 | + shh2 = T.tanh(T.dot(shh,self.W_shh_shh2)) | |
2343 | + | |
2344 | + | |
2345 | + h = T.tanh(T.concatenate([T.dot(eh2, self.W_eh2_h), T.dot(shh2,self.W_shh2_h)])) | |
2346 | + | |
2347 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
2348 | + h3 = T.tanh(T.dot(h2, self.W_h2_h3)) | |
2349 | + h4 = T.tanh(T.dot(h3, self.W_h3_h4)) | |
2350 | + | |
2351 | + current_hidden_state = hidden_states[i] | |
2352 | + hidden_states_new = T.set_subtensor(current_hidden_state, h) | |
2353 | + | |
2354 | + y_prob = T.nnet.softmax(T.dot(h4,self.W_h4_y) + self.b_y)[0] | |
2355 | + | |
2356 | + | |
2357 | + return y_prob, hidden_states_new | |
2358 | + | |
2359 | + | |
2360 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
2361 | + fn=one_step_classify, | |
2362 | + sequences = [words, children_positions,words_indexes], | |
2363 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))]) | |
2364 | + | |
2365 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
2366 | + sequences = [words_indexes]) | |
2367 | + | |
2368 | + self.classify = theano.function(inputs=[words,children_positions,words_indexes], | |
2369 | + outputs=predictions, | |
2370 | + allow_input_downcast=True, | |
2371 | + mode='FAST_RUN' | |
2372 | + ) | |
2373 | + | |
2374 | + | |
... | ... |
modules/rnn/models.pyc
0 → 100644
No preview for this file type
modules/rnn/models_with_relations.py
0 → 100644
1 | + | |
2 | +import numpy as np | |
3 | +import time | |
4 | +import sys | |
5 | +import subprocess | |
6 | +import os | |
7 | +import random | |
8 | + | |
9 | +#from modules.data import load | |
10 | +#from modules.rnn.many_models import * | |
11 | +#from modules.metrics.accuracy import conlleval | |
12 | +from modules.utils.tools import load_stanford_data4 | |
13 | + | |
14 | +from theano import pp | |
15 | + | |
16 | +import theano.tensor as T | |
17 | +import theano | |
18 | +from theano.sandbox.rng_mrg import MRG_RandomStreams #as MRG_RandomStreams | |
19 | + | |
20 | +import itertools | |
21 | + | |
22 | +import os.path | |
23 | +import pickle | |
24 | + | |
25 | +from collections import Counter | |
26 | + | |
27 | + | |
28 | + | |
29 | +from theano import tensor as T, printing | |
30 | +from collections import OrderedDict | |
31 | +from theano.ifelse import ifelse | |
32 | + | |
33 | +from keras.preprocessing import sequence as seq | |
34 | + | |
35 | +dataType = 'int64' | |
36 | + | |
37 | + | |
38 | + | |
39 | + | |
40 | +class MLP_2_1(object): | |
41 | + | |
42 | + # punkt wyjscia modelu: model55_pf1 | |
43 | + | |
44 | + def __init__(self, ne, nchd, nc, w2v_model_path, max_phrase_length, number_of_relations): | |
45 | + ''' | |
46 | + nh :: dimension of hidden state | |
47 | + nc :: number of classes | |
48 | + de :: dimension of the word embeddings | |
49 | + ds :: dimension of the sentiment state | |
50 | + ''' | |
51 | + self.max_phrase_length = max_phrase_length | |
52 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
53 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
54 | + self.words2ids = w2vecs["words2ids"] | |
55 | + | |
56 | + #ne = len(w2vecs["words2ids"]) | |
57 | + de = w2vecs["vectors"].shape[1] | |
58 | + del w2vecs | |
59 | + | |
60 | + r = 0.05 | |
61 | + self.W_e_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (de, ne)).astype(theano.config.floatX)) | |
62 | + self.W_h_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nc)).astype(theano.config.floatX)) | |
63 | + | |
64 | + self.W_sh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nchd)).astype(theano.config.floatX)) | |
65 | + | |
66 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
67 | + | |
68 | + self.relations_weights = theano.shared(r * np.random.uniform(-1.0, 1.0, (number_of_relations+1, ne+nchd, ne+nchd)).astype(theano.config.floatX)) | |
69 | + | |
70 | + def one_step(word_id, word_children_positions, y_true, relation, i, hidden_states, learning_rate): | |
71 | + | |
72 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
73 | + | |
74 | + tmp = word_children_positions>=0.0 | |
75 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
76 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
77 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
78 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
79 | + schh = schh/number_of_children | |
80 | + | |
81 | + h = T.tanh(T.concatenate([T.dot(self.emb[word_id],self.W_e_h), T.dot(schh,self.W_sh_h)])) | |
82 | + | |
83 | + current_hidden_state = hidden_states[i] | |
84 | + hidden_states_new = T.set_subtensor(current_hidden_state, T.dot(h, self.relations_weights[relation,:,:])) | |
85 | + | |
86 | + y_prob = T.nnet.softmax(T.dot(h,self.W_h_y) + self.b_y)[0] | |
87 | + | |
88 | + #l2_norm = T.sum(self.W_h_h2**2) + T.sum(self.W_h2_y**2) + T.sum(self.W_e_h**2) + T.sum(self.W_sh_h**2) + T.sum(self.emb**2) + T.sum(self.b_h**2) + T.sum(self.b_y**2) | |
89 | + | |
90 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
91 | + | |
92 | + return cross_entropy, hidden_states_new | |
93 | + | |
94 | + | |
95 | + y = T.vector('y',dtype=dataType) | |
96 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
97 | + words = T.vector(dtype=dataType) | |
98 | + children_positions = T.matrix(dtype=dataType) | |
99 | + | |
100 | + relations = T.vector(dtype=dataType) | |
101 | + | |
102 | + words_indexes = T.vector(dtype=dataType) | |
103 | + | |
104 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
105 | + sequences = [words, children_positions,y, relations, words_indexes], | |
106 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))], | |
107 | + non_sequences = learning_rate, | |
108 | + n_steps = words.shape[0]) | |
109 | + cost = T.sum(cross_entropy_vector[0]) | |
110 | + | |
111 | + updates = OrderedDict([ | |
112 | + (self.W_h_y, self.W_h_y-learning_rate*T.grad(cost, self.W_h_y)), | |
113 | + (self.W_e_h, self.W_e_h-learning_rate*T.grad(cost, self.W_e_h)), | |
114 | + (self.W_sh_h, self.W_sh_h-learning_rate*T.grad(cost, self.W_sh_h)), | |
115 | + (self.relations_weights, self.relations_weights - learning_rate*T.grad(cost,self.relations_weights)), | |
116 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
