def main():
###########################
#### create dictionary ####
###########################
if os.path.exists('./data/corpus/dictionary.dict'):
if args.lang == 'ja':
corpus = JaConvCorpus(file_path=None,
batch_size=batchsize,
size_filter=True)
else:
corpus = ConvCorpus(file_path=None,
batch_size=batchsize,
size_filter=True)
corpus.load(load_dir='./data/corpus/')
else:
if args.lang == 'ja':
corpus = JaConvCorpus(file_path=data_file,
batch_size=batchsize,
size_filter=True)
else:
corpus = ConvCorpus(file_path=data_file,
batch_size=batchsize,
size_filter=True)
corpus.save(save_dir='./data/corpus/')
print('Vocabulary Size (number of words) :', len(corpus.dic.token2id))
######################
#### create model ####
######################
model = Seq2Seq(vocab_size=len(corpus.dic.token2id),
feature_num=feature_num,
hidden_num=hidden_num,
batch_size=batchsize,
gpu_flg=args.gpu)
if args.gpu >= 0:
model.to_gpu()
optimizer = optimizers.Adam(alpha=0.001)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.GradientClipping(5))
##########################
#### create ID corpus ####
##########################
input_mat = []
output_mat = []
input_mat_rev = []
# output_wp_mat = []
max_input_ren = max_output_ren = 0
for input_text, output_text in zip(corpus.posts, corpus.cmnts):
output_text.append(corpus.dic.token2id["<eos>"])
# update max sentence length
max_input_ren = max(max_input_ren, len(input_text))
max_output_ren = max(max_output_ren, len(output_text))
input_mat.append(input_text)
output_mat.append(output_text)
# # create word prediction matrix
# wp = []
# for wid in output_text:
# if wid not in wp:
# wp.append(wid)
# output_wp_mat.append(wp)
# make reverse corpus
for input_text in input_mat:
input_mat_rev.append(input_text[::-1])
# padding
for li in input_mat:
insert_num = max_input_ren - len(li)
for _ in range(insert_num):
li.append(corpus.dic.token2id['<pad>'])
for li in output_mat:
insert_num = max_output_ren - len(li)
for _ in range(insert_num):
li.append(corpus.dic.token2id['<pad>'])
for li in input_mat_rev:
insert_num = max_input_ren - len(li)
for _ in range(insert_num):
li.insert(0, corpus.dic.token2id['<pad>'])
# create batch matrix
input_mat = np.array(input_mat, dtype=np.int32).T
input_mat_rev = np.array(input_mat_rev, dtype=np.int32).T
output_mat = np.array(output_mat, dtype=np.int32).T
# separate corpus into Train and Test
perm = np.random.permutation(len(corpus.posts))
test_input_mat = input_mat[:, perm[0:0 + testsize]]
test_output_mat = output_mat[:, perm[0:0 + testsize]]
test_input_mat_rev = input_mat_rev[:, perm[0:0 + testsize]]
train_input_mat = input_mat[:, perm[testsize:]]
train_output_mat = output_mat[:, perm[testsize:]]
train_input_mat_rev = input_mat_rev[:, perm[testsize:]]
# train_output_wp_mat = []
# for index in perm[testsize:]:
# train_output_wp_mat.append(output_wp_mat[index])
#############################
#### train seq2seq model ####
#############################
accum_loss = 0
train_loss_data = []
for num, epoch in enumerate(range(n_epoch)):
total_loss = 0
batch_num = 0
perm = np.random.permutation(len(corpus.posts) - testsize)
# for training
for i in range(0, len(corpus.posts) - testsize, batchsize):
# select batch data
input_batch = remove_extra_padding(
train_input_mat[:, perm[i:i + batchsize]], reverse_flg=False)
input_batch_rev = remove_extra_padding(
train_input_mat_rev[:, perm[i:i + batchsize]],
reverse_flg=True)
output_batch = remove_extra_padding(
train_output_mat[:, perm[i:i + batchsize]], reverse_flg=False)
