def show_heatmap(data_path, model_path):
import seaborn
import matplotlib
seaborn.set()
matplotlib.rc('font', family='sans-serif')
# call dictionary class
if args.lang == 'en':
corpus = ConvCorpus(file_path=None)
corpus.load(load_dir=data_path)
elif args.lang == 'ja':
corpus = JaConvCorpus(file_path=None)
corpus.load(load_dir=data_path)
else:
print(
'You gave wrong argument to this system. Check out your argument about languages.'
)
raise ValueError
print('Vocabulary Size (number of words) :', len(corpus.dic.token2id))
print('')
# rebuild seq2seq model
model = Seq2Seq(len(corpus.dic.token2id),
feature_num=args.feature_num,
hidden_num=args.hidden_num,
label_num=args.label_num,
label_embed_num=args.label_embed,
batch_size=1,
gpu_flg=args.gpu)
serializers.load_hdf5(model_path, model)
# sentiment matrix in the decoder
sentiment_mat = model.dec.le.W.data
cmap = seaborn.diverging_palette(
220, 10, as_cmap=True) # Generate a custom diverging colormap
seaborn.heatmap(sentiment_mat,
cmap=cmap,
center=0,
linewidths=.5,
xticklabels=False
) # square=True, cbar_kws={"orientation": "horizontal"})
plt.xlabel("Dimension (=" + str(sentiment_mat.shape[1]) + ")")
plt.ylabel("Sentiment")
plt.savefig('./data/sentiment_matrix.png')
# encoder embedding matrix
encode_mat = model.enc.xe.W
seaborn.heatmap(encode_mat,
cmap=cmap,
center=0,
linewidths=.5,
xticklabels=False
) # square=True, cbar_kws={"orientation": "horizontal"})
plt.xlabel("Dimension (=" + str(sentiment_mat.shape[1]) + ")")
plt.ylabel("Sentiment")
plt.savefig('./data/sentiment_matrix.png')
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('./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(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)
def interpreter(data_path, model_path):
"""
Run this function, if you want to talk to seq2seq model.
if you type "exit", finish to talk.
:param data_path: the path of corpus you made model learn
:param model_path: the path of model you made learn
:return:
"""
# call dictionary class
if args.lang == 'en':
corpus = ConvCorpus(file_path=None)
corpus.load(load_dir=data_path)
elif args.lang == 'ja':
corpus = JaConvCorpus(file_path=None)
corpus.load(load_dir=data_path)
else:
print(
'You gave wrong argument to this system. Check out your argument about languages.'
)
raise ValueError
print('Vocabulary Size (number of words) :', len(corpus.dic.token2id))
print('')
# rebuild seq2seq model
model = Seq2Seq(len(corpus.dic.token2id),
feature_num=args.feature_num,
hidden_num=args.hidden_num,
label_num=args.label_num,
label_embed_num=args.label_embed,
batch_size=1,
gpu_flg=args.gpu)
serializers.load_hdf5(model_path, model)
# run conversation system
print('The system is ready to run, please talk to me!')
print('( If you want to end a talk, please type "exit". )')
print('')
while True:
print('>> ', end='')
sentence = input()
if sentence == 'exit':
print('See you again!')
break
# convert to a list
if args.lang == 'en':
input_vocab = [
unicodedata.normalize('NFKC', word.lower())
for word in word_tokenize(sentence)
]
elif args.lang == 'ja':
input_vocab = parse_ja_text(sentence)
else:
print("Sorry, but your language is not supported...")
raise ValueError
# check a sentiment tag
label_id = -1
if len(input_vocab) == 0:
print('caution: you donot set any words!)')
pass
elif input_vocab[-1] == '2':
del input_vocab[-1]
label_id = 1
elif input_vocab[-1] == '1':
del input_vocab[-1]
label_id = 0
else:
print('caution: you donot set any sentiment tags!')
break
# input_vocab.reverse()
# input_vocab.insert(0, "<eos>")
# convert word into ID
input_sentence = [
corpus.dic.token2id[word] for word in input_vocab
if not corpus.dic.token2id.get(word) is None
]
model.initialize() # initialize cell
sentence = model.generate(input_sentence,
sentence_limit=len(input_sentence) + 30,
word2id=corpus.dic.token2id,
id2word=corpus.dic,
label_id=label_id)
print("-> ", sentence)
print('')
def calculate_embedding_vectors(data_path, model_path):
# call dictionary class
if args.lang == 'en':
corpus = ConvCorpus(file_path=None)
corpus.load(load_dir=data_path)
elif args.lang == 'ja':
corpus = JaConvCorpus(file_path=None)
corpus.load(load_dir=data_path)
else:
print(
'You gave wrong argument to this system. Check out your argument about languages.'