117 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
118 | + ]) | |
119 | + | |
120 | + self.train = theano.function( inputs = [words, children_positions, y, relations, words_indexes, learning_rate], | |
121 | + outputs = [], | |
122 | + updates = updates, | |
123 | + allow_input_downcast=True, | |
124 | + mode='FAST_RUN' | |
125 | + ) | |
126 | + | |
127 | + | |
128 | + def one_step_classify(word_id, word_children_positions, relation, i, hidden_states): | |
129 | + | |
130 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
131 | + | |
132 | + tmp = word_children_positions>=0.0 | |
133 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
134 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
135 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
136 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
137 | + schh = schh/number_of_children | |
138 | + | |
139 | + h = T.tanh(T.concatenate([T.dot(self.emb[word_id],self.W_e_h), T.dot(schh,self.W_sh_h)]) ) | |
140 | + | |
141 | + current_hidden_state = hidden_states[i] | |
142 | + hidden_states_new = T.set_subtensor(current_hidden_state, T.dot(h, self.relations_weights[relation,:,:])) | |
143 | + | |
144 | + y_prob = T.nnet.softmax(T.dot(h,self.W_h_y) + self.b_y)[0] | |
145 | + | |
146 | + return y_prob, hidden_states_new | |
147 | + | |
148 | + | |
149 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
150 | + fn=one_step_classify, | |
151 | + sequences = [words, children_positions, relations, words_indexes], | |
152 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))]) | |
153 | + | |
154 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
155 | + sequences = [words_indexes]) | |
156 | + | |
157 | + self.classify = theano.function(inputs=[words,children_positions,relations, words_indexes], | |
158 | + outputs=predictions, | |
159 | + allow_input_downcast=True, | |
160 | + mode='FAST_RUN' | |
161 | + ) | |
162 | + | |
163 | + | |
164 | + | |
165 | +class MLP_2_2(object): | |
166 | + | |
167 | + # punkt wyjscia modelu: model55_pf2 | |
168 | + | |
169 | + def __init__(self, ne, nchd, nh2, nc, w2v_model_path, max_phrase_length, number_of_relations): | |
170 | + ''' | |
171 | + nh :: dimension of hidden state | |
172 | + nc :: number of classes | |
173 | + de :: dimension of the word embeddings | |
174 | + ds :: dimension of the sentiment state | |
175 | + ''' | |
176 | + self.max_phrase_length = max_phrase_length | |
177 | + w2vecs = pickle.load(open(w2v_model_path,"r")) | |
178 | + self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
179 | + self.words2ids = w2vecs["words2ids"] | |
180 | + | |
181 | + de = w2vecs["vectors"].shape[1] | |
182 | + del w2vecs | |
183 | + | |
184 | + r = 0.05 | |
185 | + | |
186 | + self.W_e_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (de, ne)).astype(theano.config.floatX)) | |
187 | + | |
188 | + self.W_sh_h = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nchd)).astype(theano.config.floatX)) | |
189 | + | |
190 | + self.W_h_h2 = theano.shared(r * np.random.uniform(-1.0, 1.0, (ne+nchd, nh2)).astype(theano.config.floatX)) | |
191 | + self.W_h2_y = theano.shared(r * np.random.uniform(-1.0, 1.0, (nh2, nc)).astype(theano.config.floatX)) | |
192 | + | |
193 | + self.b_y = theano.shared(r * np.random.uniform(-1.0, 1.0, nc).astype(theano.config.floatX)) | |
194 | + | |
195 | + self.relations_weights = theano.shared(r * np.random.uniform(-1.0, 1.0, (number_of_relations+1, ne+nchd, ne+nchd)).astype(theano.config.floatX)) | |
196 | + | |
197 | + def one_step(word_id, word_children_positions, y_true, relation, i, hidden_states, learning_rate): | |
198 | + | |
199 | + schh = hidden_states[-1] #+ 0.5 # czyli wektor zerowy | |
200 | + | |
201 | + tmp = word_children_positions>=0.0 | |
202 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
203 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
204 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
205 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
206 | + schh = schh/number_of_children | |
207 | + | |
208 | + h = T.tanh(T.concatenate([T.dot(self.emb[word_id],self.W_e_h), T.dot(schh,self.W_sh_h)])) | |
209 | + #h = T.nnet.sigmoid(T.dot(self.emb[word_id],self.W_eh) + T.dot(schh,self.W_shsh) + self.bh) | |
210 | + | |
211 | + current_hidden_state = hidden_states[i] | |
212 | + hidden_states_new = T.set_subtensor(current_hidden_state, T.dot(h, self.relations_weights[relation,:,:])) | |
213 | + | |
214 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
215 | + | |
216 | + y_prob = T.nnet.softmax(T.dot(h2,self.W_h2_y) + self.b_y)[0] | |
217 | + | |
218 | + cross_entropy = -T.log(y_prob[y_true]) # + norm_coefficient * l2_norm | |
219 | + | |
220 | + return cross_entropy, hidden_states_new | |
221 | + | |
222 | + | |
223 | + y = T.vector('y',dtype=dataType) | |
224 | + learning_rate = T.scalar('lr',dtype=theano.config.floatX) | |
225 | + words = T.vector(dtype=dataType) | |
226 | + children_positions = T.matrix(dtype=dataType) | |
227 | + relations = T.vector(dtype=dataType) | |
228 | + words_indexes = T.vector(dtype=dataType) | |
229 | + | |
230 | + cross_entropy_vector, _ = theano.scan(fn=one_step, \ | |
231 | + sequences = [words, children_positions,y,relations,words_indexes], | |
232 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))], | |
233 | + non_sequences = learning_rate, | |
234 | + n_steps = words.shape[0]) | |
235 | + cost = T.sum(cross_entropy_vector[0]) | |
236 | + | |
237 | + updates = OrderedDict([ | |
238 | + (self.W_h_h2, self.W_h_h2-learning_rate*T.grad(cost, self.W_h_h2)), | |
239 | + (self.W_h2_y, self.W_h2_y-learning_rate*T.grad(cost, self.W_h2_y)), | |
240 | + (self.W_e_h, self.W_e_h-learning_rate*T.grad(cost, self.W_e_h)), | |
241 | + (self.W_sh_h, self.W_sh_h-learning_rate*T.grad(cost, self.W_sh_h)), | |
242 | + (self.relations_weights, self.relations_weights - learning_rate*T.grad(cost,self.relations_weights)), | |
243 | + (self.emb, self.emb-learning_rate*T.grad(cost, self.emb)), #updated_current_emb), # | |
244 | + (self.b_y, self.b_y-learning_rate*T.grad(cost,self.b_y)) | |
245 | + ]) | |
246 | + | |
