我是theano的初学者。我正在研究它。我想打印“价值”＆＃39;和＃＆＃39;塑造＆＃39;在操作theano.function时的TensorVariable。当我使用python的print函数时，打印函数在编译theano函数之前运行。所以我学会了使用print功能是没用的。我试着另一只手。我添加了以下语法来执行theano.printing.Print。
（cce是theano.scan的返回值。因此，也许它不是符号可变的。 实际上，我对TensorVariable和共享变量的概念感到困惑。 TensorVariable是一种共享变量？）
x = theano.tensor.tensor3() # define data type
t_print =theano.printing.Print("cce value is : ")(x)
f = theano.function([x], t_print) # define theano.function
f(cce) # call f (print value of cce)
然后，出现以下错误
TypeError: ('Bad input argument to theano function with name "seq2seq.py : 98" at index 0(0-based)', 'Expected an array-like object, but found a Variable: maybe you are trying to call a function on a (possibly shared) variable instead of a numeric array?')
您是否可以让我知道如何更正此代码以打印cce（TensorVariable）的值？或者，当theano.function正在进行时，是否无法打印TensorVariable的值？
感谢您阅读我的问题。
这是我的源代码。这是一幅大图。 theano.function（）从最后一行开始。 loss_func是＆quot; categorical_crossentropy function＆＃39;。最后4行是关于theano功能
def categorical_crossentropy(y_true, y_pred):
y_pred = T.clip(y_pred, epsilon, 1.0 - epsilon)
y_pred = y_pred.reshape( (-1, voca_dim_g) )
y_true = y_true.reshape( (-1, voca_dim_g) )
cce, updates = theano.scan(
fn=T.nnet.categorical_crossentropy,
sequences=[y_pred,y_true]
)
##### I want to print cce HERE #######
return T.mean(cce)
@staticmethod
def step(
x_t, h_tm1, c_tm1,
Ui, Wi, bi, Uf, Wf, bf,
Uo, Wo, bo, Ug, Wg, bg
):
"""
x_t.shape = (timestep=1, dim)
x_t.shape = (n_samples, timestep=1, dim)
"""
i_t = T.nnet.sigmoid(T.dot(x_t, Ui) + T.dot(h_tm1, Wi) + bi)
f_t = T.nnet.sigmoid(T.dot(x_t, Uf) + T.dot(h_tm1, Wf) + bf)
o_t = T.nnet.sigmoid(T.dot(x_t, Uo) + T.dot(h_tm1, Wo) + bo)
g_t = T.tanh(T.dot(x_t, Ug) + T.dot(h_tm1, Wg) + bg)
c_t = c_tm1 * f_t + g_t * i_t
h_t = T.tanh(c_t) * o_t
return h_t, c_t
#########################################################################################################################
def forward(self, X):
states, updates = theano.scan(
fn=self.step,
sequences=[ X ],
outputs_info=[self.h_tm1, self.c_tm1],
non_sequences=[
self.Ui, self.Wi, self.bi,
self.Uf, self.Wf, self.bf,
self.Uo, self.Wo, self.bo,
self.Ug, self.Wg, self.bg
]
)
updates = [(self.h_tm1, states[0][-1]), (self.c_tm1, states[1][-1])]
return states, updates
#########################################################################################################################
def encode(self, X):
states, updates = self.forward(X)
h_t = states[0][-1]
c_t = states[1][-1]
return h_t, c_t, updates
def decode_step(
self, y_t, h_tm1, c_tm1,
Ui, Wi, bi, Uf, Wf, bf,
Uo, Wo, bo, Ug, Wg, bg,
Wh, bh
):
h_t, c_t = self.step(
y_t, h_tm1, c_tm1,
Ui, Wi, bi, Uf, Wf, bf,
Uo, Wo, bo, Ug, Wg, bg
)
y_t = T.dot(h_t, Wh) + bh
return y_t, h_t, c_t
def decode(self, h_tm1, c_tm1, timesteps):
outputs, updates = theano.scan(
fn=self.decode_step,
outputs_info=[self.y_t, h_tm1, c_tm1],
non_sequences=[
self.Ui, self.Wi, self.bi,
self.Uf, self.Wf, self.bf,
self.Uo, self.Wo, self.bo,
self.Ug, self.Wg, self.bg,
self.Wh, self.bh
],
n_steps=timesteps
)
updates = [
(self.h_tm1, outputs[1][-1]),
(self.c_tm1, outputs[2][-1])
]
return outputs[0], updates
h_tm1, c_tm1, updates_encode = encode(seq_input)
seq_predict, updates_decode = decode(h_tm1, c_tm1, T.shape(seq_target)[0])
loss = loss_func(seq_predict, seq_target)
