def main(_):
if os.path.exists(config.forward_log_path) and config.mode == 'forward':
os.system('rm ' + config.forward_log_path)
if os.path.exists(config.backward_log_path) and config.mode == 'backward':
os.system('rm ' + config.backward_log_path)
if os.path.exists(config.use_output_path):
os.system('rm ' + config.use_output_path)
for item in config.record_time:
if os.path.exists(config.use_output_path + str(item)):
os.system('rm ' + config.use_output_path + str(item))
if os.path.exists(config.use_log_path):
os.system('rm ' + config.use_log_path)
if config.mode == 'forward' or config.mode == 'use':
with tf.name_scope("forward_train"):
with tf.variable_scope("forward", reuse=None):
m_forward = PTBModel(is_training=True)
with tf.name_scope("forward_test"):
with tf.variable_scope("forward", reuse=True):
mtest_forward = PTBModel(is_training=False)
var = tf.trainable_variables()
var_forward = [x for x in var if x.name.startswith('forward')]
saver_forward = tf.train.Saver(var_forward, max_to_keep=1)
if config.mode == 'backward' or config.mode == 'use':
with tf.name_scope("backward_train"):
with tf.variable_scope("backward", reuse=None):
m_backward = PTBModel(is_training=True)
with tf.name_scope("backward_test"):
with tf.variable_scope("backward", reuse=True):
mtest_backward = PTBModel(is_training=False)
var = tf.trainable_variables()
var_backward = [x for x in var if x.name.startswith('backward')]
saver_backward = tf.train.Saver(var_backward, max_to_keep=1)
init = tf.global_variables_initializer()
with tf.Session() as session:
session.run(init)
if config.mode == 'forward':
#train forward language model
train_data, test_data = reader.read_data(config.data_path,
config.num_steps)
test_mean_old = 15.0
for epoch in range(config.max_epoch):
train_ppl_list = []
test_ppl_list = []
for i in range(train_data.length // config.batch_size):
input, sequence_length, target = train_data(
m_forward.batch_size, i)
train_perplexity = run_epoch(session,
m_forward,
input,
sequence_length,
target,
mode='train')
train_ppl_list.append(train_perplexity)
print("Epoch:%d, Iter: %d Train NLL: %.3f" %
(epoch, i + 1, train_perplexity))
for i in range(test_data.length // config.batch_size):
input, sequence_length, target = test_data(
mtest_forward.batch_size, i)
test_perplexity = run_epoch(session,
mtest_forward,
input,
sequence_length,
target,
mode='test')
test_ppl_list.append(test_perplexity)
print("Epoch:%d, Iter: %d Test NLL: %.3f" %
(epoch, i + 1, test_perplexity))
test_mean = np.mean(test_ppl_list)
if test_mean < test_mean_old:
test_mean_old = test_mean
saver_forward.save(session, config.forward_save_path)
write_log(
'train ppl:' + str(np.mean(train_ppl_list)) + '\t' +
'test ppl:' + str(test_mean), config.forward_log_path)
if config.mode == 'backward':
#train backward language model
train_data, test_data = reader.read_data(config.data_path,
config.num_steps)
test_mean_old = 15.0
for epoch in range(config.max_epoch):
train_ppl_list = []
test_ppl_list = []
for i in range(train_data.length // config.batch_size):
input, sequence_length, target = train_data(
m_backward.batch_size, i)
input, sequence_length, target = reverse_seq(
input, sequence_length, target)
train_perplexity = run_epoch(session,
m_backward,
input,
sequence_length,
target,
mode='train')
train_ppl_list.append(train_perplexity)
print("Epoch:%d, Iter: %d Train NLL: %.3f" %
(epoch, i + 1, train_perplexity))
for i in range(test_data.length // config.batch_size):
input, sequence_length, target = test_data(
mtest_backward.batch_size, i)
input, sequence_length, target = reverse_seq(
input, sequence_length, target)
test_perplexity = run_epoch(session,
mtest_backward,
input,
sequence_length,
target,
mode='test')
test_ppl_list.append(test_perplexity)
print("Epoch:%d, Iter: %d Test NLL: %.3f" %
(epoch, i + 1, test_perplexity))
test_mean = np.mean(test_ppl_list)
if test_mean < test_mean_old:
test_mean_old = test_mean
saver_backward.save(session, config.backward_save_path)
write_log(
'train ppl:' + str(np.mean(train_ppl_list)) + '\t' +
'test ppl:' + str(test_mean), config.backward_log_path)
if config.mode == 'use':
#CGMH sampling for sentence_correction
sim = config.sim
sta_vec = list(np.zeros([config.num_steps - 1]))
saver_forward.restore(session, config.forward_save_path)
saver_backward.restore(session, config.backward_save_path)
config.shuffle = False
#erroneous sentence input
if config.keyboard_input == True:
#input from keyboard if key_input is not empty
key_input = raw_input('please input a sentence\n')
if key_input == '':
use_data = reader.read_data_use(config.use_data_path,
config.num_steps)
else:
sta_vec_list = [sen2sta_vec(key_input)]
key_input = key_input.split()
#key_input=sen2id(key_input)
use_data = [key_input]
else:
#load keywords from file
use_data = []
with open(config.use_data_path) as f:
for line in f:
use_data.append(line.strip().split())
config.batch_size = 1
for sen_id in range(len(use_data)):
#generate for each sentence
input_ = use_data[sen_id]
pos = 0
for iter in range(config.sample_time):
#ind is the index of the selected word, regardless of the beginning token.