247 | + self.train = theano.function( inputs = [words, children_positions, y, relations, words_indexes, learning_rate], | |
248 | + outputs = [], | |
249 | + updates = updates, | |
250 | + allow_input_downcast=True, | |
251 | + mode='FAST_RUN' | |
252 | + ) | |
253 | + | |
254 | + | |
255 | + def one_step_classify(word_id, word_children_positions, relation, i, hidden_states): | |
256 | + | |
257 | + schh = hidden_states[-1] #+ 0.5# czyli wektor zerowy | |
258 | + | |
259 | + tmp = word_children_positions>=0.0 | |
260 | + idx_tmp = tmp.nonzero() #indeksy realne dzieci - czyli te, gdzie nie ma -1 | |
261 | + schh = schh + hidden_states[word_children_positions[idx_tmp]].sum(axis=0) #suma stanow ukrytych dzieci | |
262 | + number_of_children = tmp.sum(dtype = theano.config.floatX) | |
263 | + number_of_children = ifelse( T.gt(number_of_children, 1.0),number_of_children, 1.0) | |
264 | + schh = schh/number_of_children | |
265 | + | |
266 | + h = T.tanh(T.concatenate([T.dot(self.emb[word_id],self.W_e_h), T.dot(schh,self.W_sh_h)])) | |
267 | + | |
268 | + current_hidden_state = hidden_states[i] | |
269 | + hidden_states_new = T.set_subtensor(current_hidden_state, T.dot(h, self.relations_weights[relation,:,:])) | |
270 | + | |
271 | + h2 = T.tanh(T.dot(h, self.W_h_h2)) | |
272 | + | |
273 | + y_prob = T.nnet.softmax(T.dot(h2,self.W_h2_y) + self.b_y)[0] | |
274 | + | |
275 | + return y_prob, hidden_states_new | |
276 | + | |
277 | + | |
278 | + [y_probs_classify, hidden_states ], _ = theano.scan( | |
279 | + fn=one_step_classify, | |
280 | + sequences = [words, children_positions, relations, words_indexes], | |
281 | + outputs_info = [None, theano.shared(np.zeros((self.max_phrase_length+1,ne+nchd), dtype = theano.config.floatX))]) | |
282 | + | |
283 | + predictions, _ = theano.scan(lambda i: T.argmax(y_probs_classify[i]), | |
284 | + sequences = [words_indexes]) | |
285 | + | |
286 | + self.classify = theano.function(inputs=[words,children_positions,relations,words_indexes], | |
287 | + outputs=predictions, | |
288 | + allow_input_downcast=True, | |
289 | + mode='FAST_RUN' | |
290 | + ) | |
291 | + | |
292 | + | |
293 | + | |
294 | + | |
... | ... |
modules/rnn/models_with_relations.pyc
0 → 100644
No preview for this file type
modules/rnn/tmp.py deleted
1 | -import theano | |
2 | -import numpy as np | |
3 | -import os | |
4 | -import pickle | |
5 | - | |
6 | -from theano import tensor as T, printing | |
7 | -from collections import OrderedDict | |
8 | -from theano.ifelse import ifelse | |
9 | - | |
10 | -theano.config.floatX = 'float64' | |
11 | -dataType = 'int64' | |
12 | - | |
13 | -class model(object): | |
14 | - | |
15 | - def __init__(self, nh, nc, ds, w2v_model_path, max_phrase_length): | |
16 | - ''' | |
17 | - nh :: dimension of hidden state | |
18 | - nc :: number of classes | |
19 | - ne :: number of word embeddings in the vocabulary | |
20 | - de :: dimension of the word embeddings | |
21 | - | |
22 | - ds :: dimension of the sentiment state | |
23 | - ''' | |
24 | - | |
25 | - | |
26 | - | |
27 | - self.max_phrase_length = max_phrase_length | |
28 | - | |
29 | - ###ne = len(model.index2word) | |
30 | - ###de = model.vector_size | |
31 | - | |
32 | - ###vectors = np.zeros((ne,de)) | |
33 | - ###self.words2ids = {} | |
34 | - ###for i in range(len(model.index2word)): | |
35 | - ### self.words2ids[model.index2word[i]] = i | |
36 | - ### vectors[i] = model[model.index2word[i]] | |
37 | - | |
38 | - w2vecs = pickle.load(open(w2v_model_path,"r")) | |
39 | - #self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
40 | - self.emb = theano.shared(np.load("saved_models4/embeddings.npy").astype(theano.config.floatX)) | |
41 | - self.words2ids = w2vecs["words2ids"] | |
42 | - | |
43 | - ne = len(w2vecs["words2ids"]) | |
44 | - de = w2vecs["vectors"].shape[1] | |
45 | - | |
46 | - del w2vecs | |
47 | - | |
48 | - #self.words2ids = {} | |
49 | - #vectors = [] | |
50 | - #i = 0 | |
51 | - #for line in open(w2v_model_path,"r"): | |
52 | - # toks = line.strip("\n").split(" ") | |
53 | - # word = toks[0] | |
54 | - # v = map(float, toks[1:]) | |
55 | - # vectors.append(v) | |
56 | - # self.words2ids[word] = i | |
57 | - # i = i + 1 | |
58 | - #vectors.append(np.zeros((len(vectors[0])))) | |
59 | - #vectors = np.array(vectors) | |
60 | - #print(vectors.shape) | |
61 | - #self.emb = theano.shared(vectors.astype(theano.config.floatX)) | |
62 | - | |
63 | - #ne = i | |
64 | - #de = len(vectors[0]) | |
65 | - | |
66 | - #bedzie trzeba obsluzyc przypadek, gdy slowo w danych nie ma embeddina w modelu | |
67 | - | |
68 | - ###del model | |
69 | - #del vectors | |
70 | - | |
71 | - #self.sent_states = theano.shared(0.2 * np.concatenate(( | |
72 | - # np.random.uniform(-1.0, 1.0,(ne, ds)),np.zeros((1,ds))),axis=0).astype(theano.config.floatX)) | |
73 | - # dodajemy jeden wektor zerowy potrzebny dla wyznaczenia sumy | |
74 | - # dzieci z liscii (czyli lisc symbolicznie ma dziecko bedace nullem - i to ma zerowy sentyment) | |
75 | - # uzyc go tez do reprezentacji rzadkich slow na zbiorze treningowym? | |
76 | - # porownac dzialanie: 1) przyjecie wektora zerowego dla nowych slow w zbiorze tren; 2) wziecie wartosci ze slowa najbardziej podobnego wzgledem embeddingu wystepujacego w zbiorze tren | |
77 | - # trzeba bedzie to uwzglednic w stosowaniu sieci | |
78 | - | |
79 | - r = 0.05 | |
80 | - | |
81 | - | |
82 | - #self.W_e_h = theano.shared(r * np.random.uniform(-1.0, 1.0,\ | |
83 | - # (de, nh)).astype(theano.config.floatX)) | |
84 | - self.W_e_h = theano.shared(np.load("saved_models4/W_eh25.npy").astype(theano.config.floatX)) | |
85 | - | |
86 | - self.W_sh = theano.shared(r * np.random.uniform(-1.0, 1.0,\ | |
87 | - (ds, nh)).astype(theano.config.floatX)) | |
88 | - | |
89 | - #self.W_h2_y = theano.shared(r * np.random.uniform(-1.0, 1.0,\ | |
90 | - # (2*nh, nc)).astype(theano.config.floatX)) | |
91 | - self.W_h2_y = theano.shared(np.load("saved_models4/W_hh225.npy").astype(theano.config.floatX)) | |
92 | - | |
93 | - #self.W_h_h2 = theano.shared(r * np.random.uniform(-1.0, 1.0,\ | |
94 | - # (2*nh, 2*nh)).astype(theano.config.floatX)) | |
