def get_sentences(fn, all_examples, word_dict, char_dict, args, examples_size):
sentence_matrix = np.zeros((examples_size, args.max_sent, 160))
for batch_x, mb_idx in all_examples:
batch_x, batch_sent, batch_doc, batch_y = zip(*batch_x)
batch_x = util.vectorization(list(batch_x),
word_dict,
char_dict,
max_char_length=args.max_char)
batch_rnn, batch_sent_mask, batch_word_mask, batch_cnn = \
util.mask_padding(batch_x, args.max_sent, args.max_word, args.max_char)
sentence_array = fn(batch_rnn, batch_cnn, batch_word_mask,
batch_sent_mask)
sentence_matrix[mb_idx, :, :] = sentence_array
return sentence_matrix
def get_attentions(fn, all_examples, word_dict, char_dict, args,
examples_size):
att_matrix = np.zeros((examples_size, args.max_sent))
for batch_x, mb_idx in all_examples:
batch_x, batch_sent, batch_doc, batch_y = zip(*batch_x)
batch_x = util.vectorization(list(batch_x),
word_dict,
char_dict,
max_char_length=args.max_char)
batch_rnn, batch_sent_mask, batch_word_mask, batch_cnn = \
util.mask_padding(batch_x, args.max_sent, args.max_word, args.max_char)
batch_sent = util.sent_ling_padding(list(batch_sent), args.max_sent,
args.max_ling)
att = fn(batch_rnn, batch_cnn, batch_word_mask, batch_sent_mask,
batch_sent)
att_matrix[mb_idx, :] = att
return att_matrix.tolist()
def main(args):
logging.info("loading data...")
fake_train, fake_dev, fake_test = du.load_fake(doc_ling=False,
sent_ling=False)
true_train, true_dev, true_test = du.load_true(doc_ling=False,
sent_ling=False)
if args.debug:
true_train = true_train[0][:100]
fake_train = fake_train[:10]
true_dev = true_dev[:100]
fake_dev = fake_dev[:10]
true_test = true_test[:100]
fake_test = fake_test[:10]
if args.rnn_type == 'gru':
args.rnn = lasagne.layers.GRULayer
elif args.rnn_type == 'lstm':
args.rnn = lasagne.layers.LSTMLayer
else:
args.rnn = lasagne.layers.RecurrentLayer
logging.info("building dictionary...")
word_dict, char_dict = util.build_dict(
None, max_words=0, dict_file=["word_dict", "char_dict"])
logging.info("creating embedding matrix...")
word_embed = util.words2embedding(word_dict, 100, args.embedding_file)
char_embed = util.char2embedding(char_dict, 30)
(args.word_vocab_size, args.word_embed_size) = word_embed.shape
(args.char_vocab_size, args.char_embed_size) = char_embed.shape
logging.info("compiling Theano function...")
att_fn, eval_fn, train_fn, params = create_theano_function(word_embed,
char_embed,
values=None)
logging.info("batching examples...")
dev_examples = mb.vec_minibatch(fake_dev + true_dev,
word_dict,
char_dict,
args,
False,
char=False,
sent_ling=False,
doc_ling=False)
test_examples = mb.vec_minibatch(fake_test + true_test,
word_dict,
char_dict,
args,
False,
char=False,
sent_ling=False,
doc_ling=False)
temp = []
for true_batch in true_train:
temp += true_batch
true_train = temp
del temp
train_examples = mb.doc_minibatch(fake_train + true_train, args.batch_size)
# train_examples = mb.train_doc_minibatch(fake_train, true_train, args)
logging.info("checking network...")
