def get_X_y(wv, wv_dim):
pmids, sentences, lbls, vectorizer = parse_summerscales.get_tokens_and_lbls()
# see: https://github.com/fchollet/keras/issues/233
# num_sentences x 1 x max_token_len x wv_dim
# number of sequences x 1 x max number of tokens (padded to max len) x word vector size
num_sentences = len(sentences)
#max_token_len = max([len(s) for s in sentences])
#X_embedded = np.zeros((num_sentences, wv_dim))
X_embedded, X_tokens = [], [] # here a sequence associated with each doc/abstract
y = []
#X_tokens = []
cur_pmid = pmids[0]
cur_x_embedded, cur_x_tokens, cur_y = [], [], []
unknown_words_to_vecs = {}
for idx, s in enumerate(sentences):
if cur_pmid != pmids[idx]:
X_embedded.append(np.vstack(cur_x_embedded))
X_tokens.append(np.vstack(cur_x_tokens))
y.append(np.array(cur_y))
cur_x_embedded, cur_x_tokens, cur_y = [], [], []
cur_pmid = pmids[idx]
for j, t in enumerate(s):
try:
v = wv[t]
except:
# or maybe use 0s???
if not t in unknown_words_to_vecs:
v = np.random.uniform(-1,1,wv_dim)
unknown_words_to_vecs[t] = v
v = unknown_words_to_vecs[t]
cur_x_embedded.append(v)
cur_x_tokens.append(vectorizer.vocabulary_[t])
cur_y.extend(lbls[idx])
#cur_x_tokens.append(v)
X_embedded.append(np.vstack(cur_x_embedded))
X_tokens.append(np.vstack(cur_x_tokens))
y.append(np.array(cur_y))
return X_embedded, X_tokens, y, vectorizer, unknown_words_to_vecs
def get_PMIDs_to_X_y(use_pickle, use_coref):
pmids_dict, token_to_features = \
parse_summerscales.get_tokens_and_lbls(use_pickle=use_pickle, use_coref=use_coref)
pmids_to_X_y = {}
for pmid in pmids_dict:
pmid_sentences, pmid_lbls = pmids_dict[pmid]
# for this sentence
X_tokens, X_features = [], []
y = []
for sent_idx, s in enumerate(pmid_sentences):
for j, token in enumerate(s):
X_features.append(token_to_features[token])
X_tokens.append(token)
y.extend(pmid_lbls[sent_idx])
pmids_to_X_y[pmid] = (np.vstack(X_tokens), np.vstack(X_features),
np.hstack(y))
return (pmids_to_X_y, X_tokens)
def main():
n_folds = 5
try:
opts, args = getopt.getopt(sys.argv[1:], '', ['window_size=', 'wiki=', 'n_feature_maps=', 'epochs=',
'undersample=', 'n_feature_maps=', 'criterion=',
'optimizer=', 'model='])
except getopt.GetoptError as error:
print error
sys.exit(2)
model_type = 'nn'
window_size = 5
wiki = True
n_feature_maps = 100
epochs = 20
undersample = False
binary_cross_entropy = False
criterion = 'categorical_crossentropy'
optimizer = 'adam'
k = 2
for opt, arg in opts:
if opt == '--window_size':
window_size = int(arg)
elif opt == '--wiki':
if arg == 0:
wiki = False
elif opt == '--epochs':
epochs = int(arg)
elif opt == '--n_feature_maps':
n_feature_maps = int(arg)
elif opt == '--undersample':
option = int(arg)
if option == 1:
undersample = True
elif opt == '--n_feature_maps':
n_feature_maps = int(arg)
elif opt == '--criterion':
criterion = arg
elif opt == '--optimizer':
optimizer = arg
elif opt == '--model':
model_type = arg
else:
print "Option {} is not valid!".format(opt)
if criterion == 'binary_crossentropy':
binary_cross_entropy = True
k = 1
print('Loading word2vec model...')
if wiki:
print 'Using wiki word2vec...'
word2vec_model = 'wikipedia-pubmed-and-PMC-w2v.bin'
else:
print 'Using non-wiki word2vec...'
word2vec_model = 'PubMed-w2v.bin'
w2v = Word2Vec.load_word2vec_format(word2vec_model, binary=True)
print('Loaded word2vec model')
pmids_dict, pmids, abstracts, lbls, vectorizer, groups_map, one_hot, dicts = \
parse_summerscales.get_tokens_and_lbls(
make_pmids_dict=True, sen=True)
all_pmids = pmids_dict.keys()
n = len(all_pmids)
kf = KFold(n, random_state=1337, shuffle=True, n_folds=n_folds)
for fold_idx, (train, test) in enumerate(kf):
print("on fold %s" % fold_idx)
train_pmids = [all_pmids[pmid_idx] for pmid_idx in train]
test_pmids = [all_pmids[pmid_idx] for pmid_idx in test]
print train_pmids
print('loading data...')
if model_type == 'cnn':
X_train, y_train = _prep_data(train_pmids, pmids_dict, w2v, window_size, model_type, binary_ce=binary_cross_entropy)
X_test, y_test = _prep_data(test_pmids, pmids_dict, w2v, window_size, model_type, binary_ce=binary_cross_entropy)
elif model_type == 'nn':
X_train, y_train = _prep_data(train_pmids, pmids_dict, w2v, window_size, model_type, binary_ce=binary_cross_entropy)
X_test, y_test = _prep_data(test_pmids, pmids_dict, w2v, window_size, model_type, binary_ce=binary_cross_entropy)
if undersample:
# Undersample the non group tags at random....probably a bad idea...
idx_undersample = numpy.where(y_train[:, 1] == 0)[0]
idx_postive = numpy.where(y_train[:, 1] == 1)[0]
random_negative_sample = numpy.random.choice(idx_undersample, idx_postive.shape[0])
X_train_postive = X_train[idx_postive, :, :, :]
y_train_postive = y_train[idx_postive, :]
X_train_negative = X_train[random_negative_sample, :, :, :]
y_train_negative = y_train[random_negative_sample, :]
X_train = numpy.vstack((X_train_postive, X_train_negative))
y_train = numpy.vstack((y_train_postive, y_train_negative))
print('loaded data...')