95 | - self.W_h_h2 = theano.shared(np.load("saved_models4/W_h2y25.npy").astype(theano.config.floatX)) | |
96 | - | |
97 | - self.W_ssy = theano.shared(r * np.random.uniform(-1.0, 1.0,\ | |
98 | - (ds, nc)).astype(theano.config.floatX)) | |
99 | - | |
100 | - #self.W_sh_h = theano.shared(r * np.random.uniform(-1.0, 1.0,\ | |
101 | - # (2*nh, nh)).astype(theano.config.floatX)) | |
102 | - self.W_sh_h = theano.shared(np.load("saved_models4/W_shsh25.npy").astype(theano.config.floatX)) | |
103 | - | |
104 | - self.bh = theano.shared(np.zeros(nh, dtype=theano.config.floatX)) | |
105 | - self.b = theano.shared(np.zeros(nc, dtype=theano.config.floatX)) | |
106 | - | |
107 | - | |
108 | - # bundle | |
109 | - self.params = [ self.W_h2_y, self.W_h_h2, self.W_e_h, self.W_sh_h,self.emb]#, self.bh, self.b ] | |
110 | - self.names = [ "W_hh2", 'W_h2y', 'W_eh', 'W_shsh', "embeddings"]#, 'bh', 'b']#, 'h0'] | |
111 | - | |
112 | - | |
113 | - # liczy sentyment obecnego slowa / predykcja | |
114 | - # word_id = obecne slowo | |
115 | - # i = indeks w zdaniu slowa word_id | |
116 | - # word_children_ids = id-ki dzieci obecnego slowa | |
117 | - # word_children_positions = pozycje word_children_ids | |
118 | - def one_step(word_id, word_children_ids, word_children_positions, i, hidden_states): | |
119 | - | |
120 | - | |
121 | - | |
122 | - idx_tmp = (word_children_positions>=0).nonzero() | |
123 | - tmp = T.zeros_like(word_children_positions) | |
124 | - tmp2 = T.set_subtensor(tmp[idx_tmp], 1) | |
125 | - number_of_children = tmp2.sum() | |
126 | - | |
127 | - #pnoc = theano.printing.Print('Number of children: ') | |
128 | - #printed_number_of_children = pnoc(number_of_children) | |
129 | - | |
130 | - | |
131 | - # sprobowac zamiast zer, wstawic wektor wartosci 0.5 | |
132 | - schh = hidden_states[word_children_positions].sum(axis=0) /( number_of_children + 0.000001) #dodane 0..1, zeby nie bawic sie w ify, gdy nie ma dzieci (wtedy suma i tak jest zero, wiece dzielenie nie ma znaczenia) | |
133 | - h = T.concatenate([T.dot(self.emb[word_id],self.W_e_h), T.dot(schh,self.W_sh_h)]) # bez biasa i sigmoida | |
134 | - | |
135 | - #h = T.nnet.sigmoid(T.dot(self.emb[word_id],self.W_eh) + T.dot(schh,self.W_shsh) + self.bh) | |
136 | - | |
137 | - #h_s = T.zeros_like(hidden_states) | |
138 | - #zeros_subtensor = h_s[i] | |
139 | - #new_h_s = T.set_subtensor(zeros_subtensor, h) | |
140 | - | |
141 | - zeros_subtensor = hidden_states[i] | |
142 | - hidden_states_new = T.set_subtensor(zeros_subtensor, h) | |
143 | - | |
144 | - h2 = T.dot(h, self.W_h_h2) | |
145 | - | |
146 | - y_prob = T.nnet.softmax(T.dot(h2,self.W_h2_y))# + self.b) | |
147 | - | |
148 | - | |
149 | - # powyzsze jest niezbyt sensownie zrobione, bo jesli jest kilka "-1" w dzieciach to tyle razy jest dodawany ten wektor | |
150 | - # czy da sie to zamienic na petle, zeby nie dodawac wektora -1 kilka razy? | |
151 | - # w tej chwili to nie ma znaczenia, bo ten wektor i tak jest stale rowny 0 - nie zmienia sie podczas uczenia | |
152 | - | |
153 | - return i+1, hidden_states_new, y_prob | |
154 | - | |
155 | - | |
156 | - | |
157 | - words = T.vector(dtype=dataType) | |
158 | - children_ids = T.matrix(dtype=dataType) | |
159 | - children_positions = T.matrix(dtype=dataType) | |
160 | - | |
161 | - y_probs, _ = theano.scan(fn=one_step, \ | |
162 | - sequences = [words, children_ids, children_positions], | |
163 | - outputs_info = [theano.shared(0), | |
164 | - theano.shared(np.zeros((self.max_phrase_length+1,2*nh), dtype = theano.config.floatX)), | |
165 | - None], | |
166 | - n_steps = words.shape[0]) | |
167 | - | |
168 | - | |
169 | - estimated_probs = y_probs[-1][-1][0] | |
170 | - | |
171 | - y_pred = T.argmax(estimated_probs) # y_probs[-1][-1][0] zwraca wektor [P(y=0), P(y=1)] -> argmax zwraca predykce klasy | |
172 | - # dostajemy sie do predykcji dla ostatniego slowa, a klasyfikacja ostatniego slowa odpowiada klasyfikacji frazy, | |
173 | - # bo slowa sa ustawione w takiej kolejnosci, ze korzen jest ostatnim slowem | |
174 | - | |
175 | - | |
176 | - y = T.scalar('y',dtype=dataType) | |
177 | - | |
178 | - # cost and gradients and learning rate | |
179 | - lr = T.scalar('lr',dtype=theano.config.floatX) | |
180 | - | |
181 | - nll = -T.log(estimated_probs)[y] #to samo co (sprawdzone): | |
182 | - #nll = T.nnet.nnet.categorical_crossentropy(estimated_probs,T.extra_ops.to_one_hot(y.dimshuffle('x'), 5)[0]) | |
183 | - | |
184 | - gradients = T.grad( nll, self.params ) | |
185 | - updates = OrderedDict(( p, p-lr*g ) for p, g in zip( self.params , gradients)) | |
186 | - | |
187 | - # uwaga: ostani rzad macierzy sent_states - wektor odpowiadajacy dziecku, ktorego nie ma - jest stale rowny zero | |
188 | - | |
189 | - | |
190 | - # theano functions | |
191 | - self.classify = theano.function(inputs=[words,children_ids,children_positions], outputs=y_pred, | |
192 | - allow_input_downcast=True, | |
193 | - mode='FAST_RUN' ) | |
194 | - | |
195 | - self.train = theano.function( inputs = [words,children_ids, children_positions, y, lr], | |
196 | - outputs = nll, | |
197 | - updates = updates, | |
198 | - allow_input_downcast=True, | |
199 | - mode='FAST_RUN' ) | |
200 | - | |
201 | - | |
202 | - #self.normalize = theano.function( inputs = [], #uwazac na dzielenie przez 0 - ostatni wiersz sent_states jest zerowy | |
203 | - # updates = {self.sent_states:\ | |
204 | - # self.sent_states/T.sqrt((self.sent_states**2).sum(axis=1))})#.dimshuffle(0,'x')}) | |
205 | - | |
206 | - def save(self, folder, e): | |
207 | - for param, name in zip(self.params, self.names): | |
208 | - np.save(os.path.join(folder, name + str(e) + '.npy'), param.get_value()) | |
209 | - | |
210 | - | |
211 | - | |
212 | - | |
213 | - | |
214 | - | |
215 | -class model2(object): | |
216 | - | |
217 | - ''' | |
218 | - | |
219 | - ''' | |
220 | - | |
221 | - | |
222 | - def __init__(self, nh, nc, ds, w2v_model_path, max_phrase_length): | |
223 | - ''' | |
224 | - nh :: dimension of hidden state | |
225 | - nc :: number of classes | |
226 | - ne :: number of word embeddings in the vocabulary | |
227 | - de :: dimension of the word embeddings | |
228 | - | |