# output_wp_batch = []
# for index in perm[i:i + batchsize]:
# output_wp_batch.append(train_output_wp_mat[index])
# output_wp_batch = create_wp_batch(vocab_size=len(corpus.dic.token2id),
# wp_lists=output_wp_batch)
# Encode a sentence
model.initialize(batch_size=input_batch.shape[1])
model.encode(input_batch, input_batch_rev, train=True)
# Decode from encoded context
end_batch = xp.array([
corpus.dic.token2id["<start>"]
for _ in range(input_batch.shape[1])
])
first_words = output_batch[0]
loss, predict_mat = model.decode(end_batch,
first_words,
train=True)
next_ids = first_words
accum_loss += loss
for w_ids in output_batch[1:]:
loss, predict_mat = model.decode(next_ids, w_ids, train=True)
next_ids = w_ids
accum_loss += loss
# learn model
model.cleargrads() # initialize all grad to zero
accum_loss.backward() # back propagation
optimizer.update()
total_loss += float(accum_loss.data)
batch_num += 1
print('Epoch: ', num, 'Batch_num', batch_num,
'batch loss: {:.2f}'.format(float(accum_loss.data)))
accum_loss = 0
train_loss_data.append(float(total_loss / batch_num))
# save model and optimizer
if (epoch + 1) % 5 == 0:
print('-----', epoch + 1, ' times -----')
print('save the model and optimizer')
serializers.save_hdf5('data/' + str(epoch) + '.model', model)
serializers.save_hdf5('data/' + str(epoch) + '.state', optimizer)
# save loss data
with open('./data/loss_train_data.pkl', 'wb') as f:
pickle.dump(train_loss_data, f)
def main():
###########################
#### create dictionary ####
###########################
if os.path.exists('./data/corpus/dictionary.dict'):
if args.lang == 'ja':
corpus = JaConvCorpus(file_path=None, batch_size=batchsize, size_filter=True)
else:
corpus = ConvCorpus(file_path=None, batch_size=batchsize)
corpus.load(load_dir='./data/corpus/')
else:
if args.lang == 'ja':
corpus = JaConvCorpus(file_path=data_file, batch_size=batchsize, size_filter=True)
else:
corpus = ConvCorpus(file_path=data_file, batch_size=batchsize)
corpus.save(save_dir='./data/corpus/')
print('Vocabulary Size (number of words) :', len(corpus.dic.token2id))
######################
#### create model ####
######################
model = Seq2Seq(len(corpus.dic.token2id), feature_num=feature_num,
hidden_num=hidden_num, label_num=label_num,
label_embed_num=label_embed_num,
batch_size=batchsize, gpu_flg=args.gpu)
if args.gpu >= 0:
model.to_gpu()
optimizer = optimizers.Adam(alpha=0.001)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.GradientClipping(5))
# optimizer.add_hook(chainer.optimizer.WeightDecay(0.0001))
##########################
#### create ID corpus ####
##########################
input_mat = []
output_mat = []
label_mat = []
max_input_ren = max_output_ren = 0
for input_text, output_text in zip(corpus.posts, corpus.cmnts):
# convert to list
# input_text.reverse() # encode words in a reverse order
# input_text.insert(0, corpus.dic.token2id["<eos>"])
output_text.append(corpus.dic.token2id["<eos>"])
# update max sentence length
max_input_ren = max(max_input_ren, len(input_text))
max_output_ren = max(max_output_ren, len(output_text))
# listのlistを作る（要修正）
label = input_text.pop(-1)
if label == corpus.dic.token2id["__label__1"]:
label_mat.append([0 for _ in range(len(output_text))])
elif label == corpus.dic.token2id["__label__2"]:
label_mat.append([1 for _ in range(len(output_text))])
else:
print('label error!: ', label)
raise ValueError
input_mat.append(input_text)
output_mat.append(output_text)
# padding
for li in input_mat:
insert_num = max_input_ren - len(li)
for _ in range(insert_num):
li.append(corpus.dic.token2id['<pad>']) # そのままの入力順序にする場合
# li.insert(0, corpus.dic.token2id['<pad>']) # 入力順序を逆にする場合
for li in output_mat:
insert_num = max_output_ren - len(li)
for _ in range(insert_num):
li.append(corpus.dic.token2id['<pad>'])