if model_type == 'cnn':
model = GroupCNN(window_size=window_size, n_feature_maps=n_feature_maps, k_output=k)
elif model_type == 'nn':
model = GroupNN(window_size=window_size, k=k)
model.train(X_train, y_train, epochs, optim_algo=optimizer, criterion=criterion)
accuracy, f1_score, precision, auc, recall = model.test(X_test, y_test)
print "Accuracy: {}".format(accuracy)
print "F1: {}".format(f1_score)
print "Precision: {}".format(precision)
print "AUC: {}".format(auc)
print "Recall: {}".format(recall)
sys.exit()
def run_crf(w2v, l2, l1, iters, shallow_parse, words_before, words_after, grid_search,tacc, name, transfer_learning=False):
pmids_dict, pmids, abstracts, lbls, vectorizer, groups_map, one_hot, dicts = \
parse_summerscales.get_tokens_and_lbls(
make_pmids_dict=True, sen=True, use_genia=shallow_parse, using_tacc=tacc)
model = pycrfsuite.Trainer(verbose=False)
all_pmids = pmids_dict.keys()
n = len(all_pmids)
n_folds = 5
kf = KFold(n, random_state=1337, shuffle=True, n_folds=n_folds)
fold_gi = []
recall_scores = []
precision_scores = []
f1_scores = []
model_type = 'nn'
binary_cross_entropy = True
for fold_idx, (train, test) in enumerate(kf):
print("on fold %s" % fold_idx)
train_pmids = [all_pmids[pmid_idx] for pmid_idx in train]
test_pmids = [all_pmids[pmid_idx] for pmid_idx in test]
print('loading data...')
if transfer_learning:
nn_model = GroupNN.load_model(model_path='NNModel.hdf5', model_info_path='NNModel.hdf5.p')
window_size = nn_model.model_info['window_size']
train_x, train_y = GroupCNNExperiment._prep_data(train_pmids, pmids_dict, w2v, window_size, model_type,
binary_ce=binary_cross_entropy, crf=True)
test_x, test_y = GroupCNNExperiment._prep_data(test_pmids, pmids_dict, w2v, window_size, model_type,
binary_ce=binary_cross_entropy, crf=True)
train_x = transform_features(nn_model, train_x)
test_x = transform_features(nn_model, test_x)
train_y = _labels_to_str(train_y)
test_y = _labels_to_str(test_y)
else:
train_x, train_y = abstracts2features(pmids_dict, train_pmids, words_before, words_after, w2v, shallow_parse)
test_x, test_y = abstracts2features(pmids_dict, test_pmids, words_before, words_after, w2v, shallow_parse)
print('loaded data...')
for x, y in zip(train_x, train_y):
model.append(x, y)
if grid_search:
model.set_params({
'c1': l1, # coefficient for L1 penalty
'c2': l2, # coefficient for L2 penalty
'max_iterations': iters, # stop earlier
# include transitions that are possible, but not observed
'feature.possible_transitions': True
})
crf = sklearn_crfsuite.CRF(
algorithm='lbfgs',
c1=l1,
c2=l2,
max_iterations=iters,
all_possible_transitions=False
)
params_space = {
'c1': scipy.stats.expon(scale=0.5),
'c2': scipy.stats.expon(scale=0.05),
}
# use the same metric for evaluation
f1_scorer = make_scorer(metrics.flat_f1_score,
average='weighted', labels=test_y)
# search
rs = RandomizedSearchCV(crf, params_space,
cv=3,
verbose=1,
n_jobs=-1,
n_iter=50,
scoring=f1_scorer)
rs.fit(train_x, train_y)
info = rs.best_estimator_.tagger_.info()
tagger = rs.best_estimator_.tagger_
else:
model.set_params({
'c1': l1, # coefficient for L1 penalty
'c2': l2, # coefficient for L2 penalty
'max_iterations': iters, # stop earlier
# include transitions that are possible, but not observed
'feature.possible_transitions': True
})
model_name = name + '_model {}'.format(fold_idx)
print('training model...')
model.train(model_name)
print('done...')
tagger = pycrfsuite.Tagger()
tagger.open(model_name)
info = tagger.info()
def print_transitions(trans_features):
for (label_from, label_to), weight in trans_features:
print("%-6s -> %-7s %0.6f" % (label_from, label_to, weight))
print("Top likely transitions:")
print_transitions(Counter(info.transitions).most_common(80))
print("\nTop unlikely transitions:")
print_transitions(Counter(info.transitions).most_common()[-80:])
def print_state_features(state_features):
for (attr, label), weight in state_features:
print("%0.6f %-6s %s" % (weight, label, attr))
print("Top positive:")
print_state_features(Counter(info.state_features).most_common(80))
print("\nTop negative:")
print_state_features(Counter(info.state_features).most_common()[-80:])
g_i = []
y_truths, predictions = [], []
abstract_predicted_mentions, true_abstract_mentions = [], []
for i, (pmid, x, y) in enumerate(zip(test_pmids, test_x, test_y)):
print(pmid)
abstract_words, abstract_labels, tagged_abstract, groups_dict, groups = pmids_dict[pmid]
vocab = groups_map[pmid]
abstract_words, _, _, _, _ = pmids_dict[pmid]
count = 0
pred_labels = tagger.tag(x)
pred_mentions = output2words(pred_labels, abstract_words)
true_mentions = output2words(abstract_labels, abstract_words)
print "Predicted: {}".format(pred_mentions)
print "True: {}".format(true_mentions)
print '\n'
"""
abstract_predicted_words.append(pred_words)
# vocab2 = output2words(x, y, vectorizer, w2v, abstract_words)
if len(test_y) > 0:
y_truths.append(y)
predictions.append(pred_labels)
for pred_word in pred_words:
for v_word in vocab:
if pred_word == v_word:
count += 1
pred_words_string = ''
true_words = " "
for word in pred_words:
pred_words_string = pred_words_string + " " + word
for word in vocab:
true_words = true_words + " " + word
print("Predicted: {}".format(pred_words_string))
print("True: {}".format(true_words))
print ""
if len(vocab) == 0:
continue
g_i1 = float(count)/float(len(vocab))
g_i.append(g_i1)
print LSTM_extraction._crf_evaluate_detection(y, pred_labels, )
"""
abstract_predicted_mentions.append(pred_mentions)
true_abstract_mentions.append(true_mentions)
fold_recall, fold_precision, fold_f1_score = eveluate(abstract_predicted_mentions, true_abstract_mentions)
recall_scores.append(fold_recall)
precision_scores.append(fold_precision)
f1_scores.append(fold_f1_score)