229 | - ds :: dimension of the sentiment state | |
230 | - ''' | |
231 | - | |
232 | - self.max_phrase_length = max_phrase_length | |
233 | - | |
234 | - w2vecs = pickle.load(open(w2v_model_path,"r")) | |
235 | - | |
236 | - self.emb = theano.shared(w2vecs["vectors"].astype(theano.config.floatX)) | |
237 | - #self.emb = theano.shared(np.load("saved_models_final1/embeddings"+str(e)+"_200.npy").astype(theano.config.floatX)) | |
238 | - | |
239 | - self.words2ids = w2vecs["words2ids"] | |
240 | - | |
241 | - ne = len(w2vecs["words2ids"]) | |
242 | - de = w2vecs["vectors"].shape[1] | |
243 | - | |
244 | - del w2vecs | |
245 | - | |
246 | - #self.sent_states = theano.shared(0.2 * np.concatenate(( | |
247 | - # np.random.uniform(-1.0, 1.0,(ne, ds)),np.zeros((1,ds))),axis=0).astype(theano.config.floatX)) | |
248 | - # dodajemy jeden wektor zerowy potrzebny dla wyznaczenia sumy | |
249 | - # dzieci z liscii (czyli lisc symbolicznie ma dziecko bedace nullem - i to ma zerowy sentyment) | |
250 | - # uzyc go tez do reprezentacji rzadkich slow na zbiorze treningowym? | |
251 | - # porownac dzialanie: 1) przyjecie wektora zerowego dla nowych slow w zbiorze tren; 2) wziecie wartosci ze slowa najbardziej podobnego wzgledem embeddingu wystepujacego w zbiorze tren | |
252 | - # trzeba bedzie to uwzglednic w stosowaniu sieci | |
253 | - | |
254 | - r = 0.05 | |
255 | - | |
256 | - | |
257 | - self.W_e_h = theano.shared(r * np.random.uniform(-1.0, 1.0,\ | |
258 | - (de, nh)).astype(theano.config.floatX)) | |
259 | - #self.W_e_h = theano.shared(np.load("saved_models_final1/W_eh"+str(e)+"_200.npy").astype(theano.config.floatX)) | |
260 | - | |
261 | - self.W_sh = theano.shared(r * np.random.uniform(-1.0, 1.0,\ | |
262 | - (ds, nh)).astype(theano.config.floatX)) | |
263 | - | |
264 | - self.W_h2_y = theano.shared(r * np.random.uniform(-1.0, 1.0,\ | |
265 | - (2*nh, nc)).astype(theano.config.floatX)) | |
266 | - #self.W_h2_y = theano.shared(np.load("saved_models_final1/W_h2y"+str(e)+"_200.npy").astype(theano.config.floatX)) | |
267 | - | |
268 | - self.W_h_h2 = theano.shared(r * np.random.uniform(-1.0, 1.0,\ | |
269 | - (2*nh, 2*nh)).astype(theano.config.floatX)) | |
270 | - #self.W_h_h2 = theano.shared(np.load("saved_models_final1/W_hh2"+str(e)+"_200.npy").astype(theano.config.floatX)) | |
271 | - | |
272 | - self.W_ssy = theano.shared(r * np.random.uniform(-1.0, 1.0,\ | |
273 | - (ds, nc)).astype(theano.config.floatX)) | |
274 | - | |
275 | - self.W_sh_h = theano.shared(r * np.random.uniform(-1.0, 1.0,\ | |
276 | - (2*nh, nh)).astype(theano.config.floatX)) | |
277 | - #self.W_sh_h = theano.shared(np.load("saved_models_final1/W_shsh"+str(e)+"_200.npy").astype(theano.config.floatX)) | |
278 | - | |
279 | - | |
280 | - self.W_h_y = theano.shared(r * np.random.uniform(-1.0, 1.0,\ | |
281 | - (2*nh, nc)).astype(theano.config.floatX)) | |
282 | - | |
283 | - self.bh = theano.shared(np.zeros(nh, dtype=theano.config.floatX)) | |
284 | - self.b = theano.shared(np.zeros(nc, dtype=theano.config.floatX)) | |
285 | - | |
286 | - | |
287 | - # bundle | |
288 | - self.params = [ self.W_h_y, self.W_e_h, self.W_sh_h, self.emb]# self.W_h2_y, self.W_h_h2, | |
289 | - self.names = [ "W_h_y", 'W_eh', 'W_shsh', "embeddings"]# 'W_h2y', "W_hh2", | |
290 | - | |
291 | - | |
292 | - shared_zero = theano.shared(0) | |
293 | - shared_one = theano.shared(1) | |
294 | - | |
295 | - # liczy sentyment obecnego slowa / predykcja | |
296 | - # word_id = obecne slowo | |
297 | - # i = indeks w zdaniu slowa word_id | |
298 | - # word_children_ids = id-ki dzieci obecnego slowa | |
299 | - # word_children_positions = pozycje word_children_ids | |
300 | - def one_step(word_id, word_children_ids, word_children_positions, y_true, i, hidden_states, learning_rate): | |
301 | - | |
302 | - p = printing.Print('word_children_positions: ') | |
303 | - word_children_positions = p(word_children_positions) | |
304 | - | |
305 | - | |
306 | - idx_tmp = (word_children_positions>=0).nonzero() | |
307 | - tmp = T.zeros_like(word_children_positions) | |
308 | - tmp2 = T.set_subtensor(tmp[idx_tmp], 1) | |
309 | - number_of_children = tmp2.sum(dtype = dataType) | |
310 | - | |
311 | - number_of_children = ifelse(T.eq(number_of_children, shared_zero), shared_one, number_of_children) | |
312 | - # sprobowac zamiast zer, wstawic wektor wartosci 0.5 | |
313 | - | |
314 | - hello_world_op = printing.Print('number_of_children: ') | |
315 | - number_of_children = hello_world_op(number_of_children) | |
316 | - | |
317 | - | |
318 | - schh = hidden_states[word_children_positions].sum(axis=0) / number_of_children#( number_of_children + 0.000001) | |
319 | -#dodane 0..1, zeby nie bawic sie w ify, gdy nie ma dzieci (wtedy suma i tak jest zero, wiece dzielenie nie ma znaczenia) | |
320 | - h = T.concatenate([T.dot(self.emb[word_id],self.W_e_h), T.dot(schh,self.W_sh_h)]) # bez biasa i sigmoida | |
321 | - | |
322 | - #h = T.nnet.sigmoid(T.dot(self.emb[word_id],self.W_eh) + T.dot(schh,self.W_shsh) + self.bh) | |
323 | - | |
324 | - #h_s = T.zeros_like(hidden_states) | |
325 | - #zeros_subtensor = h_s[i] | |
326 | - #new_h_s = T.set_subtensor(zeros_subtensor, h) | |
327 | - | |
328 | - zeros_subtensor = hidden_states[i] | |
329 | - hidden_states_new = T.set_subtensor(zeros_subtensor, h) | |
330 | - | |
331 | - #h2 = T.dot(h, self.W_h_h2) | |
332 | - | |
333 | - #y_prob = T.nnet.softmax(T.dot(h2,self.W_h2_y))# + self.b) | |
334 | - | |
335 | - y_prob = T.nnet.softmax(T.dot(h,self.W_h_y))# + self.b) | |
336 | - | |
337 | - cce = -T.log(y_prob[0][y_true]) | |
338 | - | |
339 | - #learning_rate = 0.01 | |
340 | - | |
341 | - updates = OrderedDict([#(self.W_h2_y, self.W_h2_y-learning_rate*T.grad(cce, self.W_h2_y)), | |
342 | - (self.W_h_y, self.W_h_y-learning_rate*T.grad(cce, self.W_h_y)), | |
343 | - #(self.W_h_h2, self.W_h_h2-learning_rate*T.grad(cce, self.W_h_h2)), | |
344 | - (self.W_e_h, self.W_e_h-learning_rate*T.grad(cce, self.W_e_h)), | |
345 | - (self.W_sh_h, self.W_sh_h-learning_rate*T.grad(cce, self.W_sh_h)), | |
346 | - (self.emb, self.emb-learning_rate*T.grad(cce, self.emb)) | |
347 | - ]) | |
348 | - | |
349 | - | |