)
raise ValueError
print('Vocabulary Size (number of words) :', len(corpus.dic.token2id))
print('')
# rebuild seq2seq model
model = Seq2Seq(len(corpus.dic.token2id),
feature_num=args.feature_num,
hidden_num=args.hidden_num,
label_num=args.label_num,
label_embed_num=args.label_embed,
batch_size=1,
gpu_flg=args.gpu)
serializers.load_hdf5(model_path, model)
# get embedding vectors
embed_mat = model.dec.ye.W.data
sentiment_mat = model.dec.le.W.data
neg_vec = np.array([sentiment_mat[0, :]])
pos_vec = np.array([sentiment_mat[1, :]])
# calculate cos similarity
neg_sim_dic = {}
pos_sim_dic = {}
for i in range(embed_mat.shape[0]):
word_vec = np.array([embed_mat[i, :]])
neg_sim_dic[i] = cosine_similarity(word_vec, neg_vec)
pos_sim_dic[i] = cosine_similarity(word_vec, pos_vec)
# if cosine_similarity(word_vec, pos_vec) > cosine_similarity(word_vec, neg_vec):
# print('pos: ', corpus.dic[i])
# elif cosine_similarity(word_vec, pos_vec) < cosine_similarity(word_vec, neg_vec):
# print('neg: ', corpus.dic[i])
# else:
# print('???: ', corpus.dic[i])
# raise ValueError
# sort in descending order
neg_ordered = collections.OrderedDict(
sorted(neg_sim_dic.items(), key=lambda x: x[1], reverse=True))
pos_ordered = collections.OrderedDict(
sorted(pos_sim_dic.items(), key=lambda x: x[1], reverse=True))
# show TOP50 words
print('------- The words which is similar to a NEGATIVE tag --------')
for index, w_index in enumerate(neg_ordered):
print(corpus.dic[w_index], ': ', neg_ordered[w_index][0, 0])
if index == 49:
break
print('------- The words which is similar to a POSITIVE tag --------')
for index, w_index in enumerate(pos_ordered):
print(corpus.dic[w_index], ': ', pos_ordered[w_index][0, 0])
if index == 49:
break
def interpreter(data_path, model_path):
"""
Run this function, if you want to talk to seq2seq model.
if you type "exit", finish to talk.
:param data_path: the path of corpus you made model learn
:param model_path: the path of model you made learn
:return:
"""
# call dictionary class
if args.lang == 'en':
corpus = ConvCorpus(file_path=None)
corpus.load(load_dir=data_path)
elif args.lang == 'ja':
corpus = JaConvCorpus(file_path=None)
corpus.load(load_dir=data_path)
else:
print(
'You gave wrong argument to this system. Check out your argument about languages.'
)
raise ValueError
print('Vocabulary Size (number of words) :', len(corpus.dic.token2id))
print('')
# rebuild seq2seq model
model = Seq2Seq(len(corpus.dic.token2id),
feature_num=args.feature_num,
hidden_num=args.hidden_num,
batch_size=1,
gpu_flg=args.gpu)
serializers.load_hdf5(model_path, model)
# run conversation system
print('The system is ready to run, please talk to me!')
print('( If you want to end a talk, please type "exit". )')
print('')
while True:
print('>> ', end='')
sentence = input()
if sentence == 'exit':
print('See you again!')
break
if args.lang == 'en':
input_vocab = [
unicodedata.normalize('NFKC', word.lower())
for word in word_tokenize(sentence)
]
elif args.lang == 'ja':
input_vocab = [
unicodedata.normalize('NFKC', word.lower())
for word in parse_ja_text(sentence)
]
input_vocab_rev = input_vocab[::-1]
# convert word into ID
input_sentence = [
corpus.dic.token2id[word] for word in input_vocab
if not corpus.dic.token2id.get(word) is None
]
input_sentence_rev = [
corpus.dic.token2id[word] for word in input_vocab_rev
if not corpus.dic.token2id.get(word) is None
]
model.initialize(batch_size=1) # initialize cell
sentence = model.generate(input_sentence,
input_sentence_rev,
sentence_limit=len(input_sentence) + 30,
word2id=corpus.dic.token2id,
id2word=corpus.dic)
print("-> ", sentence)
print('')
def interpreter(data_path, model_path):
"""
Run this function, if you want to talk to seq2seq model.
if you type "exit", finish to talk.
:param data_path: the path of corpus you made model learn
:param model_path: the path of model you made learn
:return:
"""
# call dictionary class
corpus = JaConvCorpus(create_flg=False)
corpus.load(load_dir=data_path)
print('Vocabulary Size (number of words) :', len(corpus.dic.token2id))
print('')
# rebuild seq2seq model
model = FineTuneSeq2Seq(all_vocab_size=len(corpus.dic.token2id), emotion_vocab_size=len(corpus.emotion_set),
feature_num=args.feature_num, hidden_num=args.hidden_num,
label_num=args.label_num, label_embed_num=args.label_embed, batch_size=1, gpu_flg=args.gpu)
serializers.load_hdf5(model_path, model)
emo_label_index = [index for index in range(args.label_num)]
topic_label_index = [index for index in range(args.topic_num)]
# run conversation system
print('The system is ready to run, please talk to me!')