fold_recall_results = "Fold recall: {}".format(fold_recall)
fold_precision_results = "Fold precision: {}".format(fold_precision)
fold_f1_results = "Fold F1 Score: {}".format(fold_f1_score)
print fold_recall_results
print fold_precision_results
print fold_f1_results
file = open(model_name + '_results.txt', 'w+')
file.write(fold_recall_results + '\n')
file.write(fold_precision_results + '\n')
file.write(fold_f1_results + '\n')
# avg_g_i = float(0)
"""
for x in g_i:
avg_g_i += x
avg_g_i = avg_g_i/len(g_i)
fold_gi.append(avg_g_i)
print('g_i: {}'.format(avg_g_i))
"""
"""
recall, precision, tp_overlapping_tokens, fp_tokens, accuracy = LSTM_extraction._crf_evaluate_detection(y_truths, predictions, abstract, vectorizer)
print('Recall: {}'.format(recall))
print('precision: {}'.format(precision))
print('')
avg_g_i = 0
for g in fold_gi:
avg_g_i+=g
avg_g_i = float(avg_g_i)/float(len(fold_gi))
avg_recall = 0
for g in fold_gi:
avg_recall+=g
avg_recall = float(avg_recall)/float(len(recalls))
avg_precision = 0
for g in precisions:
avg_precision+=g
avg_precision = float(avg_precision)/float(len(precisions))
"
print("Fold avg g_i: {}".format(avg_g_i))
"""
recall_average = _compute_average(recall_scores)
precision_average = _compute_average(precision_scores)
f1_scores = _compute_average(f1_scores)
print "Recall Average: {}".format(recall_average)
print "Precision Average: {}".format(precision_average)
print "F1 Average: {}".format(f1_scores)
def main():
n_folds = 5
try:
opts, args = getopt.getopt(sys.argv[1:], '', ['window_size=', 'wiki=', 'n_feature_maps=', 'epochs=',
'undersample=', 'n_feature_maps=', 'criterion=',
'optimizer=', 'model=', 'genia=', 'tacc=', 'layers=',
'hyperopt=', 'model_name='])
except getopt.GetoptError as error:
print error
sys.exit(2)
model_type = 'nn'
window_size = 5
wiki = True
n_feature_maps = 100
epochs = 20
undersample = False
binary_cross_entropy = False
criterion = 'categorical_crossentropy'
optimizer = 'adam'
k = 2
use_genia = False
using_tacc = False
layer_sizes = []
hyperopt = False
model_name = 'model'
for opt, arg in opts:
if opt == '--window_size':
window_size = int(arg)
elif opt == '--wiki':
if arg == 0:
wiki = False
elif opt == '--epochs':
epochs = int(arg)
elif opt == '--layers':
layer_sizes = arg.split(',')
elif opt == '--n_feature_maps':
n_feature_maps = int(arg)
elif opt == '--undersample':
option = int(arg)
if option == 1:
undersample = True
elif opt == '--n_feature_maps':
n_feature_maps = int(arg)
elif opt == '--criterion':
criterion = arg
elif opt == '--optimizer':
optimizer = arg
elif opt == '--model':
model_type = arg
elif opt == '--genia':
if int(arg) == 1:
use_genia= True
elif opt == '--tacc':
if int(arg) == 1:
using_tacc = True
elif opt == '--hyperopt':
if int(arg) == 1:
hyperopt = True
elif opt == '--model_name':
model_name = arg
else:
print "Option {} is not valid!".format(opt)
if criterion == 'binary_crossentropy':
binary_cross_entropy = True
k = 1
print('Loading word2vec model...')
if wiki:
print 'Using wiki word2vec...'
word2vec_model = 'wikipedia-pubmed-and-PMC-w2v.bin'
else:
print 'Using non-wiki word2vec...'
word2vec_model = 'PubMed-w2v.bin'
w2v = Word2Vec.load_word2vec_format(word2vec_model, binary=True)
print('Loaded word2vec model')
pmids_dict, pmids, abstracts, lbls, vectorizer, groups_map, one_hot, dicts = \
parse_summerscales.get_tokens_and_lbls(
make_pmids_dict=True, sen=True, use_genia=use_genia, using_tacc=using_tacc)
all_pmids = pmids_dict.keys()
n = len(all_pmids)
kf = KFold(n, random_state=1337, shuffle=True, n_folds=n_folds)
accuracies = []
recalls = []
precisions = []
f1_scores = []
aucs = []
global model
for fold_idx, (train, test) in enumerate(kf):
print("on fold %s" % fold_idx)
train_pmids = [all_pmids[pmid_idx] for pmid_idx in train]
test_pmids = [all_pmids[pmid_idx] for pmid_idx in test]
print train_pmids
print('loading data...')
if model_type == 'cnn':
X_train, y_train = _prep_data(train_pmids, pmids_dict, w2v, window_size, model_type, binary_ce=binary_cross_entropy)
X_test, y_test = _prep_data(test_pmids, pmids_dict, w2v, window_size, model_type, binary_ce=binary_cross_entropy)
elif model_type == 'nn':
X_train, y_train = _prep_data(train_pmids, pmids_dict, w2v, window_size, model_type, binary_ce=binary_cross_entropy)
X_test, y_test = _prep_data(test_pmids, pmids_dict, w2v, window_size, model_type, binary_ce=binary_cross_entropy)
elif model_type == 'ladder':
X_train, y_train = _prep_data(train_pmids, pmids_dict, w2v, window_size, model_type, binary_ce=binary_cross_entropy)
X_test, y_test = _prep_data(test_pmids, pmids_dict, w2v, window_size, model_type, binary_ce=binary_cross_entropy)
if undersample:
# Undersample the non group tags at random....probably a bad idea...
if binary_cross_entropy:
idx_undersample = numpy.where(y_train == 0)[0]
idx_postive = numpy.where(y_train == 1)[0]
else:
idx_undersample = numpy.where(y_train[:, 1] == 0)[0]
idx_postive = numpy.where(y_train[:, 1] == 1)[0]
random_negative_sample = numpy.random.choice(idx_undersample, idx_postive.shape[0])
if model_type == 'nn':
X_train_postive = X_train[idx_postive, :]
X_train_negative = X_train[random_negative_sample, :]
else:
X_train_postive = X_train[idx_postive, :, :, :]
X_train_negative = X_train[random_negative_sample, :, :, :]
if binary_cross_entropy:
y_train_postive = y_train[idx_postive]
y_train_negative = y_train[random_negative_sample]
else:
y_train_postive = y_train[idx_postive, :]
y_train_negative = y_train[random_negative_sample, :]
X_train = numpy.vstack((X_train_postive, X_train_negative))
if binary_cross_entropy:
y_train = numpy.hstack((y_train_postive, y_train_negative))
else:
y_train = numpy.vstack((y_train_postive, y_train_negative))
print('loaded data...')