350 | - return (i+1,hidden_states_new, y_prob), updates | |
351 | - | |
352 | - | |
353 | - | |
354 | - | |
355 | - y = T.vector('y',dtype=dataType) | |
356 | - | |
357 | - lr = T.scalar('lr',dtype=theano.config.floatX) | |
358 | - | |
359 | - words = T.vector(dtype=dataType) | |
360 | - children_ids = T.matrix(dtype=dataType) | |
361 | - children_positions = T.matrix(dtype=dataType) | |
362 | - #words_indexes = T.vector(dtype=dataType) | |
363 | - | |
364 | - y_probs, upd = theano.scan(fn=one_step, \ | |
365 | - sequences = [words, children_ids, children_positions,y],#,words_indexes], | |
366 | - outputs_info = [theano.shared(0), | |
367 | - theano.shared(np.zeros((self.max_phrase_length+1,2*nh), dtype = theano.config.floatX)), | |
368 | - None], | |
369 | - non_sequences = lr, | |
370 | - n_steps = words.shape[0]) | |
371 | - | |
372 | - | |
373 | - def one_step_classify(word_id, word_children_ids, word_children_positions, i, hidden_states): | |
374 | - | |
375 | - | |
376 | - idx_tmp = (word_children_positions>=0).nonzero() | |
377 | - tmp = T.zeros_like(word_children_positions) | |
378 | - tmp2 = T.set_subtensor(tmp[idx_tmp], 1) | |
379 | - number_of_children = tmp2.sum() | |
380 | - | |
381 | - schh = hidden_states[word_children_positions].sum(axis=0) / ifelse(T.eq(number_of_children, shared_zero), shared_one, number_of_children) | |
382 | - h = T.concatenate([T.dot(self.emb[word_id],self.W_e_h), T.dot(schh,self.W_sh_h)]) # bez biasa i sigmoida | |
383 | - | |
384 | - zeros_subtensor = hidden_states[i] | |
385 | - hidden_states_new = T.set_subtensor(zeros_subtensor, h) | |
386 | - | |
387 | - #h2 = T.dot(h, self.W_h_h2) | |
388 | - #y_prob = T.nnet.softmax(T.dot(h2,self.W_h2_y))# + self.b) | |
389 | - y_prob = T.nnet.softmax(T.dot(h,self.W_h_y)) | |
390 | - | |
391 | - return i+1, hidden_states_new, y_prob | |
392 | - | |
393 | - | |
394 | - | |
395 | - y_probs_classify, _ = theano.scan(fn=one_step_classify, \ | |
396 | - sequences = [words, children_ids, children_positions], | |
397 | - outputs_info = [theano.shared(0), | |
398 | - theano.shared(np.zeros((self.max_phrase_length+1,2*nh), dtype = theano.config.floatX)), | |
399 | - None], | |
400 | - n_steps = words.shape[0]) | |
401 | - | |
402 | - | |
403 | - | |
404 | - | |
405 | - predictions, _ = theano.scan(lambda i: (i+1, T.argmax(y_probs_classify[2][i][0])), outputs_info = [theano.shared(0), None], n_steps = y_probs_classify[2].shape[0]) | |
406 | - | |
407 | - #res2 , _ = theano.scan(lambda x,i : (i+1, T.argmax(x)), | |
408 | - # sequences = [estimated_probs[1]], | |
409 | - # outputs_info = [theano.shared(0), None] | |
410 | - # ) | |
411 | - | |
412 | - | |
413 | -# minus_log_true_class_prob = res[1] | |
414 | - #prediction_class = res2[1] | |
415 | - | |
416 | - | |
417 | -# nll = minus_log_true_class_prob.sum() | |
418 | - | |
419 | - #y_pred = T.argmax(estimated_probs) # y_probs[-1][-1][0] zwraca wektor [P(y=0), P(y=1), ...] -> argmax zwraca predykce klasy | |
420 | - # dostajemy sie do predykcji dla ostatniego slowa, a klasyfikacja ostatniego slowa odpowiada klasyfikacji frazy, | |
421 | - # bo slowa sa ustawione w takiej kolejnosci, ze korzen jest ostatnim slowem | |
422 | - | |
423 | - | |
424 | - | |
425 | - | |
426 | - # cost and gradients and learning rate | |
427 | - #nll = -T.log(estimated_probs[1])[y] #to samo co (sprawdzone): | |
428 | - #nll = T.nnet.nnet.categorical_crossentropy(estimated_probs,T.extra_ops.to_one_hot(y.dimshuffle('x'), 5)[0]) | |
429 | - | |
430 | -# gradients = T.grad( nll, self.params ) | |
431 | -# updates = OrderedDict(( p, p-lr*g ) for p, g in zip( self.params , gradients)) | |
432 | - | |
433 | - # uwaga: ostani rzad macierzy sent_states - wektor odpowiadajacy dziecku, ktorego nie ma - jest stale rowny zero | |
434 | - | |
435 | - | |
436 | - # theano functions | |
437 | - self.classify = theano.function(inputs=[words,children_ids,children_positions], outputs=predictions[1], | |
438 | - allow_input_downcast=True, | |
439 | - mode='FAST_RUN' ) | |
440 | - | |
441 | - self.train = theano.function( inputs = [words,children_ids, children_positions, y, lr],#, words_indexes | |
442 | - outputs = [],#nll, | |
443 | - updates = upd,#updates, | |
444 | - allow_input_downcast=True, | |
445 | - mode='FAST_RUN' ) | |
446 | - | |
447 | - | |
448 | - #self.normalize = theano.function( inputs = [], #uwazac na dzielenie przez 0 - ostatni wiersz sent_states jest zerowy | |
449 | - # updates = {self.sent_states:\ | |
450 | - # self.sent_states/T.sqrt((self.sent_states**2).sum(axis=1))})#.dimshuffle(0,'x')}) | |
451 | - | |
452 | - def save(self, folder, e, i): | |
453 | - for param, name in zip(self.params, self.names): | |
454 | - np.save(os.path.join(folder, name + str(e) + "_" + str(i) + '.npy'), param.get_value()) |
modules/rnn/tmp.pyc deleted
No preview for this file type
modules/utils/tools.py
... | ... | @@ -3,8 +3,12 @@ import numpy |
3 | 3 | from keras.preprocessing import sequence as seq |
4 | 4 | import theano |
5 | 5 | |
6 | +from collections import Counter | |
7 | + | |
6 | 8 | import pickle |
7 | 9 | |
10 | + | |
11 | + | |
8 | 12 | def shuffle(lol, seed): |
9 | 13 | ''' |
10 | 14 | lol :: list of list as input |
... | ... | @@ -70,7 +74,6 @@ def filter_embeddings(datasets, embedding_path, destination): |
70 | 74 | |
71 | 75 | |
72 | 76 | |
73 | - | |
74 | 77 | def words_in_from_down_to_top_order(sentence_tree): |
75 | 78 | #print sentence_tree |
76 | 79 | levels = numpy.setdiff1d(range(len(sentence_tree)),numpy.unique(sentence_tree)) # - zwraca slowo/a, ktore nie jest niczyim dzieckiem - czyli powinno/y byc korzeniem/korzeniami frazy/fraz |
... | ... | @@ -81,7 +84,8 @@ def words_in_from_down_to_top_order(sentence_tree): |
81 | 84 | for i in range(len(sentence_tree)): |
82 | 85 | #print i |
83 | 86 | #print levels[i] |
84 | - levels.extend(numpy.setdiff1d(sentence_tree[levels[i]],-1)) | |
87 | + x = numpy.setdiff1d(sentence_tree[levels[i]],-1) | |
88 | + levels.extend(x[x<len(sentence_tree)]) | |
85 | 89 | |
86 | 90 | ordered_words = numpy.array(levels)[levels != numpy.array(-1)][::-1] #odwaracmy kolejnosc na poczatku beda slowa znajdujace sie najglebiej |