self._train = theano.function([seq_input, seq_target], loss, updates = updates)
下面是完整的源代码
# -*- coding: utf-8 -*-
__modifier__ = "Lee Guk Beom, Lee Jae Sang, Jang Jae Kwang (alphabetical Order)"
import readFile
import numpy as np
import theano
import theano.tensor as T
from six.moves import zip
from theano.compile.debugmode import DebugMode
import nltk
import sys
import os
from nltk.tokenize import sent_tokenize
import codecs
#theano.config.optimizer='fast_compile'
#theano.config.exception_verbosity='high'
#theano.config.compute_test_value = 'warn'
epsilon = 1e-6
dtype = theano.config.floatX
minibatch_size_g = 0
longest_seq_g = 0
voca_dim_g = 0
n_time_step_input_g = 0
n_timestep_target_g = 0
word_to_index_input_g = dict()
word_to_index_targrt_g = dict()
index_to_word_target_g = dict()
#########################################################################################################################
def shared(value, name=None):
return theano.shared(value.astype(dtype), name=name)
#########################################################################################################################
def shared_zeros(shape, name=None):
return shared(value=np.zeros(shape), name=name)
#########################################################################################################################
def shared_zeros_like(x, name=None):
return shared_zeros(shape=x.shape, name=name)
#########################################################################################################################
def init_weights(shape, name=None):
bound = np.sqrt(1.0/shape[1])
w = np.random.uniform(-bound, bound, shape)
return shared(value=w, name=name)
#########################################################################################################################
def adadelta(params, cost, lr=1.0, rho=0.95):
# from https://github.com/fchollet/keras/blob/master/keras/optimizers.py
cost = cost.astype('float32')
grads = T.grad(cost, params)
accus = [shared_zeros_like(p.get_value()) for p in params]
delta_accus = [shared_zeros_like(p.get_value()) for p in params]
updates = []
for p, g, a, d_a in zip(params, grads, accus, delta_accus):
new_a = rho * a + (1.0 - rho) * T.square(g)
updates.append((a, new_a))
update = g * T.sqrt(d_a + epsilon) / T.sqrt(new_a + epsilon)
new_p = p - lr * update
updates.append((p, new_p))
new_d_a = rho * d_a + (1.0 - rho) * T.square(update)
updates.append((d_a, new_d_a))
return updates
#########################################################################################################################
def categorical_crossentropy(y_true, y_pred):
# from https://github.com/fchollet/keras/blob/master/keras/objectives.py
y_pred = T.clip(y_pred, epsilon, 1.0 - epsilon)
# y_true = y_true.reshape( (-1, minibatch_size_g, voca_dim_g) )
'''
cce = T.nnet.categorical_crossentropy(y_pred,y_true)
# only matrix can be calculated
'''
# Y_PRED SOFTMAX
y_pred = y_pred.reshape( (-1, voca_dim_g) )
# y_pred_flat = T.nnet.softmax(y_pred)
y_true = y_true.reshape( (-1, voca_dim_g) )
cce, updates = theano.scan(
fn=T.nnet.categorical_crossentropy,
sequences=[y_pred,y_true]
)
return T.mean(cce)
#########################################################################################################################