sta_vec = sen2sta_vec(' '.join(input_))
input__ = reader.array_data([sen2id(input_)],
config.num_steps,
config.dict_size)
input, sequence_length, _ = input__(1, 0)
input_original = input[0]
ind = pos % (sequence_length[0] - 1)
print(' '.join(input_))
if iter in config.record_time:
with open(config.use_output_path + str(iter),
'a') as g:
g.write(' '.join(input_) + '\n')
if True:
prob_old = run_epoch(session,
mtest_forward,
input,
sequence_length,
mode='use')
if config.double_LM == True:
input_backward, _, _ = reverse_seq(
input, sequence_length, input)
prob_old = (prob_old + run_epoch(session,
mtest_backward,
input_backward,
sequence_length,
mode='use')) * 0.5
tem = 1
for j in range(sequence_length[0] - 1):
tem *= prob_old[0][j][input[0][j + 1]]
tem *= prob_old[0][j + 1][config.dict_size + 1]
prob_old_prob = tem
if sim != None:
similarity_old = similarity(
input[0], input_original)
prob_old_prob *= similarity_old
else:
similarity_old = -1
input_candidate_ = generate_change_candidate(
input_, ind)
tem = reader.array_data(
[sen2id(x) for x in input_candidate_],
config.num_steps, config.dict_size)
input_candidate, sequence_length_candidate, _ = tem(
len(input_candidate_), 0)
prob_candidate_pre = run_epoch(
session,
mtest_forward,
input_candidate,
sequence_length_candidate,
mode='use')
if config.double_LM == True:
input_candidate_backward, _, _ = reverse_seq(
input_candidate, sequence_length_candidate,
input_candidate)
prob_candidate_pre = (
prob_candidate_pre +
run_epoch(session,
mtest_backward,
input_candidate_backward,
sequence_length_candidate,
mode='use')) * 0.5
prob_candidate = []
for i in range(len(input_candidate_)):
tem = 1
for j in range(sequence_length[0] - 1):
tem *= prob_candidate_pre[i][j][
input_candidate[i][j + 1]]
tem *= prob_candidate_pre[i][j +
1][config.dict_size +
1]
prob_candidate.append(tem)
prob_candidate = np.array(prob_candidate)
if sim != None:
similarity_candidate = similarity_batch(
input_candidate, input_original)
prob_candidate = prob_candidate * similarity_candidate
prob_candidate_norm = normalize(prob_candidate)
prob_candidate_ind = sample_from_candidate(
prob_candidate_norm)
prob_change_prob = prob_candidate[prob_candidate_ind]
input_change_ = input_candidate_[prob_candidate_ind]
#word replacement (action: 0)
if True:
if False:
pass
else:
input_forward, input_backward, sequence_length_forward, sequence_length_backward = cut_from_point(
input, sequence_length, ind, mode=0)
prob_forward = run_epoch(
session,
mtest_forward,
input_forward,
sequence_length_forward,
mode='use')[0,
ind % (sequence_length[0] - 1), :]
prob_backward = run_epoch(
session,
mtest_backward,
input_backward,
sequence_length_backward,
mode='use')[0, sequence_length[0] - 1 - ind %
(sequence_length[0] - 1), :]
prob_mul = (prob_forward * prob_backward)
input_candidate, sequence_length_candidate = generate_candidate_input(
input,
sequence_length,
ind,
prob_mul,
config.search_size,
mode=1)
prob_candidate_pre = run_epoch(
session,
mtest_forward,
input_candidate,
sequence_length_candidate,
mode='use')
if config.double_LM == True:
input_candidate_backward, _, _ = reverse_seq(
input_candidate, sequence_length_candidate,
input_candidate)
prob_candidate_pre = (
prob_candidate_pre +
run_epoch(session,
mtest_backward,
input_candidate_backward,
sequence_length_candidate,
mode='use')) * 0.5
prob_candidate = []
for i in range(config.search_size):
tem = 1
for j in range(sequence_length_candidate[0] -
1):
tem *= prob_candidate_pre[i][j][
input_candidate[i][j + 1]]
tem *= prob_candidate_pre[i][j + 1][
config.dict_size + 1]
prob_candidate.append(tem)
prob_candidate = np.array(prob_candidate)
if config.sim_word == True:
similarity_candidate = similarity_batch(
input_candidate[:, ind + 1:ind + 2],
input_original[ind + 1:ind + 2])
prob_candidate = prob_candidate * similarity_candidate
prob_candidate_norm = normalize(prob_candidate)
prob_candidate_ind = sample_from_candidate(
prob_candidate_norm)
prob_candidate_prob = prob_candidate[
prob_candidate_ind]
prob_changeanother_prob = prob_candidate_prob
word = id2sen(
input_candidate[prob_candidate_ind])[ind]
input_changeanother_ = input_[:ind] + [
word
] + input_[ind + 1:]
#word insertion(action:1)
if True:
if sequence_length[0] >= config.num_steps:
prob_add_prob = 0
pass
else:
input_forward, input_backward, sequence_length_forward, sequence_length_backward = cut_from_point(
input, sequence_length, ind, mode=1)
prob_forward = run_epoch(
session,
mtest_forward,
input_forward,
sequence_length_forward,
mode='use')[0,
ind % (sequence_length[0] - 1), :]
prob_backward = run_epoch(
session,
mtest_backward,
input_backward,
sequence_length_backward,
mode='use')[0, sequence_length[0] - 1 - ind %
(sequence_length[0] - 1), :]
prob_mul = (prob_forward * prob_backward)
input_candidate, sequence_length_candidate = generate_candidate_input(
input,
sequence_length,
ind,
prob_mul,
config.search_size,
mode=1)
prob_candidate_pre = run_epoch(
session,
mtest_forward,
input_candidate,
sequence_length_candidate,
mode='use')
if config.double_LM == True:
input_candidate_backward, _, _ = reverse_seq(
input_candidate, sequence_length_candidate,
input_candidate)
prob_candidate_pre = (
prob_candidate_pre +
run_epoch(session,
mtest_backward,
input_candidate_backward,
sequence_length_candidate,
mode='use')) * 0.5
prob_candidate = []
for i in range(config.search_size):
tem = 1
for j in range(sequence_length_candidate[0] -
1):
tem *= prob_candidate_pre[i][j][
input_candidate[i][j + 1]]
tem *= prob_candidate_pre[i][j + 1][
config.dict_size + 1]
prob_candidate.append(tem)
prob_candidate = np.array(prob_candidate)
#similarity_candidate=np.array([similarity(x, input_original) for x in input_candidate])
if sim != None:
similarity_candidate = similarity_batch(
input_candidate, input_original)
prob_candidate = prob_candidate * similarity_candidate
prob_candidate_norm = normalize(prob_candidate)
prob_candidate_ind = sample_from_candidate(
prob_candidate_norm)
prob_candidate_prob = prob_candidate[
prob_candidate_ind]
prob_add_prob = prob_candidate_prob
word = id2sen(
input_candidate[prob_candidate_ind])[ind]
input_add_ = input_[:ind] + [word] + input_[ind:]
#word deletion(action: 2)