if model_type == 'cnn':
model = GroupCNN(window_size=window_size, n_feature_maps=n_feature_maps, k_output=k, name=model_name)
elif model_type == 'nn':
model = GroupNN(window_size=window_size, k=k, hyperparameter_search=hyperopt, name=model_name)
if hyperopt:
best_run, best_model = optim.minimize(model=_model,
data=_data,
algo=tpe.suggest,
max_evals=5,
trials=Trials())
model.model = best_model
else:
model.train(X_train, y_train, epochs, optim_algo=optimizer, criterion=criterion)
words = []
for pmid in test_pmids:
words.extend(pmids_dict[pmid][0])
predictions = model.predict_classes(X_test)
predicted_words = crf.output2words(predictions, words)
y_test_arg_max = numpy.argmax(y_test, axis=1)
true_words = crf.output2words(y_test_arg_max, words)
accuracy, f1_score, precision, auc, recall = model.test(X_test, y_test)
recall, precision, f1_score = crf.eveluate(predicted_words, true_words)
print "Accuracy: {}".format(accuracy)
print "F1: {}".format(f1_score)
print "Precision: {}".format(precision)
print "AUC: {}".format(auc)
print "Recall: {}".format(recall)
accuracies.append(accuracy)
f1_scores.append(f1_score)
precisions.append(precision)
aucs.append(auc)
recalls.append(recall)
mean_accuracy = numpy.mean(accuracies)
mean_f1_score = numpy.mean(f1_scores)
mean_precision = numpy.mean(precisions)
mean_auc_score = numpy.mean(aucs)
mean_recall = numpy.mean(recalls)
mean_accuracy_string = "Mean Accuracy: {}".format(mean_accuracy)
mean_f1_score_string = "Mean F1: {}".format(mean_f1_score)
mean_precision_string = "Mean Precision: {}".format(mean_precision)
mean_auc_score_string = "Mean AUC: {}".format(mean_auc_score)
mean_recall_string = "Mean Recall: {}".format(mean_recall)
print mean_accuracy_string
print mean_f1_score_string
print mean_precision_string
print mean_auc_score_string
print mean_recall_string
results = open('{}_fold_results'.format(model.model_name), 'w+')
results.write(mean_accuracy_string)
results.write(mean_f1_score_string)
results.write(mean_precision_string)
results.write(mean_auc_score_string)
results.write(mean_recall_string)
def get_X_y(wv, wv_dim, vectorizer=None, distant=False, n=None):
pmids, sentences, lbls, vectorizer = [None] * 4
if distant:
#pmids, sentences, lbls, vectorizer = distant_intervention_tag.get_tokens_and_lbls(N=N)
pmids, tagged_abstracts, tokens_and_lbls, intervention_texts = \
distant_intervention_tag.distantly_annotate(n=n)
else:
pmids, sentences, lbls, vectorizer = parse_summerscales.get_tokens_and_lbls(
)
#pdb.set_trace()
# see: https://github.com/fchollet/keras/issues/233
# num_sentences x 1 x max_token_len x wv_dim
# number of sequences x 1 x max number of tokens (padded to max len) x word vector size
num_sentences = len(sentences)
#max_token_len = max([len(s) for s in sentences])
#X_embedded = np.zeros((num_sentences, wv_dim))
X_embedded, X_tokens = [], [
] # here a sequence associated with each doc/abstract
y = []
#X_tokens = []
cur_pmid = pmids[0]
cur_x_embedded, cur_x_tokens, cur_y, token_pmid_list = [], [], [], []
unknown_words_to_vecs = {}
for idx, s in enumerate(sentences):
if cur_pmid != pmids[idx]:
X_embedded.append(np.vstack(cur_x_embedded))
X_tokens.append(np.vstack(cur_x_tokens))
y.append(np.array(cur_y))
cur_x_embedded, cur_x_tokens, cur_y = [], [], []
cur_pmid = pmids[idx]
for j, t in enumerate(s):
try:
v = wv[t]
except:
print("%s not known!" % t)
# or maybe use 0s???
if not t in unknown_words_to_vecs:
v = np.random.uniform(-1, 1, wv_dim)
unknown_words_to_vecs[t] = v
v = unknown_words_to_vecs[t]
cur_x_embedded.append(v)
cur_x_tokens.append(vectorizer.vocabulary_[t])
token_pmid_list.append(cur_pmid)
cur_y.extend(lbls[idx])
X_embedded.append(np.vstack(cur_x_embedded))
X_tokens.append(np.vstack(cur_x_tokens))
y.append(np.array(cur_y))
X_embedded = np.vstack(X_embedded)
X_tokens = np.vstack(X_tokens)
y = np.hstack(y)
return X_embedded, X_tokens, y, vectorizer, unknown_words_to_vecs, token_pmid_list
def run_crf(w2v, words_before, words_after, shallow_parse):
pmids_dict, pmids, abstracts, lbls, vectorizer, groups_map, one_hot, dicts = \
parse_summerscales.get_tokens_and_lbls(
make_pmids_dict=True, sen=True)
"""
Create model
"""
model = ChainCRF(directed=False)
ssvm = FrankWolfeSSVM(model=model, C=.1, max_iter=30)
all_pmids = pmids_dict.keys()
n = len(all_pmids)
n_folds = 5
kf = KFold(n, random_state=1337, shuffle=True, n_folds=n_folds)
fold_gi = []
for fold_idx, (train, test) in enumerate(kf):
print("on fold %s" % fold_idx)
train_pmids = [all_pmids[pmid_idx] for pmid_idx in train]
test_pmids = [all_pmids[pmid_idx] for pmid_idx in test]
print('loading data...')
train_x, train_y = abstract2features(pmids_dict, words_before, w2v, shallow_parse)
test_x, test_y = abstract2features(pmids_dict, words_after, w2v, shallow_parse)
print('loaded data...')
print 'training...'
ssvm.fit(train_x, train_y)
print ssvm.score(test_x, test_y)
for i, (pmid, x, y) in enumerate(zip(test_pmids, test_x, test_y)):
abstract_words, _, _= pmids_dict[pmid]
print(pmid)
# predict() takes in a list returns another list
prediction = ssvm.predict([x]).pop(0)
predicted = ''
output = ''
if len(prediction) > 0:
for p in prediction:
if p == 1:
print "word: {}".format(abstract_words[p])
if n == 0:
predicted += abstract_words[p]
else:
predicted += ' ' + abstract_words[p]
if not predicted == '':
output = 'predicted: {}'.format(predicted)
else:
output = 'Predicted nothing!'
else:
output = 'Predicted nothing!'