dev_acc = evals.eval_vec_batch(eval_fn,
dev_examples,
char=False,
sent_ling=False,
doc_ling=False)
print('Dev A: %.2f P:%.2f R:%.2f F:%.2f' % dev_acc)
test_acc = evals.eval_vec_batch(eval_fn,
test_examples,
char=False,
sent_ling=False,
doc_ling=False)
print('Performance on Test set: A: %.2f P:%.2f R:%.2f F:%.2f' % test_acc)
prev_fsc = 0
stop_count = 0
best_fsc = 0
best_acc = 0
logging.info("training %d examples" % len(train_examples))
start_time = time.time()
n_updates = 0
for epoch in range(args.epoches):
np.random.shuffle(train_examples)
# if epoch > 3:
# logging.info("compiling Theano function again...")
# args.learning_rate *= 0.9
# att_fn, eval_fn, train_fn, params = create_theano_function(
# word_embed, char_embed, values=[x.get_value() for x in params])
for batch_x, _ in train_examples:
batch_x, batch_y = zip(*batch_x)
batch_x = util.vectorization(list(batch_x),
word_dict,
char_dict,
max_char_length=args.max_char)
batch_rnn, batch_sent_mask, batch_word_mask, _ = \
util.mask_padding(batch_x, args.max_sent, args.max_word, args.max_char)
batch_y = np.array(list(batch_y))
train_loss = train_fn(batch_rnn, batch_word_mask, batch_sent_mask,
batch_y)
n_updates += 1
if n_updates % 100 == 0 and epoch > 7:
logging.info(
'Epoch = %d, loss = %.2f, elapsed time = %.2f (s)' %
(epoch, train_loss, time.time() - start_time))
# dev_acc = evals.eval_batch(eval_fn, dev_examples, word_dict, char_dict, args)
dev_acc = evals.eval_vec_batch(eval_fn,
dev_examples,
char=False,
sent_ling=False,
doc_ling=False)
logging.info('Dev A: %.2f P:%.2f R:%.2f F:%.2f' % dev_acc)
if dev_acc[3] > best_fsc and dev_acc[0] > best_acc:
best_fsc = dev_acc[3]
best_acc = dev_acc[0]
logging.info(
'Best dev f1: epoch = %d, n_udpates = %d, f1 = %.2f %%'
% (epoch, n_updates, dev_acc[3]))
record = 'Best dev accuracy: epoch = %d, n_udpates = %d ' % \
(epoch, n_updates) + ' Dev A: %.2f P:%.2f R:%.2f F:%.2f' % dev_acc
test_acc = evals.eval_vec_batch(eval_fn,
test_examples,
char=False,
sent_ling=False,
doc_ling=False)
print(
'Performance on Test set: A: %.2f P:%.2f R:%.2f F:%.2f'
% test_acc)
# util.save_params('char_not_params_%.2f' % test_acc[3], params,
# epoch=epoch, n_updates=n_updates)
if prev_fsc > dev_acc[3]:
stop_count += 1
else:
stop_count = 0
if stop_count == 6:
print("stopped")
prev_fsc = dev_acc[3]
print(record)
print('Performance on Test set: A: %.2f P:%.2f R:%.2f F:%.2f' % test_acc)
return
def eval_batch(fn,
examples,
word_dict,
char_dict,
args,
char=True,
sent_ling=True,
doc_ling=True):
a = b = c = d = 0
for batch_x, _ in examples:
if char and sent_ling and doc_ling:
batch_x, batch_sent, batch_doc, batch_y = zip(*batch_x)
batch_x = util.vectorization(list(batch_x),
word_dict,
char_dict,
max_char_length=args.max_char)
batch_rnn, batch_sent_mask, batch_word_mask, batch_cnn = \
util.mask_padding(batch_x, args.max_sent, args.max_word, args.max_char)
batch_sent = util.sent_ling_padding(list(batch_sent),
args.max_sent, args.max_ling)
batch_doc = util.doc_ling_padding(list(batch_doc), args.max_ling)
batch_y = np.array(list(batch_y))
predict = fn(batch_rnn, batch_cnn, batch_word_mask,