if True:
if sequence_length[0] <= 2:
prob_delete_prob = 0
pass
else:
input_candidate, sequence_length_candidate = generate_candidate_input(
input,
sequence_length,
ind,
None,
config.search_size,
mode=2)
prob_new = run_epoch(session,
mtest_forward,
input_candidate,
sequence_length_candidate,
mode='use')
tem = 1
for j in range(sequence_length_candidate[0] - 1):
tem *= prob_new[0][j][input_candidate[0][j +
1]]
tem *= prob_new[0][j + 1][config.dict_size + 1]
prob_new_prob = tem
if sim != None:
similarity_new = similarity_batch(
input_candidate, input_original)
prob_new_prob = prob_new_prob * similarity_new
prob_delete_prob = prob_new_prob
input_delete_ = input_[:ind] + input_[ind + 1:]
b = np.argmax([
prob_old_prob, prob_change_prob,
prob_changeanother_prob * 0.3, prob_add_prob * 0.1,
prob_delete_prob * 0.001
])
print([
prob_old_prob, prob_change_prob,
prob_changeanother_prob, prob_add_prob,
prob_delete_prob
])
print([
input_, input_change_, input_changeanother_,
input_add_, input_delete_
])
input_ = [
input_, input_change_, input_changeanother_,
input_add_, input_delete_
][b]
pos += 1
def main(_):
if os.path.exists(config.forward_log_path) and config.mode=='forward':
os.system('rm '+config.forward_log_path)
if os.path.exists(config.backward_log_path) and config.mode=='backward':
os.system('rm '+config.backward_log_path)
if os.path.exists(config.use_output_path):
os.system('rm '+config.use_output_path)
for item in config.record_time:
if os.path.exists(config.use_output_path+str(item)):
os.system('rm '+config.use_output_path+str(item))
if os.path.exists(config.use_log_path):
os.system('rm '+config.use_log_path)
if config.mode=='forward' or config.mode=='use':
with tf.name_scope("forward_train"):
with tf.variable_scope("forward", reuse=None):
m_forward = PTBModel(is_training=True)
with tf.name_scope("forward_test"):
with tf.variable_scope("forward", reuse=True):
mtest_forward = PTBModel(is_training=False)
var=tf.trainable_variables()
var_forward=[x for x in var if x.name.startswith('forward')]
saver_forward=tf.train.Saver(var_forward, max_to_keep=1)
if config.mode=='backward' or config.mode=='use':
with tf.name_scope("backward_train"):
with tf.variable_scope("backward", reuse=None):
m_backward = PTBModel(is_training=True)
with tf.name_scope("backward_test"):
with tf.variable_scope("backward", reuse=True):
mtest_backward = PTBModel(is_training=False)
var=tf.trainable_variables()
var_backward=[x for x in var if x.name.startswith('backward')]
saver_backward=tf.train.Saver(var_backward, max_to_keep=1)
init = tf.global_variables_initializer()
with tf.Session() as session:
session.run(init)
if config.mode=='forward':
train_data, test_data = reader.read_data(config.data_path, config.num_steps)
test_mean_old=15.0
for epoch in range(config.max_epoch):
train_ppl_list=[]
test_ppl_list=[]
for i in range(train_data.length//config.batch_size):
input, sequence_length, target=train_data(m_forward.batch_size, i)
train_perplexity = run_epoch(session, m_forward,input, sequence_length, target, mode='train')
train_ppl_list.append(train_perplexity)
print("Epoch:%d, Iter: %d Train NLL: %.3f" % (epoch, i + 1, train_perplexity))
for i in range(test_data.length//config.batch_size):
input, sequence_length, target=test_data(mtest_forward.batch_size, i)
test_perplexity = run_epoch(session, mtest_forward, input, sequence_length, target, mode='test')
test_ppl_list.append(test_perplexity)
print("Epoch:%d, Iter: %d Test NLL: %.3f" % (epoch, i + 1, test_perplexity))
test_mean=np.mean(test_ppl_list)
if test_mean<test_mean_old:
test_mean_old=test_mean
saver_forward.save(session, config.forward_save_path)
write_log('train ppl:'+str(np.mean(train_ppl_list))+'\t'+'test ppl:'+str(test_mean), config.forward_log_path)
if config.mode=='backward':
train_data, test_data = reader.read_data(config.data_path, config.num_steps)
test_mean_old=15.0
for epoch in range(config.max_epoch):
train_ppl_list=[]
test_ppl_list=[]
for i in range(train_data.length//config.batch_size):
input, sequence_length, target=train_data(m_backward.batch_size, i)
input, sequence_length, target=reverse_seq(input, sequence_length, target)
train_perplexity = run_epoch(session, m_backward,input, sequence_length, target, mode='train')
train_ppl_list.append(train_perplexity)
print("Epoch:%d, Iter: %d Train NLL: %.3f" % (epoch, i + 1, train_perplexity))
for i in range(test_data.length//config.batch_size):
input, sequence_length, target=test_data(mtest_backward.batch_size, i)
input, sequence_length, target=reverse_seq(input, sequence_length, target)
test_perplexity = run_epoch(session, mtest_backward, input, sequence_length, target, mode='test')
test_ppl_list.append(test_perplexity)
print("Epoch:%d, Iter: %d Test NLL: %.3f" % (epoch, i + 1, test_perplexity))
test_mean=np.mean(test_ppl_list)
if test_mean<test_mean_old:
test_mean_old=test_mean
saver_backward.save(session, config.backward_save_path)
write_log('train ppl:'+str(np.mean(train_ppl_list))+'\t'+'test ppl:'+str(test_mean), config.backward_log_path)
if config.mode=='use':
sim=config.sim
#keyword stable
sta_vec=list(np.zeros([config.num_steps-1]))
saver_forward.restore(session, config.forward_save_path)
saver_backward.restore(session, config.backward_save_path)
config.shuffle=False
if config.keyboard_input==True:
key_input=raw_input('please input a sentence in lower case\n')
if key_input=='':
use_data = reader.read_data_use(config.use_data_path, config.num_steps)
else:
key_input=key_input.split()
key_input=sen2id(key_input)
use_data = reader.array_data([key_input], config.num_steps, config.dict_size)
else:
use_data, sta_vec_list = reader.read_data_use(config.use_data_path, config.num_steps)
config.batch_size=1
#use_data.length=1 #######################################
for sen_id in range(use_data.length):
if config.keyboard_input==False:
sta_vec=sta_vec_list[sen_id%len(sta_vec)]
print(sta_vec)
input, sequence_length, _=use_data(1, sen_id)
input_original=input[0]
for i in range(1,config.num_steps):
if input[0][i]>config.rare_since and input[0][i]<config.dict_size:
sta_vec[i-1]=1
pos=0
for iter in range(config.sample_time):
#ind is the index of the selected word, regardless of the beginning token.