for li in label_mat:
insert_num = max_output_ren - len(li)
for _ in range(insert_num):
li.append(corpus.dic.token2id['<pad>'])
# create batch matrix
input_mat = np.array(input_mat, dtype=np.int32).T
output_mat = np.array(output_mat, dtype=np.int32).T
label_mat = np.array(label_mat, dtype=np.int32).T
# separate corpus into Train and Test
perm = np.random.permutation(len(corpus.posts))
test_input_mat = input_mat[:, perm[0:0 + testsize]]
test_output_mat = output_mat[:, perm[0:0 + testsize]]
test_label_mat = label_mat[:, perm[0:0 + testsize]]
train_input_mat = input_mat[:, perm[testsize:]]
train_output_mat = output_mat[:, perm[testsize:]]
train_label_mat = label_mat[:, perm[testsize:]]
list_of_references = []
for text_ndarray in test_output_mat.T:
reference = text_ndarray.tolist()
references = [[w_id for w_id in reference if w_id is not -1]]
list_of_references.append(references)
#############################
#### train seq2seq model ####
#############################
accum_loss = 0
train_loss_data = []
test_loss_data = []
bleu_score_data = []
wer_score_data = []
for num, epoch in enumerate(range(n_epoch)):
total_loss = test_loss = 0
batch_num = 0
perm = np.random.permutation(len(corpus.posts) - testsize)
# for training
for i in range(0, len(corpus.posts) - testsize, batchsize):
# select batch data
input_batch = train_input_mat[:, perm[i:i + batchsize]]
output_batch = train_output_mat[:, perm[i:i + batchsize]]
label_batch = train_label_mat[:, perm[i:i + batchsize]]
# Encode a sentence
model.initialize() # initialize cell
model.encode(input_batch, train=True) # encode (output: hidden Variable)
# Decode from encoded context
next_ids = xp.array([corpus.dic.token2id["<eos>"] for _ in range(batchsize)])
accum_loss = 0
for w_ids, l_ids in zip(output_batch, label_batch):
loss, predict_mat = model.decode(next_ids, w_ids, l_ids, train=True)
next_ids = w_ids
accum_loss += loss
# learn model
model.cleargrads() # initialize all grad to zero
accum_loss.backward() # back propagation
optimizer.update()
total_loss += float(accum_loss.data)
batch_num += 1
print('Epoch: ', num, 'Batch_num', batch_num, 'batch loss: {:.2f}'.format(float(accum_loss.data)))
# for testing
list_of_hypotheses = []
for i in range(0, testsize, batchsize):
# select test batch data
input_batch = test_input_mat[:, i:i + batchsize]
output_batch = test_output_mat[:, i:i + batchsize]
label_batch = test_label_mat[:, i:i + batchsize]
# Encode a sentence
model.initialize() # initialize cell
model.encode(input_batch, train=True) # encode (output: hidden Variable)
# Decode from encoded context
next_ids = xp.array([corpus.dic.token2id["<start>"] for _ in range(batchsize)])
if args.gpu >= 0:
hypotheses = [cuda.to_cpu(next_ids)]
else:
hypotheses = [next_ids]
for w_ids, l_ids in zip(output_batch, label_batch):
loss, predict_mat = model.decode(next_ids, w_ids, l_ids, train=True)
next_ids = xp.argmax(predict_mat.data, axis=1)
test_loss += loss.data
if args.gpu >= 0:
hypotheses.append(cuda.to_cpu(next_ids))
else:
hypotheses.append(next_ids)
# collect hypotheses for calculating BLEU score
hypotheses = np.array(hypotheses).T
for hypothesis in hypotheses:
text_list = hypothesis.tolist()
list_of_hypotheses.append([w_id for w_id in text_list if w_id is not -1])
# calculate BLEU score from test (develop) data
bleu_score = nltk.translate.bleu_score.corpus_bleu(list_of_references, list_of_hypotheses,
weights=(0.25, 0.25, 0.25, 0.25))
bleu_score_data.append(bleu_score)
print('Epoch: ', num, 'BLEU SCORE: ', bleu_score)
# calculate WER score from test (develop) data
wer_score = 0
for index, references in enumerate(list_of_references):