batch_sent_mask, batch_sent, batch_doc)
elif char and not sent_ling and doc_ling:
batch_x, batch_doc, batch_y = zip(*batch_x)
batch_x = util.vectorization(list(batch_x),
word_dict,
char_dict,
max_char_length=args.max_char)
batch_rnn, batch_sent_mask, batch_word_mask, batch_cnn = \
util.mask_padding(batch_x, args.max_sent, args.max_word, args.max_char)
batch_doc = util.doc_ling_padding(list(batch_doc), args.max_ling)
batch_y = np.array(list(batch_y))
predict = fn(batch_rnn, batch_cnn, batch_word_mask,
batch_sent_mask, batch_doc)
elif char and sent_ling and not doc_ling:
batch_x, batch_sent, batch_y = zip(*batch_x)
batch_x = util.vectorization(list(batch_x),
word_dict,
char_dict,
max_char_length=args.max_char)
batch_rnn, batch_sent_mask, batch_word_mask, batch_cnn = \
util.mask_padding(batch_x, args.max_sent, args.max_word, args.max_char)
batch_sent = util.sent_ling_padding(list(batch_sent),
args.max_sent, args.max_ling)
batch_y = np.array(list(batch_y))
predict = fn(batch_rnn, batch_cnn, batch_word_mask,
batch_sent_mask, batch_sent)
elif char and not sent_ling and not doc_ling:
batch_x, batch_y = zip(*batch_x)
batch_x = util.vectorization(list(batch_x),
word_dict,
char_dict,
max_char_length=args.max_char)
batch_rnn, batch_sent_mask, batch_word_mask, batch_cnn = \
util.mask_padding(batch_x, args.max_sent, args.max_word, args.max_char)
batch_y = np.array(list(batch_y))
predict = fn(batch_rnn, batch_cnn, batch_word_mask,
batch_sent_mask)
elif not char and not sent_ling and not doc_ling:
batch_x, batch_y = zip(*batch_x)
batch_x = util.vectorization(list(batch_x),
word_dict,
char_dict,
max_char_length=args.max_char)
batch_rnn, batch_sent_mask, batch_word_mask, _ = \
util.mask_padding(batch_x, args.max_sent, args.max_word, args.max_char)
batch_y = np.array(list(batch_y))
predict = fn(batch_rnn, batch_word_mask, batch_sent_mask)
elif not char and sent_ling and doc_ling:
batch_x, batch_sent, batch_doc, batch_y = zip(*batch_x)
batch_x = util.vectorization(list(batch_x),
word_dict,
char_dict,
max_char_length=args.max_char)
batch_rnn, batch_sent_mask, batch_word_mask, _ = \
util.mask_padding(batch_x, args.max_sent, args.max_word, args.max_char)
batch_sent = util.sent_ling_padding(list(batch_sent),
args.max_sent, args.max_ling)
batch_doc = util.doc_ling_padding(list(batch_doc), args.max_ling)
batch_y = np.array(list(batch_y))
predict = fn(batch_rnn, batch_word_mask, batch_sent_mask,
batch_sent, batch_doc)
matrix = confusion_matrix(predict, batch_y)
a += matrix[0]
b += matrix[1]
c += matrix[2]
d += matrix[3]
acc = 100.0 * (a + d) / (a + b + c + d)
pre = 100.0 * a / (a + c)
rec = 100.0 * a / (a + b)
fsc = 2 * pre * rec / (pre + rec)
return acc, pre, rec, fsc
def vec_minibatch(docs,
word_dict,
char_dict,
args,
shuffle=True,
char=True,
sent_ling=True,
doc_ling=True):
examples = []
if shuffle:
random.shuffle(docs)
doc_length = len(docs)
id_list = np.arange(0, doc_length, args.batch_size)
if shuffle:
np.random.shuffle(id_list)
mbs = [
np.arange(id, min(id + args.batch_size, doc_length)) for id in id_list
]
for mb in mbs:
batch_x = [docs[i] for i in mb]