ind=pos%(sequence_length[0]-1)
action=choose_action(config.action_prob)
#tem
print(' '.join(id2sen(input[0])))
if iter in config.record_time:
with open(config.use_output_path+str(iter), 'a') as g:
g.write(' '.join(id2sen(input[0]))+'\n')
#tem_end
#print(sta_vec, sequence_length[0], ind)
'''
if sta_vec[ind]==1 and action in [0, 2]: #stop skipping words
action=3
'''
#change word
if action==0:
prob_old=run_epoch(session, mtest_forward, input, sequence_length, mode='use')
if config.double_LM==True:
input_backward, _, _ =reverse_seq(input, sequence_length, input)
prob_old=(prob_old+run_epoch(session, mtest_backward, input_backward, sequence_length, mode='use'))*0.5
tem=1
for j in range(sequence_length[0]-1):
tem*=prob_old[0][j][input[0][j+1]]
tem*=prob_old[0][j+1][config.dict_size+1]
prob_old_prob=tem
if sim!=None:
similarity_old=similarity(input[0], input_original, sta_vec)
prob_old_prob*=similarity_old
else:
similarity_old=-1
input_forward, input_backward, sequence_length_forward, sequence_length_backward = cut_from_point(input, sequence_length, ind, mode=action)
prob_forward=run_epoch(session, mtest_forward, input_forward, sequence_length_forward, mode='use')[0, ind%(sequence_length[0]-1),:]
prob_backward=run_epoch(session, mtest_backward, input_backward, sequence_length_backward, mode='use')[0, sequence_length[0]-1-ind%(sequence_length[0]-1),:]
prob_mul=(prob_forward*prob_backward)
input_candidate, sequence_length_candidate=generate_candidate_input(input, sequence_length, ind, prob_mul, config.search_size, mode=action)
prob_candidate_pre=run_epoch(session, mtest_forward, input_candidate, sequence_length_candidate, mode='use')
if config.double_LM==True:
input_candidate_backward, _, _ =reverse_seq(input_candidate, sequence_length_candidate, input_candidate)
prob_candidate_pre=(prob_candidate_pre+run_epoch(session, mtest_backward, input_candidate_backward, sequence_length_candidate, mode='use'))*0.5
prob_candidate=[]
for i in range(config.search_size):
tem=1
for j in range(sequence_length[0]-1):
tem*=prob_candidate_pre[i][j][input_candidate[i][j+1]]
tem*=prob_candidate_pre[i][j+1][config.dict_size+1]
prob_candidate.append(tem)
prob_candidate=np.array(prob_candidate)
#similarity_candidate=np.array([similarity(x, input_original) for x in input_candidate])
if sim!=None:
similarity_candidate=similarity_batch(input_candidate, input_original,sta_vec)
prob_candidate=prob_candidate*similarity_candidate
prob_candidate_norm=normalize(prob_candidate)
prob_candidate_ind=sample_from_candidate(prob_candidate_norm)
prob_candidate_prob=prob_candidate[prob_candidate_ind]
if input_candidate[prob_candidate_ind][ind+1]<config.dict_size and ( prob_candidate_prob>prob_old_prob*config.threshold or just_acc()==0):
input=input_candidate[prob_candidate_ind:prob_candidate_ind+1]
pos+=1
#old_place=len(prob_mul)-list(np.argsort(prob_mul)).index(input[0][ind+1])
#write_log('step:'+str(iter)+'action:0 prob_old:'+str(prob_old_prob)+' prob_new:'+str(prob_candidate_prob)+' '+str(old_place)+' '+str(sta_vec.index(1))+' '+str(ind), config.use_log_path)
print('action:0', 1, prob_old_prob, prob_candidate_prob, prob_candidate_norm[prob_candidate_ind], similarity_old)
#add word
if action==1:
if sequence_length[0]>=config.num_steps:
action=3
else:
input_forward, input_backward, sequence_length_forward, sequence_length_backward = cut_from_point(input, sequence_length, ind, mode=action)
prob_forward=run_epoch(session, mtest_forward, input_forward, sequence_length_forward, mode='use')[0, ind%(sequence_length[0]-1),:]
prob_backward=run_epoch(session, mtest_backward, input_backward, sequence_length_backward, mode='use')[0, sequence_length[0]-1-ind%(sequence_length[0]-1),:]
prob_mul=(prob_forward*prob_backward)
input_candidate, sequence_length_candidate=generate_candidate_input(input, sequence_length, ind, prob_mul, config.search_size, mode=action)
prob_candidate_pre=run_epoch(session, mtest_forward, input_candidate, sequence_length_candidate, mode='use')
if config.double_LM==True:
input_candidate_backward, _, _ =reverse_seq(input_candidate, sequence_length_candidate, input_candidate)
prob_candidate_pre=(prob_candidate_pre+run_epoch(session, mtest_backward, input_candidate_backward, sequence_length_candidate, mode='use'))*0.5
prob_candidate=[]
for i in range(config.search_size):
tem=1
for j in range(sequence_length_candidate[0]-1):
tem*=prob_candidate_pre[i][j][input_candidate[i][j+1]]
tem*=prob_candidate_pre[i][j+1][config.dict_size+1]
prob_candidate.append(tem)
prob_candidate=np.array(prob_candidate)
#similarity_candidate=np.array([similarity(x, input_original) for x in input_candidate])
if sim!=None:
similarity_candidate=similarity_batch(input_candidate, input_original,sta_vec)
prob_candidate=prob_candidate*similarity_candidate
prob_candidate_norm=normalize(prob_candidate)
prob_candidate_ind=sample_from_candidate(prob_candidate_norm)
prob_candidate_prob=prob_candidate[prob_candidate_ind]
prob_old=run_epoch(session, mtest_forward, input, sequence_length, mode='use')
if config.double_LM==True:
input_backward, _, _ =reverse_seq(input, sequence_length, input)