print output
def get_PMIDs_to_X_y(wv, wv_dim, max_length=None, distant=False, n=200):
unknown_words_to_vecs = {}
tokens_DS = None
if distant:
tokens_and_lbls, X_DS_embedded, y_DS, tokens_DS, unknown_words_to_vecs = \
_get_distantly_lbled_tokens(n=n, wv=wv, wv_dim=wv_dim)
# we pass tokens_DS -- the unique tokens in the DS
# data -- to go into our vectorizer!
"""
pmids_dict, pmids, sentences, lbls, vectorizer, groups_map = \
parse_summerscales.get_tokens_and_lbls(
make_pmids_dict=True)
"""
pmids_dict, pmids, abstracts, lbls, vectorizer, groups_map, one_hot, dicts = \
parse_summerscales.get_tokens_and_lbls(
make_pmids_dict=True, sen=True)
###
# now loop through and get X_tokens representation!
if distant:
# really token_indices maybe more correct
X_DS_tokens = []
for abs_idx, abs_tokens_and_lbls in enumerate(tokens_and_lbls):
for token_idx, token_and_lbl in enumerate(abs_tokens_and_lbls):
t, lbl = token_and_lbl
X_DS_tokens.append(vectorizer.vocabulary_[t])
# see: https://github.com/fchollet/keras/issues/233
# num_sentences x 1 x max_token_len x wv_dim
# number of sequences x 1 x max number of tokens (padded to max len) x word vector size
# num_sentences = len(sentences)
#max_token_len = max([len(s) for s in sentences])
#X_embedded = np.zeros((num_sentences, wv_dim))
X_embedded, X_tokens = [], [
] # here a sequence associated with each doc/abstract
#unknown_words_to_vecs = {}
pmids_to_X_y = {}
for pmid in pmids_dict:
# pmid_sentences, pmid_lbls = pmids_dict[pmid]
abstract_tokens, abstract_output_labels, _ = pmids_dict[pmid]
# for this sentence
X_embedded = []
X_tokens = []
y = []
for w_i, word_token in enumerate(abstract_tokens):
try:
v = wv[word_token]
except:
# or maybe use 0s???
if word_token not in unknown_words_to_vecs:
print("word '%s' not known!" % word_token)
v = np.random.uniform(-1, 1, wv_dim)
unknown_words_to_vecs[word_token] = v
v = unknown_words_to_vecs[word_token]
X_embedded.append(v)
X_tokens.append(vectorizer.vocabulary_[word_token])
#pmids_to_X_y[pmid] = (np.vstack(X_embedded), np.vstack(X_tokens), np.hstack(y))
if len(abstract_output_labels) > max_length:
abstract_output_labels = abstract_output_labels[:max_length]
elif len(abstract_output_labels) < max_length:
padding = []
for i in range(max_length - len(abstract_output_labels)):
padding.append(0)
abstract_output_labels = padding + abstract_output_labels
assert len(abstract_output_labels) == max_length, 'Must be same size'
pmids_to_X_y[pmid] = (X_embedded, X_tokens, abstract_output_labels)
"""
for sent_idx, s in enumerate(pmid_sentences):
for j, t in enumerate(s):
try:
v = wv[t]
except:
# or maybe use 0s???
if not t in unknown_words_to_vecs:
print("word '%s' not known!" % t)
v = np.random.uniform(-1,1,wv_dim)
unknown_words_to_vecs[t] = v
v = unknown_words_to_vecs[t]
X_embedded.append(v)
X_tokens.append(vectorizer.vocabulary_[t])
y.extend(pmid_lbls[sent_idx])
pmids_to_X_y[pmid] = (np.vstack(X_embedded), np.vstack(X_tokens), np.hstack(y))
"""
if distant:
return pmids_to_X_y, vectorizer, unknown_words_to_vecs, X_DS_embedded, X_DS_tokens, y_DS
return pmids_to_X_y, vectorizer, unknown_words_to_vecs, groups_map, pmids_dict
def run_crf(w2v, words_before, words_after, shallow_parse):
pmids_dict, pmids, abstracts, lbls, vectorizer, groups_map, one_hot, dicts = \
parse_summerscales.get_tokens_and_lbls(
make_pmids_dict=True, sen=True)
"""
Create model
"""
model = ChainCRF(directed=False)
ssvm = FrankWolfeSSVM(model=model, C=.1, max_iter=30)
all_pmids = pmids_dict.keys()
n = len(all_pmids)
n_folds = 5
kf = KFold(n, random_state=1337, shuffle=True, n_folds=n_folds)
fold_gi = []
for fold_idx, (train, test) in enumerate(kf):
print("on fold %s" % fold_idx)
train_pmids = [all_pmids[pmid_idx] for pmid_idx in train]
test_pmids = [all_pmids[pmid_idx] for pmid_idx in test]
print('loading data...')
train_x, train_y = abstract2features(pmids_dict, words_before, w2v,
shallow_parse)
test_x, test_y = abstract2features(pmids_dict, words_after, w2v,
shallow_parse)
print('loaded data...')
print 'training...'
ssvm.fit(train_x, train_y)
print ssvm.score(test_x, test_y)
for i, (pmid, x, y) in enumerate(zip(test_pmids, test_x, test_y)):
abstract_words, _, _ = pmids_dict[pmid]
print(pmid)
# predict() takes in a list returns another list
prediction = ssvm.predict([x]).pop(0)
predicted = ''
output = ''
if len(prediction) > 0:
for p in prediction:
if p == 1:
print "word: {}".format(abstract_words[p])
if n == 0:
predicted += abstract_words[p]
else:
predicted += ' ' + abstract_words[p]
if not predicted == '':
output = 'predicted: {}'.format(predicted)
else:
output = 'Predicted nothing!'
else:
output = 'Predicted nothing!'
print output
def main():
n_folds = 5
try:
opts, args = getopt.getopt(sys.argv[1:], '', [
'window_size=', 'wiki=', 'n_feature_maps=', 'epochs=',
'undersample=', 'n_feature_maps=', 'criterion=', 'optimizer=',
'model=', 'genia=', 'tacc=', 'layers=', 'hyperopt=', 'model_name='
])
except getopt.GetoptError as error:
print error
sys.exit(2)
model_type = 'nn'
window_size = 5
wiki = True
n_feature_maps = 100
epochs = 20
undersample = False
binary_cross_entropy = False
criterion = 'categorical_crossentropy'
optimizer = 'adam'
k = 2
use_genia = False
using_tacc = False
layer_sizes = []
hyperopt = False
model_name = 'model'
for opt, arg in opts:
if opt == '--window_size':
window_size = int(arg)
elif opt == '--wiki':
if arg == 0:
wiki = False
elif opt == '--epochs':
epochs = int(arg)
elif opt == '--layers':
layer_sizes = arg.split(',')
elif opt == '--n_feature_maps':
n_feature_maps = int(arg)
elif opt == '--undersample':
option = int(arg)
if option == 1:
undersample = True
elif opt == '--n_feature_maps':
n_feature_maps = int(arg)
elif opt == '--criterion':
criterion = arg
elif opt == '--optimizer':
optimizer = arg
elif opt == '--model':
model_type = arg
elif opt == '--genia':
if int(arg) == 1:
use_genia = True
elif opt == '--tacc':
if int(arg) == 1:
using_tacc = True
elif opt == '--hyperopt':
if int(arg) == 1:
hyperopt = True
elif opt == '--model_name':
model_name = arg
else:
print "Option {} is not valid!".format(opt)
if criterion == 'binary_crossentropy':
binary_cross_entropy = True
k = 1
print('Loading word2vec model...')