wer_score += wer(references[0], list_of_hypotheses[index])
wer_score /= len(list_of_references)
wer_score_data.append(wer_score)
print('Epoch: ', num, 'WER SCORE: ', wer_score)
# save model and optimizer
if (epoch + 1) % 10 == 0:
print('-----', epoch + 1, ' times -----')
print('save the model and optimizer')
serializers.save_hdf5('data/' + str(epoch) + '.model', model)
serializers.save_hdf5('data/' + str(epoch) + '.state', optimizer)
# display the on-going status
print('Epoch: ', num,
'Train loss: {:.2f}'.format(total_loss),
'Test loss: {:.2f}'.format(float(test_loss)))
train_loss_data.append(float(total_loss / batch_num))
test_loss_data.append(float(test_loss))
# evaluate a test loss
check_loss = test_loss_data[-10:] # check out the last 10 loss data
end_flg = [j for j in range(len(check_loss) - 1) if check_loss[j] < check_loss[j + 1]]
if len(end_flg) > 9:
print('Probably it is over-fitting. So stop to learn...')
break
# save loss data
with open('./data/loss_train_data.pkl', 'wb') as f:
pickle.dump(train_loss_data, f)
with open('./data/loss_test_data.pkl', 'wb') as f:
pickle.dump(test_loss_data, f)
with open('./data/bleu_score_data.pkl', 'wb') as f:
pickle.dump(bleu_score_data, f)
with open('./data/wer_score_data.pkl', 'wb') as f:
pickle.dump(wer_score_data, f)
def main():
###########################
#### create dictionary ####
###########################
if os.path.exists(CORPUS_DIR + 'dictionary.dict'):
corpus = JaConvCorpus(create_flg=False,
batch_size=batchsize,
size_filter=True)
corpus.load(load_dir=CORPUS_DIR)
else:
corpus = JaConvCorpus(create_flg=True,
batch_size=batchsize,
size_filter=True)
corpus.save(save_dir=CORPUS_DIR)
print('Vocabulary Size (number of words) :', len(corpus.dic.token2id))
print('Emotion size: ', len(corpus.emotion_set))
# search word_threshold (general or emotional)
ma = 0
mi = 999999
for word in corpus.emotion_set:
wid = corpus.dic.token2id[word]
if wid > ma:
ma = wid
if wid < mi:
mi = wid
word_threshold = mi
######################
#### create model ####
######################
model = PreTrainSeq2Seq(all_vocab_size=len(corpus.dic.token2id),
emotion_vocab_size=len(corpus.emotion_set),
feature_num=feature_num,
hidden_num=hidden_num,
batch_size=batchsize,
label_num=label_num,
label_embed_num=label_embed,
gpu_flg=args.gpu)
if args.gpu >= 0:
model.to_gpu()
optimizer = optimizers.Adam(alpha=0.001)
optimizer.setup(model)
optimizer.add_hook(chainer.optimizer.WeightDecay(0.0001))
##########################
#### create ID corpus ####
##########################
input_mat = []
output_mat = []
input_mat_rev = []
label_mat = []
max_input_ren = max_output_ren = 0
print('start making corpus matrix...')
for input_text, output_text in zip(corpus.rough_posts, corpus.rough_cmnts):
# reverse an input and add eos tag
output_text.append(corpus.dic.token2id["<eos>"]) # 出力の最後にeosを挿入
# update max sentence length
max_input_ren = max(max_input_ren, len(input_text))
max_output_ren = max(max_output_ren, len(output_text))
# make a list of lists
input_mat.append(input_text)
output_mat.append(output_text)
# make label lists TODO: 3値分類（pos, neg, neu）のみの対応なので可変にする
n_num = p_num = 0
for word in output_text:
if corpus.dic[word] in corpus.neg_words:
n_num += 1
if corpus.dic[word] in corpus.pos_words:
p_num += 1
if (n_num + p_num) == 0:
label_mat.append([1 for _ in range(len(output_text))])
elif n_num <= p_num:
label_mat.append([2 for _ in range(len(output_text))])
elif n_num > p_num:
label_mat.append([0 for _ in range(len(output_text))])
else:
raise ValueError
# make reverse corpus
for input_text in input_mat:
input_mat_rev.append(input_text[::-1])
# padding (inputの文頭・outputの文末にパディングを挿入する)
print('start labeling...')