prob_old=(prob_old+run_epoch(session, mtest_backward, input_backward, sequence_length, mode='use'))*0.5
tem=1
for j in range(sequence_length[0]-1):
tem*=prob_old[0][j][input[0][j+1]]
tem*=prob_old[0][j+1][config.dict_size+1]
prob_old_prob=tem
if sim!=None:
similarity_old=similarity(input[0], input_original,sta_vec)
prob_old_prob=prob_old_prob*similarity_old
else:
similarity_old=-1
alpha=min(1, prob_candidate_prob*config.action_prob[2]/(prob_old_prob*config.action_prob[1]*prob_candidate_norm[prob_candidate_ind]))
#alpha=min(1, prob_candidate_prob*config.action_prob[2]/(prob_old_prob*config.action_prob[1]))
print ('action:1',alpha, prob_old_prob,prob_candidate_prob, prob_candidate_norm[prob_candidate_ind], similarity_old)
if choose_action([alpha, 1-alpha])==0 and input_candidate[prob_candidate_ind][ind]<config.dict_size and (prob_candidate_prob>prob_old_prob* config.threshold or just_acc()==0):
#write_log('step:'+str(iter)+'action:1 prob_old:'+str(prob_old_prob)+' prob_new:'+str(prob_candidate_prob)+' '+str(sta_vec.index(1))+' '+str(ind), config.use_log_path)
input=input_candidate[prob_candidate_ind:prob_candidate_ind+1]
sequence_length+=1
pos+=2
sta_vec.insert(ind, 0.0)
del(sta_vec[-1])
else:
action=3
#delete word
if action==2:
if sequence_length[0]<=2:
action=3
else:
prob_old=run_epoch(session, mtest_forward, input, sequence_length, mode='use')
if config.double_LM==True:
input_backward, _, _ =reverse_seq(input, sequence_length, input)
prob_old=(prob_old+run_epoch(session, mtest_backward, input_backward, sequence_length, mode='use'))*0.5
tem=1
for j in range(sequence_length[0]-1):
tem*=prob_old[0][j][input[0][j+1]]
tem*=prob_old[0][j+1][config.dict_size+1]
prob_old_prob=tem
if sim!=None:
similarity_old=similarity(input[0], input_original,sta_vec)
prob_old_prob=prob_old_prob*similarity_old
else:
similarity_old=-1
input_candidate, sequence_length_candidate=generate_candidate_input(input, sequence_length, ind, None , config.search_size, mode=2)
prob_new=run_epoch(session, mtest_forward, input_candidate, sequence_length_candidate, mode='use')
tem=1
for j in range(sequence_length_candidate[0]-1):
tem*=prob_new[0][j][input_candidate[0][j+1]]
tem*=prob_new[0][j+1][config.dict_size+1]
prob_new_prob=tem
if sim!=None:
similarity_new=similarity_batch(input_candidate, input_original,sta_vec)
prob_new_prob=prob_new_prob*similarity_new
input_forward, input_backward, sequence_length_forward, sequence_length_backward = cut_from_point(input, sequence_length, ind, mode=0)
prob_forward=run_epoch(session, mtest_forward, input_forward, sequence_length_forward, mode='use')[0, ind%(sequence_length[0]-1),:]
prob_backward=run_epoch(session, mtest_backward, input_backward, sequence_length_backward, mode='use')[0, sequence_length[0]-1-ind%(sequence_length[0]-1),:]
prob_mul=(prob_forward*prob_backward)
input_candidate, sequence_length_candidate=generate_candidate_input(input, sequence_length, ind, prob_mul, config.search_size, mode=0)
prob_candidate_pre=run_epoch(session, mtest_forward, input_candidate, sequence_length_candidate, mode='use')
if config.double_LM==True:
input_candidate_backward, _, _ =reverse_seq(input_candidate, sequence_length_candidate, input_candidate)
prob_candidate_pre=(prob_candidate_pre+run_epoch(session, mtest_backward, input_candidate_backward, sequence_length_candidate, mode='use'))*0.5
prob_candidate=[]
for i in range(config.search_size):
tem=1
for j in range(sequence_length[0]-1):
tem*=prob_candidate_pre[i][j][input_candidate[i][j+1]]
tem*=prob_candidate_pre[i][j+1][config.dict_size+1]
prob_candidate.append(tem)
prob_candidate=np.array(prob_candidate)
#similarity_candidate=np.array([similarity(x, input_original) for x in input_candidate])
if sim!=None:
similarity_candidate=similarity_batch(input_candidate, input_original,sta_vec)
prob_candidate=prob_candidate*similarity_candidate
#####There is a unsolved problem
prob_candidate_norm=normalize(prob_candidate)
if input[0] in input_candidate:
for candidate_ind in range(len(input_candidate)):
if input[0] in input_candidate[candidate_ind: candidate_ind+1]:
break
pass
alpha=min(prob_candidate_norm[candidate_ind]*prob_new_prob*config.action_prob[1]/(config.action_prob[2]*prob_old_prob), 1)
else:
pass
alpha=0
#alpha=min(prob_new_prob*config.action_prob[1]/(config.action_prob[2]*prob_old_prob), 1)
print('action:2', alpha, prob_old_prob, prob_new_prob, prob_candidate_norm[candidate_ind], similarity_old)
if choose_action([alpha, 1-alpha])==0 and (prob_new_prob> prob_old_prob*config.threshold or just_acc()==0):
#write_log('step:'+str(iter)+'action:2 prob_old:'+str(prob_old_prob)+' prob_new:'+str(prob_new_prob)+' '+str(sta_vec.index(1))+' '+str(ind), config.use_log_path)
input=np.concatenate([input[:,:ind+1], input[:,ind+2:], input[:,:1]*0+config.dict_size+1], axis=1)
sequence_length-=1
pos+=0
del(sta_vec[ind])
sta_vec.append(0)
else:
action=3
#do nothing
if action==3:
#write_log('step:'+str(iter)+'action:3', config.use_log_path)
pos+=1
def main(_):
if os.path.exists(config.forward_log_path) and config.mode == 'forward':
os.system('rm ' + config.forward_log_path)
if os.path.exists(config.backward_log_path) and config.mode == 'backward':