87 | 91 | |
... | ... | @@ -94,7 +98,6 @@ def words_in_from_down_to_top_order(sentence_tree): |
94 | 98 | |
95 | 99 | |
96 | 100 | |
97 | - | |
98 | 101 | def load_conll_data(conll_format_data, words2ids): |
99 | 102 | |
100 | 103 | |
... | ... | @@ -633,12 +636,7 @@ def load_stanford_data3(labels, parents, tokens, words2ids, use_batch, batch_siz |
633 | 636 | |
634 | 637 | def load_stanford_data4(labels, parents, tokens, words2ids, use_batch, batch_size, nb_classes): |
635 | 638 | |
636 | - | |
637 | - | |
638 | - | |
639 | 639 | def transform_labels(x, nb_classes): |
640 | - | |
641 | - | |
642 | 640 | if nb_classes == 3: |
643 | 641 | if x =='#' or int(x) == 0: |
644 | 642 | return 1 |
... | ... | @@ -685,19 +683,12 @@ def load_stanford_data4(labels, parents, tokens, words2ids, use_batch, batch_siz |
685 | 683 | |
686 | 684 | for labels_i,parents_i,tokens_i in zip(labels,parents,tokens): |
687 | 685 | |
688 | - | |
689 | - | |
690 | 686 | k = k + 1 |
691 | - | |
692 | - | |
687 | + | |
693 | 688 | s = [] |
694 | 689 | for i in range(len(tokens_i)): |
695 | 690 | s.append([i,int(parents_i[i]),labels_i[i],tokens_i[i]]) |
696 | 691 | |
697 | - | |
698 | - | |
699 | - | |
700 | - | |
701 | 692 | if len(s) == 1 and use_batch == False: #przypadek gdy fraza sklada sie z jednego tokena |
702 | 693 | |
703 | 694 | #if nb_classes == 2: |
... | ... | @@ -743,7 +734,6 @@ def load_stanford_data4(labels, parents, tokens, words2ids, use_batch, batch_siz |
743 | 734 | # if current_sentence[3][-1] <0: |
744 | 735 | # continue |
745 | 736 | |
746 | - | |
747 | 737 | if use_batch == True: |
748 | 738 | |
749 | 739 | # w tej chwili len(current_sentence[0]) nie jest nigdzie wykorzystywane |
... | ... | @@ -770,8 +760,7 @@ def load_stanford_data4(labels, parents, tokens, words2ids, use_batch, batch_siz |
770 | 760 | #wyrzucamy macierz id dzieci batch_children_ids.append(current_batch[sent][0][1][tok]) |
771 | 761 | batch_labels.append(current_batch[sent][0][2][tok]) |
772 | 762 | batch_words.append(current_batch[sent][0][3][tok]) |
773 | - | |
774 | - | |
763 | + | |
775 | 764 | #wyrzucamy macierz id dzieci batch_children_ids = seq.pad_sequences(batch_children_ids, padding='post', value = -1) |
776 | 765 | batch_children_positions = seq.pad_sequences(batch_children_positions, padding='post', value = -1) |
777 | 766 | |
... | ... | @@ -785,8 +774,7 @@ def load_stanford_data4(labels, parents, tokens, words2ids, use_batch, batch_siz |
785 | 774 | |
786 | 775 | current_batch, batch_tokens, batch_children_positions, batch_labels = [], [], [], [] |
787 | 776 | batch_words = [] |
788 | - | |
789 | - | |
777 | + | |
790 | 778 | else: |
791 | 779 | |
792 | 780 | sentences.append(current_sentence) |
... | ... | @@ -826,15 +814,13 @@ def load_stanford_data4(labels, parents, tokens, words2ids, use_batch, batch_siz |
826 | 814 | numpy.array(batch_labels) |
827 | 815 | ,numpy.array(batch_words) |
828 | 816 | ]) |
829 | - | |
830 | - | |
831 | - | |
832 | - | |
817 | + | |
833 | 818 | return sentences |
834 | 819 | |
835 | 820 | |
836 | 821 | |
837 | 822 | |
823 | + | |
838 | 824 | def load_stanford_data5(labels, parents, tokens, words2ids, use_batch, batch_size, nb_classes): |
839 | 825 | |
840 | 826 | |
... | ... | @@ -1033,5 +1019,200 @@ def load_stanford_data5(labels, parents, tokens, words2ids, use_batch, batch_siz |
1033 | 1019 | |
1034 | 1020 | |
1035 | 1021 | |
1022 | +def load_stanford_data6(labels, parents, tokens, relations, words2ids, use_batch, batch_size, nb_classes, k_most_common_relations): | |
1023 | + | |
1024 | + def transform_labels(x, nb_classes): | |
1025 | + if nb_classes == 3: | |
1026 | + if x =='#' or int(x) == 0: | |
1027 | + return 1 | |
1028 | + elif int(x) < 0: | |
1029 | + return 0 | |
1030 | + else: | |
1031 | + return 2 | |
1032 | + elif nb_classes == 5: | |
1033 | + if x =='#': | |
1034 | + return 2 | |
1035 | + else: | |
1036 | + return int(x)+2 | |
1037 | + # elif nb_classes == 2: #jesli chcemy miec dwie klasy to neutralne wyrzucamy ze zbioru, | |
1038 | + # if x =='#' or int(x) == 0: | |
1039 | + # return -1 | |
1040 | + # elif int(x) < 0: | |
1041 | + # return 0 | |
1042 | + # else: | |
1043 | + # return 1 | |
1044 | + | |
1045 | + sentences = [] | |
1046 | + | |
1047 | + l = open(labels, "r") | |
1048 | + # 5 klas: labels = [[2 if y=='#' else int(y)+2 for y in x.split()] for x in l.readlines()] | |
1049 | + | |
1050 | + # Na ten moment przyjmujemy wartosc "2" w miejsce "#" | |
1051 | + | |
1052 | + labels = [[transform_labels(y,nb_classes) for y in x.split()] for x in l.readlines()] | |
1053 | + l.close() | |
1054 | + | |
1055 | + p = open(parents,"r") | |
1056 | + parents = [[int(y) for y in x.split()] for x in p.readlines()] | |
1057 | + p.close() | |
1058 | + | |
1059 | + t = open(tokens,"r") | |
1060 | + tokens = [x.split() for x in t.readlines()] | |
1061 | + t.close() | |
1062 | + | |
1063 | + | |
1064 | + rels = open(relations,"r") | |
1065 | + relations = [[y for y in x.split()] for x in rels.readlines()] | |
1066 | + rels.close() | |
1067 | + most_common_rels = [x[0] for x in Counter(numpy.concatenate(relations)).most_common(k_most_common_relations)] | |
1068 | + transform_rels = dict(zip(most_common_rels,range(len(most_common_rels)))) | |
1069 | + relations = [[transform_rels.get(x, k_most_common_relations) for x in sent] for sent in relations] | |
1070 | + | |
1071 | + | |
1072 | + k = 0 | |
1073 | + sentence_length = 0 | |
1074 | + current_batch, batch_tokens, batch_children_ids, batch_children_positions, batch_labels, batch_relations = [], [], [], [], [], [] | |
1075 | + batch_words = [] | |
1076 | + | |
1077 | + for labels_i, parents_i, tokens_i, relations_i in zip(labels,parents,tokens,relations): | |
1078 | + | |
1079 | + k = k + 1 | |
1080 | + | |
1081 | + s = [] | |
1082 | + for i in range(len(tokens_i)): | |
1083 | + s.append([i,int(parents_i[i]),labels_i[i],tokens_i[i],relations_i[i]]) | |
1084 | + | |
1085 | + if len(s) == 1 and use_batch == False: #przypadek gdy fraza sklada sie z jednego tokena | |
1086 | + | |
1087 | + #if nb_classes == 2: | |
1088 | + # if s[0][-1] < 0: | |
1089 | + # continue | |