if wiki:
print 'Using wiki word2vec...'
word2vec_model = 'wikipedia-pubmed-and-PMC-w2v.bin'
else:
print 'Using non-wiki word2vec...'
word2vec_model = 'PubMed-w2v.bin'
w2v = Word2Vec.load_word2vec_format(word2vec_model, binary=True)
print('Loaded word2vec model')
pmids_dict, pmids, abstracts, lbls, vectorizer, groups_map, one_hot, dicts = \
parse_summerscales.get_tokens_and_lbls(
make_pmids_dict=True, sen=True, use_genia=use_genia, using_tacc=using_tacc)
all_pmids = pmids_dict.keys()
n = len(all_pmids)
kf = KFold(n, random_state=1337, shuffle=True, n_folds=n_folds)
accuracies = []
recalls = []
precisions = []
f1_scores = []
aucs = []
global model
for fold_idx, (train, test) in enumerate(kf):
print("on fold %s" % fold_idx)
train_pmids = [all_pmids[pmid_idx] for pmid_idx in train]
test_pmids = [all_pmids[pmid_idx] for pmid_idx in test]
print train_pmids
print('loading data...')
if model_type == 'cnn':
X_train, y_train = _prep_data(train_pmids,
pmids_dict,
w2v,
window_size,
model_type,
binary_ce=binary_cross_entropy)
X_test, y_test = _prep_data(test_pmids,
pmids_dict,
w2v,
window_size,
model_type,
binary_ce=binary_cross_entropy)
elif model_type == 'nn':
X_train, y_train = _prep_data(train_pmids,
pmids_dict,
w2v,
window_size,
model_type,
binary_ce=binary_cross_entropy)
X_test, y_test = _prep_data(test_pmids,
pmids_dict,
w2v,
window_size,
model_type,
binary_ce=binary_cross_entropy)
elif model_type == 'ladder':
X_train, y_train = _prep_data(train_pmids,
pmids_dict,
w2v,
window_size,
model_type,
binary_ce=binary_cross_entropy)
X_test, y_test = _prep_data(test_pmids,
pmids_dict,
w2v,
window_size,
model_type,
binary_ce=binary_cross_entropy)
if undersample:
# Undersample the non group tags at random....probably a bad idea...
if binary_cross_entropy:
idx_undersample = numpy.where(y_train == 0)[0]
idx_postive = numpy.where(y_train == 1)[0]
else:
idx_undersample = numpy.where(y_train[:, 1] == 0)[0]
idx_postive = numpy.where(y_train[:, 1] == 1)[0]
random_negative_sample = numpy.random.choice(
idx_undersample, idx_postive.shape[0])
if model_type == 'nn':
X_train_postive = X_train[idx_postive, :]
X_train_negative = X_train[random_negative_sample, :]
else:
X_train_postive = X_train[idx_postive, :, :, :]
X_train_negative = X_train[random_negative_sample, :, :, :]
if binary_cross_entropy:
y_train_postive = y_train[idx_postive]
y_train_negative = y_train[random_negative_sample]
else:
y_train_postive = y_train[idx_postive, :]
y_train_negative = y_train[random_negative_sample, :]
X_train = numpy.vstack((X_train_postive, X_train_negative))
if binary_cross_entropy:
y_train = numpy.hstack((y_train_postive, y_train_negative))
else:
y_train = numpy.vstack((y_train_postive, y_train_negative))
print('loaded data...')
if model_type == 'cnn':
model = GroupCNN(window_size=window_size,
n_feature_maps=n_feature_maps,
k_output=k,
name=model_name)
elif model_type == 'nn':
model = GroupNN(window_size=window_size,
k=k,
hyperparameter_search=hyperopt,
name=model_name)
if hyperopt:
best_run, best_model = optim.minimize(model=_model,
data=_data,
algo=tpe.suggest,
max_evals=5,
trials=Trials())
model.model = best_model
else:
model.train(X_train,
y_train,
epochs,
optim_algo=optimizer,
criterion=criterion)
words = []
for pmid in test_pmids:
words.extend(pmids_dict[pmid][0])
predictions = model.predict_classes(X_test)
predicted_words = crf.output2words(predictions, words)
y_test_arg_max = numpy.argmax(y_test, axis=1)
true_words = crf.output2words(y_test_arg_max, words)
accuracy, f1_score, precision, auc, recall = model.test(X_test, y_test)
recall, precision, f1_score = crf.eveluate(predicted_words, true_words)
print "Accuracy: {}".format(accuracy)
print "F1: {}".format(f1_score)
print "Precision: {}".format(precision)
print "AUC: {}".format(auc)
print "Recall: {}".format(recall)
accuracies.append(accuracy)
f1_scores.append(f1_score)
precisions.append(precision)
aucs.append(auc)
recalls.append(recall)
mean_accuracy = numpy.mean(accuracies)
mean_f1_score = numpy.mean(f1_scores)
mean_precision = numpy.mean(precisions)
mean_auc_score = numpy.mean(aucs)
mean_recall = numpy.mean(recalls)
mean_accuracy_string = "Mean Accuracy: {}".format(mean_accuracy)
mean_f1_score_string = "Mean F1: {}".format(mean_f1_score)
mean_precision_string = "Mean Precision: {}".format(mean_precision)
mean_auc_score_string = "Mean AUC: {}".format(mean_auc_score)
mean_recall_string = "Mean Recall: {}".format(mean_recall)
print mean_accuracy_string
print mean_f1_score_string
print mean_precision_string
print mean_auc_score_string
print mean_recall_string
results = open('{}_fold_results'.format(model.model_name), 'w+')
results.write(mean_accuracy_string)
results.write(mean_f1_score_string)
results.write(mean_precision_string)
results.write(mean_auc_score_string)
results.write(mean_recall_string)
def run_crf(w2v,
l2,
l1,
iters,
shallow_parse,
words_before,
words_after,
grid_search,
tacc,
name,
transfer_learning=False):
pmids_dict, pmids, abstracts, lbls, vectorizer, groups_map, one_hot, dicts = \
parse_summerscales.get_tokens_and_lbls(
make_pmids_dict=True, sen=True, use_genia=shallow_parse, using_tacc=tacc)
model = pycrfsuite.Trainer(verbose=False)
all_pmids = pmids_dict.keys()
n = len(all_pmids)
n_folds = 5
kf = KFold(n, random_state=1337, shuffle=True, n_folds=n_folds)
fold_gi = []
recall_scores = []
precision_scores = []
f1_scores = []
model_type = 'nn'
binary_cross_entropy = True
for fold_idx, (train, test) in enumerate(kf):
print("on fold %s" % fold_idx)
train_pmids = [all_pmids[pmid_idx] for pmid_idx in train]
test_pmids = [all_pmids[pmid_idx] for pmid_idx in test]
print('loading data...')