os.system('rm ' + config.backward_log_path)
if os.path.exists(config.use_output_path):
os.system('rm ' + config.use_output_path)
if os.path.exists(config.use_output_path):
os.system('rm ' + config.use_output_path)
if os.path.exists(config.use_log_path):
os.system('rm ' + config.use_log_path)
if config.mode == 'forward' or config.mode == 'use':
with tf.name_scope("forward_train"):
with tf.variable_scope("forward", reuse=None):
m_forward = PTBModel(is_training=True)
with tf.name_scope("forward_test"):
with tf.variable_scope("forward", reuse=True):
mtest_forward = PTBModel(is_training=False)
var = tf.trainable_variables()
var_forward = [x for x in var if x.name.startswith('forward')]
saver_forward = tf.train.Saver(var_forward, max_to_keep=1)
if config.mode == 'backward' or config.mode == 'use':
with tf.name_scope("backward_train"):
with tf.variable_scope("backward", reuse=None):
m_backward = PTBModel(is_training=True)
with tf.name_scope("backward_test"):
with tf.variable_scope("backward", reuse=True):
mtest_backward = PTBModel(is_training=False)
var = tf.trainable_variables()
var_backward = [x for x in var if x.name.startswith('backward')]
saver_backward = tf.train.Saver(var_backward, max_to_keep=1)
init = tf.global_variables_initializer()
configs = tf.ConfigProto()
configs.gpu_options.allow_growth = True
with tf.Session(config=configs) as session:
session.run(init)
if config.mode == 'forward':
#train forward language model
train_data, test_data = reader.read_data(config.data_path,
config.num_steps)
test_mean_old = 15.0
for epoch in range(config.max_epoch):
train_ppl_list = []
test_ppl_list = []
for i in range(train_data.length // config.batch_size):
input, sequence_length, target = train_data(
m_forward.batch_size, i)
train_perplexity = run_epoch(session,
m_forward,
input,
sequence_length,
target,
mode='train')
train_ppl_list.append(train_perplexity)
print("Epoch:%d, Iter: %d Train NLL: %.3f" %
(epoch, i + 1, train_perplexity))
for i in range(test_data.length // config.batch_size):
input, sequence_length, target = test_data(
mtest_forward.batch_size, i)
test_perplexity = run_epoch(session,
mtest_forward,
input,
sequence_length,
target,
mode='test')
test_ppl_list.append(test_perplexity)
print("Epoch:%d, Iter: %d Test NLL: %.3f" %
(epoch, i + 1, test_perplexity))
test_mean = np.mean(test_ppl_list)
if test_mean < test_mean_old:
test_mean_old = test_mean
saver_forward.save(session, config.forward_save_path)
write_log(
'train ppl:' + str(np.mean(train_ppl_list)) + '\t' +
'test ppl:' + str(test_mean), config.forward_log_path)
if config.mode == 'backward':
#train backward language model
train_data, test_data = reader.read_data(config.data_path,
config.num_steps)
test_mean_old = 15.0
for epoch in range(config.max_epoch):
train_ppl_list = []
test_ppl_list = []
for i in range(train_data.length // config.batch_size):
input, sequence_length, target = train_data(
m_backward.batch_size, i)
input, sequence_length, target = reverse_seq(
input, sequence_length, target)
train_perplexity = run_epoch(session,
m_backward,
input,
sequence_length,
target,
mode='train')
train_ppl_list.append(train_perplexity)
print("Epoch:%d, Iter: %d Train NLL: %.3f" %
(epoch, i + 1, train_perplexity))
for i in range(test_data.length // config.batch_size):
input, sequence_length, target = test_data(
mtest_backward.batch_size, i)
input, sequence_length, target = reverse_seq(
input, sequence_length, target)
test_perplexity = run_epoch(session,
mtest_backward,
input,
sequence_length,
target,
mode='test')
test_ppl_list.append(test_perplexity)
print("Epoch:%d, Iter: %d Test NLL: %.3f" %
(epoch, i + 1, test_perplexity))
test_mean = np.mean(test_ppl_list)
if test_mean < test_mean_old:
test_mean_old = test_mean
saver_backward.save(session, config.backward_save_path)
write_log(
'train ppl:' + str(np.mean(train_ppl_list)) + '\t' +
'test ppl:' + str(test_mean), config.backward_log_path)
if config.mode == 'use':
#CGMH sampling for key_gen
sim = config.sim
saver_forward.restore(session, config.forward_save_path)
saver_backward.restore(session, config.backward_save_path)
config.shuffle = False
#keyword input
if config.keyboard_input == True:
#input from keyboard if key_input is not empty
key_input = raw_input('please input a sentence\n')
if key_input == '':
use_data = reader.read_data_use(config.use_data_path,
config.num_steps)
else:
key_input = key_input.split()
key_input = sen2id(key_input)
sta_vec = list(np.zeros([config.num_steps - 1]))
for i in range(len(key_input)):
sta_vec[i] = 1
use_data = reader.array_data([key_input], config.num_steps,
config.dict_size)
else:
#load keywords from file
use_data, sta_vec_list = reader.read_data_use(
config.use_data_path, config.num_steps)
config.batch_size = 1
for sen_id in range(use_data.length):
#generate for each sequence of keywords
if config.keyboard_input == False:
sta_vec = sta_vec_list[sen_id % (config.num_steps - 1)]
print(sta_vec)
input, sequence_length, _ = use_data(1, sen_id)
input_original = input[0]
pos = 0
outputs = []
output_p = []
for iter in range(config.sample_time):
#ind is the index of the selected word, regardless of the beginning token.