def mean_square_error(y_true, y_pred):
return T.mean(T.square(y_pred - y_true))
#########################################################################################################################
class LSTM(object):
def __init__(self, size, dim):
self.size = size
self.dim = dim
shape_b = (minibatch_size_g, size)
shape_U = (dim, size)
shape_W = (size, size)
self.h_tm1 = shared_zeros(shape_b, "h_tm1")
self.c_tm1 = shared_zeros(shape_b, "c_tm1")
self.Ui = init_weights(shape_U, "Ui")
self.Wi = init_weights(shape_W, "Wi")
self.bi = shared_zeros(shape_b, "bi")
self.Uf = init_weights(shape_U, "Uf")
self.Wf = init_weights(shape_W, "Wf")
self.bf = shared_zeros(shape_b, "bf")
self.Uo = init_weights(shape_U, "Uo")
self.Wo = init_weights(shape_W, "Wo")
self.bo = shared_zeros(shape_b, "bo")
self.Ug = init_weights(shape_U, "Ug")
self.Wg = init_weights(shape_W, "Wg")
self.bg = shared_zeros(shape_b, "bg")
self.params = [
self.Ui, self.Wi, self.bi,
self.Uf, self.Wf, self.bf,
self.Uo, self.Wo, self.bo,
self.Ug, self.Wg, self.bg
]
def set_state(self, h, c):
self.h_tm1.set_value(h.get_value())
self.c_tm1.set_value(c.get_value())
def reset_state(self):
self.h_tm1 = shared_zeros((1, self.size), "h_tm1")
self.c_tm1 = shared_zeros((1, self.size), "c_tm1")
#########################################################################################################################
@staticmethod
def step(
x_t, h_tm1, c_tm1,
Ui, Wi, bi, Uf, Wf, bf,
Uo, Wo, bo, Ug, Wg, bg
):
"""
x_t.shape = (timestep=1, dim)
x_t.shape = (n_samples, timestep=1, dim)
"""
i_t = T.nnet.sigmoid(T.dot(x_t, Ui) + T.dot(h_tm1, Wi) + bi)
f_t = T.nnet.sigmoid(T.dot(x_t, Uf) + T.dot(h_tm1, Wf) + bf)
o_t = T.nnet.sigmoid(T.dot(x_t, Uo) + T.dot(h_tm1, Wo) + bo)
g_t = T.tanh(T.dot(x_t, Ug) + T.dot(h_tm1, Wg) + bg)
c_t = c_tm1 * f_t + g_t * i_t
h_t = T.tanh(c_t) * o_t
return h_t, c_t
#########################################################################################################################
def forward(self, X):
states, updates = theano.scan(
fn=self.step,
sequences=[ X ],
outputs_info=[self.h_tm1, self.c_tm1],
non_sequences=[
self.Ui, self.Wi, self.bi,
self.Uf, self.Wf, self.bf,
self.Uo, self.Wo, self.bo,
self.Ug, self.Wg, self.bg
]
)
updates = [(self.h_tm1, states[0][-1]), (self.c_tm1, states[1][-1])]
return states, updates
#########################################################################################################################
class LSTMEncoder(LSTM):
def encode(self, X):
states, updates = self.forward(X)
h_t = states[0][-1]
c_t = states[1][-1]
return h_t, c_t, updates
class LSTMDecoder(LSTM):
def __init__(self, size, dim, h_tm1=None, c_tm1=None):
super(LSTMDecoder, self).__init__(size=size, dim=dim)
self.Wh = init_weights((size, dim), "Wh")
self.bh = shared_zeros((minibatch_size_g, dim), "bh")
self.h_tm1 = h_tm1 or shared_zeros((minibatch_size_g, size), "h_tm1")
self.c_tm1 = c_tm1 or shared_zeros((minibatch_size_g, size), "c_tm1")
self.y_t = shared_zeros((minibatch_size_g, dim), "y_t")
# self.decode_length = theano.shared(decode_length)
self.params.append(self.Wh)
self.params.append(self.bh)
def decode_step(
self, y_t, h_tm1, c_tm1,
Ui, Wi, bi, Uf, Wf, bf,
Uo, Wo, bo, Ug, Wg, bg,
Wh, bh
):
h_t, c_t = self.step(
y_t, h_tm1, c_tm1,