if transfer_learning:
nn_model = GroupNN.load_model(model_path='NNModel.hdf5',
model_info_path='NNModel.hdf5.p')
window_size = nn_model.model_info['window_size']
train_x, train_y = GroupCNNExperiment._prep_data(
train_pmids,
pmids_dict,
w2v,
window_size,
model_type,
binary_ce=binary_cross_entropy,
crf=True)
test_x, test_y = GroupCNNExperiment._prep_data(
test_pmids,
pmids_dict,
w2v,
window_size,
model_type,
binary_ce=binary_cross_entropy,
crf=True)
train_x = transform_features(nn_model, train_x)
test_x = transform_features(nn_model, test_x)
train_y = _labels_to_str(train_y)
test_y = _labels_to_str(test_y)
else:
train_x, train_y = abstracts2features(pmids_dict, train_pmids,
words_before, words_after,
w2v, shallow_parse)
test_x, test_y = abstracts2features(pmids_dict, test_pmids,
words_before, words_after, w2v,
shallow_parse)
print('loaded data...')
for x, y in zip(train_x, train_y):
model.append(x, y)
if grid_search:
model.set_params({
'c1': l1, # coefficient for L1 penalty
'c2': l2, # coefficient for L2 penalty
'max_iterations': iters, # stop earlier
# include transitions that are possible, but not observed
'feature.possible_transitions': True
})
crf = sklearn_crfsuite.CRF(algorithm='lbfgs',
c1=l1,
c2=l2,
max_iterations=iters,
all_possible_transitions=False)
params_space = {
'c1': scipy.stats.expon(scale=0.5),
'c2': scipy.stats.expon(scale=0.05),
}
# use the same metric for evaluation
f1_scorer = make_scorer(metrics.flat_f1_score,
average='weighted',
labels=test_y)
# search
rs = RandomizedSearchCV(crf,
params_space,
cv=3,
verbose=1,
n_jobs=-1,
n_iter=50,
scoring=f1_scorer)
rs.fit(train_x, train_y)
info = rs.best_estimator_.tagger_.info()
tagger = rs.best_estimator_.tagger_
else:
model.set_params({
'c1': l1, # coefficient for L1 penalty
'c2': l2, # coefficient for L2 penalty
'max_iterations': iters, # stop earlier
# include transitions that are possible, but not observed
'feature.possible_transitions': True
})
model_name = name + '_model {}'.format(fold_idx)
print('training model...')
model.train(model_name)
print('done...')
tagger = pycrfsuite.Tagger()
tagger.open(model_name)
info = tagger.info()
def print_transitions(trans_features):
for (label_from, label_to), weight in trans_features:
print("%-6s -> %-7s %0.6f" % (label_from, label_to, weight))
print("Top likely transitions:")
print_transitions(Counter(info.transitions).most_common(80))
print("\nTop unlikely transitions:")
print_transitions(Counter(info.transitions).most_common()[-80:])
def print_state_features(state_features):
for (attr, label), weight in state_features:
print("%0.6f %-6s %s" % (weight, label, attr))
print("Top positive:")
print_state_features(Counter(info.state_features).most_common(80))
print("\nTop negative:")
print_state_features(Counter(info.state_features).most_common()[-80:])
g_i = []
y_truths, predictions = [], []
abstract_predicted_mentions, true_abstract_mentions = [], []
for i, (pmid, x, y) in enumerate(zip(test_pmids, test_x, test_y)):
print(pmid)
abstract_words, abstract_labels, tagged_abstract, groups_dict, groups = pmids_dict[
pmid]
vocab = groups_map[pmid]
abstract_words, _, _, _, _ = pmids_dict[pmid]
count = 0
pred_labels = tagger.tag(x)
pred_mentions = output2words(pred_labels, abstract_words)
true_mentions = output2words(abstract_labels, abstract_words)
print "Predicted: {}".format(pred_mentions)
print "True: {}".format(true_mentions)
print '\n'
"""
abstract_predicted_words.append(pred_words)
# vocab2 = output2words(x, y, vectorizer, w2v, abstract_words)
if len(test_y) > 0:
y_truths.append(y)
predictions.append(pred_labels)
for pred_word in pred_words:
for v_word in vocab:
if pred_word == v_word:
count += 1
pred_words_string = ''
true_words = " "
for word in pred_words:
pred_words_string = pred_words_string + " " + word
for word in vocab:
true_words = true_words + " " + word
print("Predicted: {}".format(pred_words_string))
print("True: {}".format(true_words))
print ""
if len(vocab) == 0:
continue
g_i1 = float(count)/float(len(vocab))
g_i.append(g_i1)
print LSTM_extraction._crf_evaluate_detection(y, pred_labels, )
"""
abstract_predicted_mentions.append(pred_mentions)
true_abstract_mentions.append(true_mentions)
fold_recall, fold_precision, fold_f1_score = eveluate(
abstract_predicted_mentions, true_abstract_mentions)
recall_scores.append(fold_recall)
precision_scores.append(fold_precision)
f1_scores.append(fold_f1_score)
fold_recall_results = "Fold recall: {}".format(fold_recall)
fold_precision_results = "Fold precision: {}".format(fold_precision)
fold_f1_results = "Fold F1 Score: {}".format(fold_f1_score)
print fold_recall_results
print fold_precision_results
print fold_f1_results
file = open(model_name + '_results.txt', 'w+')
file.write(fold_recall_results + '\n')
file.write(fold_precision_results + '\n')
file.write(fold_f1_results + '\n')
# avg_g_i = float(0)
"""
for x in g_i:
avg_g_i += x
avg_g_i = avg_g_i/len(g_i)
fold_gi.append(avg_g_i)
print('g_i: {}'.format(avg_g_i))
"""
"""
recall, precision, tp_overlapping_tokens, fp_tokens, accuracy = LSTM_extraction._crf_evaluate_detection(y_truths, predictions, abstract, vectorizer)
print('Recall: {}'.format(recall))
print('precision: {}'.format(precision))
print('')
avg_g_i = 0
for g in fold_gi:
avg_g_i+=g
avg_g_i = float(avg_g_i)/float(len(fold_gi))
avg_recall = 0
for g in fold_gi:
avg_recall+=g
avg_recall = float(avg_recall)/float(len(recalls))
avg_precision = 0