#sample config.sample_time times for each set of keywords
config.sample_prior = [1, 10.0 / sequence_length[0], 1, 1]
if iter % 20 < 10:
config.threshold = 0
else:
config.threshold = 0.5
ind = pos % (sequence_length[0])
action = choose_action(config.action_prob)
print(' '.join(id2sen(input[0])))
if sta_vec[ind] == 1 and action in [0, 2]:
#skip words that we do not change(original keywords)
action = 3
#word replacement (action: 0)
if action == 0 and ind < sequence_length[0] - 1:
prob_old = run_epoch(session,
mtest_forward,
input,
sequence_length,
mode='use')
if config.double_LM == True:
input_backward, _, _ = reverse_seq(
input, sequence_length, input)
prob_old = (prob_old + run_epoch(session,
mtest_backward,
input_backward,
sequence_length,
mode='use')) * 0.5
tem = 1
for j in range(sequence_length[0] - 1):
tem *= prob_old[0][j][input[0][j + 1]]
tem *= prob_old[0][j + 1][config.dict_size + 1]
prob_old_prob = tem
if sim != None:
similarity_old = similarity(
input[0], input_original, sta_vec)
prob_old_prob *= similarity_old
else:
similarity_old = -1
input_forward, input_backward, sequence_length_forward, sequence_length_backward = cut_from_point(
input, sequence_length, ind, mode=action)
prob_forward = run_epoch(
session,
mtest_forward,
input_forward,
sequence_length_forward,
mode='use')[0, ind % (sequence_length[0] - 1), :]
prob_backward = run_epoch(
session,
mtest_backward,
input_backward,
sequence_length_backward,
mode='use')[0, sequence_length[0] - 1 - ind %
(sequence_length[0] - 1), :]
prob_mul = (prob_forward * prob_backward)
input_candidate, sequence_length_candidate = generate_candidate_input(
input,
sequence_length,
ind,
prob_mul,
config.search_size,
mode=action)
prob_candidate_pre = run_epoch(
session,
mtest_forward,
input_candidate,
sequence_length_candidate,
mode='use')
if config.double_LM == True:
input_candidate_backward, _, _ = reverse_seq(
input_candidate, sequence_length_candidate,
input_candidate)
prob_candidate_pre = (
prob_candidate_pre +
run_epoch(session,
mtest_backward,
input_candidate_backward,
sequence_length_candidate,
mode='use')) * 0.5
prob_candidate = []
for i in range(config.search_size):
tem = 1
for j in range(sequence_length[0] - 1):
tem *= prob_candidate_pre[i][j][
input_candidate[i][j + 1]]
tem *= prob_candidate_pre[i][j +
1][config.dict_size +
1]
prob_candidate.append(tem)
prob_candidate = np.array(prob_candidate)
if sim != None:
similarity_candidate = similarity_batch(
input_candidate, input_original, sta_vec)
prob_candidate = prob_candidate * similarity_candidate
prob_candidate_norm = normalize(prob_candidate)
prob_candidate_ind = sample_from_candidate(
prob_candidate_norm)
prob_candidate_prob = prob_candidate[
prob_candidate_ind]
if input_candidate[prob_candidate_ind][
ind + 1] < config.dict_size and (
prob_candidate_prob > prob_old_prob *
config.threshold or just_acc() == 0):
input = input_candidate[
prob_candidate_ind:prob_candidate_ind + 1]
pos += 1
print('action:0', 1, prob_old_prob,
prob_candidate_prob,
prob_candidate_norm[prob_candidate_ind],
similarity_old)
if ' '.join(id2sen(input[0])) not in output_p:
outputs.append(
[' '.join(id2sen(input[0])), prob_old_prob])
#word insertion(action:1)
if action == 1:
if sequence_length[0] >= config.num_steps:
action = 3
else:
input_forward, input_backward, sequence_length_forward, sequence_length_backward = cut_from_point(
input, sequence_length, ind, mode=action)
prob_forward = run_epoch(
session,
mtest_forward,
input_forward,
sequence_length_forward,
mode='use')[0,
ind % (sequence_length[0] - 1), :]
prob_backward = run_epoch(
session,
mtest_backward,
input_backward,
sequence_length_backward,
mode='use')[0, sequence_length[0] - 1 - ind %
(sequence_length[0] - 1), :]
prob_mul = (prob_forward * prob_backward)
input_candidate, sequence_length_candidate = generate_candidate_input(
input,
sequence_length,
ind,
prob_mul,
config.search_size,
mode=action)
prob_candidate_pre = run_epoch(
session,
mtest_forward,
input_candidate,
sequence_length_candidate,
mode='use')
if config.double_LM == True:
input_candidate_backward, _, _ = reverse_seq(
input_candidate, sequence_length_candidate,
input_candidate)
prob_candidate_pre = (
prob_candidate_pre +
run_epoch(session,
mtest_backward,
input_candidate_backward,
sequence_length_candidate,
mode='use')) * 0.5
prob_candidate = []
for i in range(config.search_size):
tem = 1
for j in range(sequence_length_candidate[0] -
1):
tem *= prob_candidate_pre[i][j][
input_candidate[i][j + 1]]
tem *= prob_candidate_pre[i][j + 1][
config.dict_size + 1]
prob_candidate.append(tem)
prob_candidate = np.array(
prob_candidate) * config.sample_prior[1]
if sim != None:
similarity_candidate = similarity_batch(
input_candidate, input_original, sta_vec)
prob_candidate = prob_candidate * similarity_candidate
prob_candidate_norm = normalize(prob_candidate)
prob_candidate_ind = sample_from_candidate(
prob_candidate_norm)
prob_candidate_prob = prob_candidate[
prob_candidate_ind]
prob_old = run_epoch(session,
mtest_forward,
input,
sequence_length,
mode='use')
if config.double_LM == True:
input_backward, _, _ = reverse_seq(
input, sequence_length, input)
prob_old = (prob_old +
run_epoch(session,
mtest_backward,
input_backward,
sequence_length,
mode='use')) * 0.5
tem = 1
for j in range(sequence_length[0] - 1):