print('( If you want to end a talk, please type "exit". )')
print('')
while True:
print('>> ', end='')
sentence = input()
if sentence == 'exit':
print('See you again!')
break
# check a sentiment tag
input_vocab = sentence.split(' ')
emo_label_id = input_vocab.pop(-1)
topic_label_id = input_vocab.pop(-1)
label_false_flg = 1
for index in emo_label_index:
if emo_label_id == str(index):
emo_label_id = index # TODO: ラベルのインデックスに注意．今は3値分類 (0, 1, 2)
label_false_flg = 0
break
if label_false_flg:
print('caution: you donot set any enable tags! (emotion label)')
emo_label_id = -1
# check a topic tag # TODO: 本当はユーザ側の指定ではなく，tweet2vecの判定から決定する
label_false_flg = 1
for index in topic_label_index:
if topic_label_id == str(index):
topic_label_id = index # TODO: ラベルのインデックスに注意．今は3値分類 (0, 1, 2)
label_false_flg = 0
break
if label_false_flg:
print('caution: you donot set any enable tags! (topic label)')
topic_label_id = -1
input_vocab = [unicodedata.normalize('NFKC', word.lower()) for word in parse_ja_text(sentence)]
input_vocab_rev = input_vocab[::-1]
# convert word into ID
input_sentence = [corpus.dic.token2id[word] for word in input_vocab if not corpus.dic.token2id.get(word) is None]
input_sentence_rev = [corpus.dic.token2id[word] for word in input_vocab_rev if not corpus.dic.token2id.get(word) is None]
model.initialize(batch_size=1)
if args.beam_search:
hypotheses = model.beam_search(model.initial_state_function, model.generate_function,
input_sentence, input_sentence_rev, start_id=corpus.dic.token2id['<start>'],
end_id=corpus.dic.token2id['<eos>'], emo_label_id=emo_label_id,
topic_label_id=topic_label_id)
for hypothesis in hypotheses:
generated_indices = hypothesis.to_sequence_of_values()
generated_tokens = [corpus.dic[i] for i in generated_indices]
print("--> ", " ".join(generated_tokens))
else:
sentence = model.generate(input_sentence, input_sentence_rev, sentence_limit=len(input_sentence) + 20,
emo_label_id=emo_label_id, topic_label_id=topic_label_id,
word2id=corpus.dic.token2id, id2word=corpus.dic)
print("-> ", sentence)
print('')
def output_file(data_path, model_path):
"""
:param data_path: the path of corpus you made model learn
:param model_path: the path of model you made learn
:return:
"""
# call dictionary class
corpus = JaConvCorpus(create_flg=False)
corpus.load(load_dir=data_path)
print('Vocabulary Size (number of words) :', len(corpus.dic.token2id))
print('')
# rebuild seq2seq model
model = FineTuneSeq2Seq(all_vocab_size=len(corpus.dic.token2id), emotion_vocab_size=len(corpus.emotion_set),
feature_num=args.feature_num, hidden_num=args.hidden_num,
label_num=args.label_num, label_embed_num=args.label_embed, batch_size=1, gpu_flg=args.gpu)
serializers.load_hdf5(model_path, model)
emo_label_index = [index for index in range(args.label_num)]
topic_label_index = [index for index in range(args.topic_num)]
# run conversation system
r_label = re.compile("(__label__)([0-9]+)")
pattern = "(.+?)(\t)(.+?)(\n|\r\n)"
r = re.compile(pattern)
for line in open(T2V_OUTPUT, 'r', encoding='utf-8'):
m = r.search(line)
if m is not None:
topic_label = m.group(1)
sentence = m.group(3)
# check a topic tag
label_info = r_label.search(topic_label)
if int(label_info.group(2)) < len(topic_label_index):
topic_label_id = int(label_info.group(2))
else:
print('domain label がドメイン数の上限を超えています．')
raise ValueError
# parse text by mecab
input_vocab = [unicodedata.normalize('NFKC', word.lower()) for word in parse_ja_text(sentence)]
input_vocab_rev = input_vocab[::-1]
# convert word into ID
input_sentence = [corpus.dic.token2id[word] for word in input_vocab if not corpus.dic.token2id.get(word) is None]
input_sentence_rev = [corpus.dic.token2id[word] for word in input_vocab_rev if not corpus.dic.token2id.get(word) is None]
print("input -> ", sentence, "domain:", topic_label_id)
model.initialize(batch_size=1)
for emo_label in range(LABEL_NUM):
sentence = model.generate(input_sentence, input_sentence_rev, sentence_limit=len(input_sentence) + 20,
emo_label_id=emo_label, topic_label_id=topic_label_id,
word2id=corpus.dic.token2id, id2word=corpus.dic)
if emo_label == 0:
print("neg -> ", sentence)
elif emo_label == 1:
print("neu -> ", sentence)
elif emo_label == 2:
print("pos -> ", sentence)
else:
raise ValueError
print('')