for g in precisions:
avg_precision+=g
avg_precision = float(avg_precision)/float(len(precisions))
"
print("Fold avg g_i: {}".format(avg_g_i))
"""
recall_average = _compute_average(recall_scores)
precision_average = _compute_average(precision_scores)
f1_scores = _compute_average(f1_scores)
print "Recall Average: {}".format(recall_average)
print "Precision Average: {}".format(precision_average)
print "F1 Average: {}".format(f1_scores)
def get_X_y(wv, wv_dim, vectorizer=None, distant=False, n=None):
pmids, sentences, lbls, vectorizer = [None]*4
if distant:
#pmids, sentences, lbls, vectorizer = distant_intervention_tag.get_tokens_and_lbls(N=N)
pmids, tagged_abstracts, tokens_and_lbls, intervention_texts = \
distant_intervention_tag.distantly_annotate(n=n)
else:
pmids, sentences, lbls, vectorizer = parse_summerscales.get_tokens_and_lbls()
#pdb.set_trace()
# see: https://github.com/fchollet/keras/issues/233
# num_sentences x 1 x max_token_len x wv_dim
# number of sequences x 1 x max number of tokens (padded to max len) x word vector size
num_sentences = len(sentences)
#max_token_len = max([len(s) for s in sentences])
#X_embedded = np.zeros((num_sentences, wv_dim))
X_embedded, X_tokens = [], [] # here a sequence associated with each doc/abstract
y = []
#X_tokens = []
cur_pmid = pmids[0]
cur_x_embedded, cur_x_tokens, cur_y, token_pmid_list = [], [], [], []
unknown_words_to_vecs = {}
for idx, s in enumerate(sentences):
if cur_pmid != pmids[idx]:
X_embedded.append(np.vstack(cur_x_embedded))
X_tokens.append(np.vstack(cur_x_tokens))
y.append(np.array(cur_y))
cur_x_embedded, cur_x_tokens, cur_y = [], [], []
cur_pmid = pmids[idx]
for j, t in enumerate(s):
try:
v = wv[t]
except:
print("%s not known!" % t)
# or maybe use 0s???
if not t in unknown_words_to_vecs:
v = np.random.uniform(-1,1,wv_dim)
unknown_words_to_vecs[t] = v
v = unknown_words_to_vecs[t]
cur_x_embedded.append(v)
cur_x_tokens.append(vectorizer.vocabulary_[t])
token_pmid_list.append(cur_pmid)
cur_y.extend(lbls[idx])
X_embedded.append(np.vstack(cur_x_embedded))
X_tokens.append(np.vstack(cur_x_tokens))
y.append(np.array(cur_y))
X_embedded = np.vstack(X_embedded)
X_tokens = np.vstack(X_tokens)
y = np.hstack(y)
return X_embedded, X_tokens, y, vectorizer, unknown_words_to_vecs, token_pmid_list
def get_PMIDs_to_X_y(wv, wv_dim, max_length=None, distant=False, n=200):
unknown_words_to_vecs = {}
tokens_DS = None
if distant:
tokens_and_lbls, X_DS_embedded, y_DS, tokens_DS, unknown_words_to_vecs = \
_get_distantly_lbled_tokens(n=n, wv=wv, wv_dim=wv_dim)
# we pass tokens_DS -- the unique tokens in the DS
# data -- to go into our vectorizer!
"""
pmids_dict, pmids, sentences, lbls, vectorizer, groups_map = \
parse_summerscales.get_tokens_and_lbls(
make_pmids_dict=True)
"""
pmids_dict, pmids, abstracts, lbls, vectorizer, groups_map, one_hot, dicts = \
parse_summerscales.get_tokens_and_lbls(
make_pmids_dict=True, sen=True)
###
# now loop through and get X_tokens representation!
if distant:
# really token_indices maybe more correct
X_DS_tokens = []
for abs_idx, abs_tokens_and_lbls in enumerate(tokens_and_lbls):
for token_idx, token_and_lbl in enumerate(abs_tokens_and_lbls):
t, lbl = token_and_lbl
X_DS_tokens.append(vectorizer.vocabulary_[t])
# see: https://github.com/fchollet/keras/issues/233
# num_sentences x 1 x max_token_len x wv_dim
# number of sequences x 1 x max number of tokens (padded to max len) x word vector size
# num_sentences = len(sentences)
#max_token_len = max([len(s) for s in sentences])
#X_embedded = np.zeros((num_sentences, wv_dim))
X_embedded, X_tokens = [], [] # here a sequence associated with each doc/abstract
#unknown_words_to_vecs = {}
pmids_to_X_y = {}
for pmid in pmids_dict:
# pmid_sentences, pmid_lbls = pmids_dict[pmid]
abstract_tokens, abstract_output_labels, _ = pmids_dict[pmid]
# for this sentence
X_embedded = []
X_tokens = []
y = []
for w_i, word_token in enumerate(abstract_tokens):
try:
v = wv[word_token]
except:
# or maybe use 0s???
if word_token not in unknown_words_to_vecs:
print("word '%s' not known!" % word_token)
v = np.random.uniform(-1,1,wv_dim)
unknown_words_to_vecs[word_token] = v
v = unknown_words_to_vecs[word_token]
X_embedded.append(v)
X_tokens.append(vectorizer.vocabulary_[word_token])
#pmids_to_X_y[pmid] = (np.vstack(X_embedded), np.vstack(X_tokens), np.hstack(y))
if len(abstract_output_labels) > max_length:
abstract_output_labels = abstract_output_labels[:max_length]
elif len(abstract_output_labels) < max_length:
padding = []
for i in range(max_length - len(abstract_output_labels)):
padding.append(0)
abstract_output_labels = padding + abstract_output_labels
assert len(abstract_output_labels) == max_length, 'Must be same size'
pmids_to_X_y[pmid] = (X_embedded, X_tokens, abstract_output_labels)
"""
for sent_idx, s in enumerate(pmid_sentences):
for j, t in enumerate(s):
try:
v = wv[t]
except:
# or maybe use 0s???
if not t in unknown_words_to_vecs:
print("word '%s' not known!" % t)
v = np.random.uniform(-1,1,wv_dim)
unknown_words_to_vecs[t] = v
v = unknown_words_to_vecs[t]
X_embedded.append(v)
X_tokens.append(vectorizer.vocabulary_[t])
y.extend(pmid_lbls[sent_idx])
pmids_to_X_y[pmid] = (np.vstack(X_embedded), np.vstack(X_tokens), np.hstack(y))
"""
if distant:
return pmids_to_X_y, vectorizer, unknown_words_to_vecs, X_DS_embedded, X_DS_tokens, y_DS
return pmids_to_X_y, vectorizer, unknown_words_to_vecs, groups_map, pmids_dict