tem *= prob_old[0][j][input[0][j + 1]]
tem *= prob_old[0][j + 1][config.dict_size + 1]
prob_old_prob = tem
if sim != None:
similarity_old = similarity(
input[0], input_original, sta_vec)
prob_old_prob = prob_old_prob * similarity_old
else:
similarity_old = -1
#alpha is acceptance ratio of current proposal
alpha = min(
1,
prob_candidate_prob * config.action_prob[2] /
(prob_old_prob * config.action_prob[1] *
prob_candidate_norm[prob_candidate_ind]))
print('action:1', alpha, prob_old_prob,
prob_candidate_prob,
prob_candidate_norm[prob_candidate_ind],
similarity_old)
if ' '.join(id2sen(input[0])) not in output_p:
outputs.append([
' '.join(id2sen(input[0])), prob_old_prob
])
if choose_action([
alpha, 1 - alpha
]) == 0 and input_candidate[prob_candidate_ind][
ind + 1] < config.dict_size and (
prob_candidate_prob > prob_old_prob *
config.threshold or just_acc() == 0):
input = input_candidate[
prob_candidate_ind:prob_candidate_ind + 1]
sequence_length += 1
pos += 2
sta_vec.insert(ind, 0.0)
del (sta_vec[-1])
else:
action = 3
#word deletion(action: 2)
if action == 2 and ind < sequence_length[0] - 1:
if sequence_length[0] <= 2:
action = 3
else:
prob_old = run_epoch(session,
mtest_forward,
input,
sequence_length,
mode='use')
if config.double_LM == True:
input_backward, _, _ = reverse_seq(
input, sequence_length, input)
prob_old = (prob_old +
run_epoch(session,
mtest_backward,
input_backward,
sequence_length,
mode='use')) * 0.5
tem = 1
for j in range(sequence_length[0] - 1):
tem *= prob_old[0][j][input[0][j + 1]]
tem *= prob_old[0][j + 1][config.dict_size + 1]
prob_old_prob = tem
if sim != None:
similarity_old = similarity(
input[0], input_original, sta_vec)
prob_old_prob = prob_old_prob * similarity_old
else:
similarity_old = -1
input_candidate, sequence_length_candidate = generate_candidate_input(
input,
sequence_length,
ind,
None,
config.search_size,
mode=2)
prob_new = run_epoch(session,
mtest_forward,
input_candidate,
sequence_length_candidate,
mode='use')
tem = 1
for j in range(sequence_length_candidate[0] - 1):
tem *= prob_new[0][j][input_candidate[0][j +
1]]
tem *= prob_new[0][j + 1][config.dict_size + 1]
prob_new_prob = tem
if sim != None:
similarity_new = similarity_batch(
input_candidate, input_original, sta_vec)
prob_new_prob = prob_new_prob * similarity_new
input_forward, input_backward, sequence_length_forward, sequence_length_backward = cut_from_point(
input, sequence_length, ind, mode=0)
prob_forward = run_epoch(
session,
mtest_forward,
input_forward,
sequence_length_forward,
mode='use')[0,
ind % (sequence_length[0] - 1), :]
prob_backward = run_epoch(
session,
mtest_backward,
input_backward,
sequence_length_backward,
mode='use')[0, sequence_length[0] - 1 - ind %
(sequence_length[0] - 1), :]
prob_mul = (prob_forward * prob_backward)
input_candidate, sequence_length_candidate = generate_candidate_input(
input,
sequence_length,
ind,
prob_mul,
config.search_size,
mode=0)
prob_candidate_pre = run_epoch(
session,
mtest_forward,
input_candidate,
sequence_length_candidate,
mode='use')
if config.double_LM == True:
input_candidate_backward, _, _ = reverse_seq(
input_candidate, sequence_length_candidate,
input_candidate)
prob_candidate_pre = (
prob_candidate_pre +
run_epoch(session,
mtest_backward,
input_candidate_backward,
sequence_length_candidate,
mode='use')) * 0.5
prob_candidate = []
for i in range(config.search_size):
tem = 1
for j in range(sequence_length[0] - 1):
tem *= prob_candidate_pre[i][j][
input_candidate[i][j + 1]]
tem *= prob_candidate_pre[i][j + 1][
config.dict_size + 1]
prob_candidate.append(tem)
prob_candidate = np.array(prob_candidate)
if sim != None:
similarity_candidate = similarity_batch(
input_candidate, input_original, sta_vec)
prob_candidate = prob_candidate * similarity_candidate
#alpha is acceptance ratio of current proposal
prob_candidate_norm = normalize(prob_candidate)
if input[0] in input_candidate:
for candidate_ind in range(
len(input_candidate)):
if input[0] in input_candidate[
candidate_ind:candidate_ind + 1]:
break
pass
alpha = min(
prob_candidate_norm[candidate_ind] *
prob_new_prob * config.action_prob[1] /
(config.action_prob[2] * prob_old_prob), 1)
else:
pass
alpha = 0
print('action:2', alpha, prob_old_prob,
prob_new_prob,
prob_candidate_norm[candidate_ind],
similarity_old)
if ' '.join(id2sen(input[0])) not in output_p:
outputs.append([
' '.join(id2sen(input[0])), prob_old_prob
])
if choose_action([
alpha, 1 - alpha
]) == 0 and (prob_new_prob > prob_old_prob *
config.threshold or just_acc() == 0):
input = np.concatenate([
input[:, :ind + 1], input[:, ind + 2:],
input[:, :1] * 0 + config.dict_size + 1
],
axis=1)
sequence_length -= 1
pos += 0
del (sta_vec[ind])
sta_vec.append(0)
else:
action = 3
#skip word (action: 3)
if action == 3:
#write_log('step:'+str(iter)+'action:3', config.use_log_path)
pos += 1
print(outputs)
if outputs != []:
output_p.append(outputs[-1][0])
#choose output from samples
for num in range(config.min_length, 0, -1):
outputss = [x for x in outputs if len(x[0].split()) >= num]
print(num, outputss)
if outputss != []:
break
if outputss == []:
outputss.append([' '.join(id2sen(input[0])), 1])
outputss = sorted(outputss, key=lambda x: x[1])[::-1]
with open(config.use_output_path, 'a') as g:
g.write(outputss[0][0] + '\n')