if char and sent_ling and doc_ling:
batch_x, batch_sent, batch_doc, batch_y = zip(*batch_x)
batch_x = util.vectorization(list(batch_x),
word_dict,
char_dict,
max_char_length=args.max_char)
batch_rnn, batch_sent_mask, batch_word_mask, batch_cnn = \
util.mask_padding(batch_x, args.max_sent, args.max_word, args.max_char)
batch_sent = util.sent_ling_padding(list(batch_sent),
args.max_sent, args.max_ling)
batch_doc = util.doc_ling_padding(list(batch_doc), args.max_ling)
batch_y = np.array(list(batch_y))
examples.append((batch_rnn, batch_sent_mask, batch_word_mask,
batch_cnn, batch_sent, batch_doc, batch_y, mb))
elif char and sent_ling and not doc_ling:
batch_x, batch_sent, batch_y = zip(*batch_x)
batch_x = util.vectorization(list(batch_x),
word_dict,
char_dict,
max_char_length=args.max_char)
batch_rnn, batch_sent_mask, batch_word_mask, batch_cnn = \
util.mask_padding(batch_x, args.max_sent, args.max_word, args.max_char)
batch_sent = util.sent_ling_padding(list(batch_sent),
args.max_sent, args.max_ling)
batch_y = np.array(list(batch_y))
examples.append((batch_rnn, batch_sent_mask, batch_word_mask,
batch_cnn, batch_sent, batch_y, mb))
elif char and not sent_ling and doc_ling:
batch_x, batch_doc, batch_y = zip(*batch_x)
batch_x = util.vectorization(list(batch_x),
word_dict,
char_dict,
max_char_length=args.max_char)
batch_rnn, batch_sent_mask, batch_word_mask, batch_cnn = \
util.mask_padding(batch_x, args.max_sent, args.max_word, args.max_char)
batch_doc = util.doc_ling_padding(list(batch_doc), args.max_ling)
batch_y = np.array(list(batch_y))
examples.append((batch_rnn, batch_sent_mask, batch_word_mask,
batch_cnn, batch_doc, batch_y, mb))
elif char and not sent_ling and not doc_ling:
batch_x, batch_y = zip(*batch_x)
batch_x = util.vectorization(list(batch_x),
word_dict,
char_dict,
max_char_length=args.max_char)
batch_rnn, batch_sent_mask, batch_word_mask, batch_cnn = \
util.mask_padding(batch_x, args.max_sent, args.max_word, args.max_char)
batch_y = np.array(list(batch_y))
examples.append((batch_rnn, batch_sent_mask, batch_word_mask,
batch_cnn, batch_y, mb))
elif not char and not sent_ling and not doc_ling:
batch_x, batch_y = zip(*batch_x)
batch_x = util.vectorization(list(batch_x),
word_dict,
char_dict,
max_char_length=args.max_char)
batch_rnn, batch_sent_mask, batch_word_mask, _ = \
util.mask_padding(batch_x, args.max_sent, args.max_word, args.max_char)
batch_y = np.array(list(batch_y))
examples.append(
(batch_rnn, batch_sent_mask, batch_word_mask, batch_y, mb))
elif not char and sent_ling and doc_ling:
batch_x, batch_sent, batch_doc, batch_y = zip(*batch_x)
batch_x = util.vectorization(list(batch_x),
word_dict,
char_dict,
max_char_length=args.max_char)
batch_rnn, batch_sent_mask, batch_word_mask, _ = \
util.mask_padding(batch_x, args.max_sent, args.max_word, args.max_char)
batch_sent = util.sent_ling_padding(list(batch_sent),
args.max_sent, args.max_ling)
batch_doc = util.doc_ling_padding(list(batch_doc), args.max_ling)
batch_y = np.array(list(batch_y))
examples.append((batch_rnn, batch_sent_mask, batch_word_mask,
batch_sent, batch_doc, batch_y, mb))
return examples