1090 | + | |
1091 | + sentences.append([\ | |
1092 | + numpy.array([words2ids.get(tokens[0], -1)]),\ | |
1093 | + #wyrzucamy macierz id dzieci numpy.array([-1], ndmin=2),\ | |
1094 | + numpy.array([-1], ndmin=2), \ | |
1095 | + numpy.array(labels_i[0]), \ | |
1096 | + numpy.array(relations_i[0]) | |
1097 | + ]) | |
1098 | + | |
1099 | + else: | |
1100 | + | |
1101 | + for i in range(len(s)): # nie wiem czy sie nie wywali dla frazy dlugosci 1 | |
1102 | + children = [] | |
1103 | + for j in range(len(s)): | |
1104 | + if s[j][1] == i+1: | |
1105 | + children.append(s[j][0]) | |
1106 | + s[i].append(children) | |
1107 | + | |
1108 | + words = [x[0] for x in s] | |
1109 | + children = seq.pad_sequences([x[-1] for x in s], padding='post', value = -1) | |
1110 | + tokens = [x[3] for x in s] | |
1111 | + labels_in_batch = [x[2] for x in s] | |
1112 | + relations = [x[4] for x in s] | |
1113 | + | |
1114 | + ordered_words, order = words_in_from_down_to_top_order(children) | |
1115 | + | |
1116 | + if ordered_words is None: | |
1117 | + continue | |
1118 | + | |
1119 | + current_sentence = [ | |
1120 | + numpy.array([words2ids.get(x,-1) for x in tokens])[ordered_words], | |
1121 | + #wyrzucamy macierz id dzieci numpy.array([[words2ids.get(tokens[w],-1) if w>=0 else -1 for w in x] | |
1122 | + # for x in children[ordered_words]]), | |
1123 | + numpy.array([[order[w] if w>= 0 else -1 for w in x] for x in children[ordered_words]]), | |
1124 | + numpy.array(labels_in_batch)[ordered_words], | |
1125 | + numpy.array(relations)[ordered_words] , | |
1126 | + numpy.array(words) | |
1127 | + ] | |
1128 | + #if nb_classes == 2: | |
1129 | + # if current_sentence[3][-1] <0: | |
1130 | + # continue | |
1131 | + | |
1132 | + if use_batch == True: | |
1133 | + | |
1134 | + # w tej chwili len(current_sentence[0]) nie jest nigdzie wykorzystywane | |
1135 | + current_batch.append((current_sentence, len(current_sentence[0]))) | |
1136 | + | |
1137 | + if len(current_batch) % batch_size == 0: | |
1138 | + | |
1139 | + shift = 0 | |
1140 | + | |
1141 | + for sent in range(batch_size): | |
1142 | + | |
1143 | + ##if sent > 0: | |
1144 | + ## shift = shift + current_batch[sent-1][1] | |
1145 | + | |
1146 | + for tok in range(len(current_batch[sent][0][0])): | |
1147 | + | |
1148 | + if sent == 0: | |
1149 | + batch_children_positions.append(current_batch[sent][0][1][tok]) | |
1150 | + else: | |
1151 | + batch_children_positions.append([chd+shift if chd>=0 else -1 for chd in current_batch[sent][0][1][tok]]) | |
1152 | + #batch_children_positions.append(current_batch[sent][0][2][tok]) | |
1036 | 1153 | |
1154 | + batch_tokens.append(current_batch[sent][0][0][tok]) | |
1155 | + #wyrzucamy macierz id dzieci batch_children_ids.append(current_batch[sent][0][1][tok]) | |
1156 | + batch_labels.append(current_batch[sent][0][2][tok]) | |
1157 | + batch_relations.append(current_batch[sent][0][3][tok]) | |
1158 | + batch_words.append(current_batch[sent][0][4][tok]) | |
1159 | + | |
1160 | + #wyrzucamy macierz id dzieci batch_children_ids = seq.pad_sequences(batch_children_ids, padding='post', value = -1) | |
1161 | + batch_children_positions = seq.pad_sequences(batch_children_positions, padding='post', value = -1) | |
1162 | + | |
1163 | + sentences.append([ | |
1164 | + numpy.array(batch_tokens), | |
1165 | + #wyrzucamy macierz id dzieci numpy.array(batch_children_ids), | |
1166 | + numpy.array(batch_children_positions), | |
1167 | + numpy.array(batch_labels), | |
1168 | + numpy.array(batch_relations) | |
1169 | + ,numpy.array(batch_words) | |
1170 | + ]) | |
1171 | + | |
1172 | + current_batch, batch_tokens, batch_children_positions, batch_labels, batch_relations = [], [], [], [], [] | |
1173 | + batch_words = [] | |
1174 | + | |
1175 | + else: | |
1176 | + | |
1177 | + sentences.append(current_sentence) | |
1178 | + | |
1179 | + | |
1180 | + # gdy liczba zdan nie jest wilokrotnosci licznosci batch, to na koncu trzeba dodac pozostale zdania: | |
1181 | + if use_batch == True and len(current_batch) > 0: | |
1182 | + | |
1183 | + shift = 0 | |
1184 | + | |
1185 | + for sent in range(len(current_batch)): | |
1186 | + | |
1187 | + #if sent > 0: | |
1188 | + # shift = shift + current_batch[sent-1][1] | |
1189 | + | |
1190 | + for tok in range(len(current_batch[sent][0][0])): | |
1191 | + | |
1192 | + if sent == 0: | |
1193 | + batch_children_positions.append(current_batch[sent][0][1][tok]) | |
1194 | + else: | |
1195 | + batch_children_positions.append([chd+shift if chd>=0 else -1 for chd in current_batch[sent][0][1][tok]]) | |
1196 | + #batch_children_positions.append(current_batch[sent][0][2][tok]) | |
1197 | + | |
1198 | + batch_tokens.append(current_batch[sent][0][0][tok]) | |
1199 | + #wyrzucamy macierz id dzieci batch_children_ids.append(current_batch[sent][0][1][tok]) | |
1200 | + batch_labels.append(current_batch[sent][0][2][tok]) | |
1201 | + batch_relations.append(current_batch[sent][0][3][tok]) | |
1202 | + batch_words.append(current_batch[sent][0][4][tok]) | |
1203 | + | |
1204 | + | |
1205 | + #wyrzucamy macierz id dzieci batch_children_ids = seq.pad_sequences(batch_children_ids, padding='post', value = -1) | |
1206 | + batch_children_positions = seq.pad_sequences(batch_children_positions, padding='post', value = -1) | |
1207 | + | |
1208 | + sentences.append([ | |
1209 | + numpy.array(batch_tokens), | |
1210 | + #wyrzucamy macierz id dzieci numpy.array(batch_children_ids), | |
1211 | + numpy.array(batch_children_positions), | |
1212 | + numpy.array(batch_labels), | |
1213 | + numpy.array(batch_relations) | |
1214 | + ,numpy.array(batch_words) | |
1215 | + ]) | |
1216 | + | |
1217 | + return sentences | |
1037 | 1218 | |
... | ... |
modules/utils/tools.pyc
No preview for this file type
modules/rnn/many_models.py renamed to nieaktualne/many_models.py
modules/rnn/many_models.pyc renamed to nieaktualne/many_models.pyc
No preview for this file type
modules/rnn/nnet_for_dependency_trees.py renamed to nieaktualne/nnet_for_dependency_trees.py
modules/rnn/nnet_for_dependency_trees.pyc renamed to nieaktualne/nnet_for_dependency_trees.pyc
No preview for this file type