for li in input_mat:
insert_num = max_input_ren - len(li)
for _ in range(insert_num):
li.append(corpus.dic.token2id['<pad>'])
for li in output_mat:
insert_num = max_output_ren - len(li)
for _ in range(insert_num):
li.append(corpus.dic.token2id['<pad>'])
for li in input_mat_rev:
insert_num = max_input_ren - len(li)
for _ in range(insert_num):
li.insert(0, corpus.dic.token2id['<pad>'])
for li in label_mat:
insert_num = max_output_ren - len(li)
for _ in range(insert_num):
li.append(corpus.dic.token2id['<pad>'])
if len(output_mat) != len(label_mat):
print('Output matrix and label matrix should have the same dimension.')
raise ValueError
# create batch matrix
print('transpose...')
input_mat = np.array(input_mat, dtype=np.int32).T
input_mat_rev = np.array(input_mat_rev, dtype=np.int32).T
output_mat = np.array(output_mat, dtype=np.int32).T
label_mat = np.array(label_mat, dtype=np.int32).T
# separate corpus into Train and Test TODO:実験時はテストデータとトレーニングデータに分離する
print('split train and test...')
train_input_mat = input_mat
train_output_mat = output_mat
train_input_mat_rev = input_mat_rev
train_label_mat = label_mat
#############################
#### train seq2seq model ####
#############################
accum_loss = 0
train_loss_data = []
print('start training...')
for num, epoch in enumerate(range(n_epoch)):
total_loss = 0
batch_num = 0
perm = np.random.permutation(len(corpus.rough_posts))
# for training
for i in range(0, len(corpus.rough_posts), batchsize):
# select batch data
input_batch = remove_extra_padding(
train_input_mat[:, perm[i:i + batchsize]], reverse_flg=False)
input_batch_rev = remove_extra_padding(
train_input_mat_rev[:, perm[i:i + batchsize]],
reverse_flg=True)
output_batch = remove_extra_padding(
train_output_mat[:, perm[i:i + batchsize]], reverse_flg=False)
label_batch = remove_extra_padding(
train_label_mat[:, perm[i:i + batchsize]], reverse_flg=False)
# Encode a sentence
model.initialize(
batch_size=input_batch.shape[1]) # initialize cell
model.encode(input_batch, input_batch_rev,
train=True) # encode (output: hidden Variable)
# Decode from encoded context
input_ids = xp.array([
corpus.dic.token2id["<start>"]
for _ in range(input_batch.shape[1])
])
for w_ids, l_ids in zip(output_batch, label_batch):
loss, predict_mat = model.decode(input_ids,
w_ids,
label_id=l_ids,
word_th=word_threshold,
train=True)
input_ids = w_ids
accum_loss += loss
# learn model
model.cleargrads() # initialize all grad to zero
accum_loss.backward() # back propagation
optimizer.update()
total_loss += float(accum_loss.data)
batch_num += 1
print('Epoch: ', num, 'Batch_num', batch_num,
'batch loss: {:.2f}'.format(float(accum_loss.data)))
accum_loss = 0
train_loss_data.append(float(total_loss / batch_num))
# save model and optimizer
print('-----', epoch + 1, ' times -----')
print('save the model and optimizer')
serializers.save_hdf5('../data/seq2seq/' + str(epoch) + '_rough.model',
model)
serializers.save_hdf5('../data/seq2seq/' + str(epoch) + '_rough.state',
optimizer)
# save loss data
with open('./data/loss_train_data.pkl', 'wb') as f:
pickle.dump(train_loss_data, f)