Ui, Wi, bi, Uf, Wf, bf,
Uo, Wo, bo, Ug, Wg, bg
)
y_t = T.dot(h_t, Wh) + bh
return y_t, h_t, c_t
def decode(self, h_tm1, c_tm1, timesteps):
outputs, updates = theano.scan(
fn=self.decode_step,
outputs_info=[self.y_t, h_tm1, c_tm1],
non_sequences=[
self.Ui, self.Wi, self.bi,
self.Uf, self.Wf, self.bf,
self.Uo, self.Wo, self.bo,
self.Ug, self.Wg, self.bg,
self.Wh, self.bh
],
n_steps=timesteps
)
updates = [
(self.h_tm1, outputs[1][-1]),
(self.c_tm1, outputs[2][-1])
]
# return T.flatten(outputs[0], 3), updates
return outputs[0], updates
@staticmethod
def argmax(seq):
seq = T.argmax(seq, axis=2)
return seq
#########################################################################################################################
class Seq2Seq(object):
def __init__(self, size, dim):
self.encoder = LSTMEncoder(size, dim)
self.decoder = LSTMDecoder(size, dim)
self.params = []
self.params += self.encoder.params
self.params += self.decoder.params
self._predict = None
self._train = None
self._test = None
def compile(self, loss_func, optimizer):
seq_input = T.tensor3()
seq_target = T.tensor3()
decode_timesteps = T.iscalar()
h_tm1, c_tm1, updates_encode = self.encoder.encode(seq_input)
seq_predict_flex, updates_decode_flex = self.decoder.decode(h_tm1, c_tm1, decode_timesteps)
seq_argmax = self.decoder.argmax(seq_predict_flex)
seq_predict, updates_decode = self.decoder.decode(h_tm1, c_tm1, T.shape(seq_target)[0])
loss = loss_func(seq_predict, seq_target)
self._predict = theano.function([seq_input, decode_timesteps], seq_argmax, updates=updates_encode+updates_decode_flex)
self._test = theano.function([seq_input, seq_target], loss, updates=updates_encode+updates_decode)
updates = []
updates += updates_encode
updates += updates_decode
updates += optimizer(self.params, loss)
self._train = theano.function([seq_input, seq_target], loss, updates = updates)
def predict(self, seq_input, decode_timesteps):
self.encoder.reset_state()
self.decoder.reset_state()
return self._predict(seq_input, decode_timesteps)
def train(self, seq_input, seq_target):
self.encoder.reset_state()
self.decoder.reset_state()
return self._train(seq_input, seq_target)
def test(self, seq_input, seq_target):
self.encoder.reset_state()
self.decoder.reset_state()
return self._test(seq_input, seq_target)
#########################################################################################################################
def train(x, target):
for mini_batch, target in zip(x,target):
mini_batch = mini_batch.astype(dtype)
target = target.astype(dtype)
print("result of train function(loss or update) :", seq2seq.train(mini_batch, target))
#########################################################################################################################
# make weight information to pickle file
# information of Encooder class and decoder class of Seq2Seq class
# Encooder and decoder class should have function that returns value of weight variables
# one list contains elements that save weights' information
def save_weight():
None
#########################################################################################################################
def gen_processed_seq(input_sentence):
tokenized_seq = nltk.word_tokenize( input_sentence )
input_sentences = [ None for _ in range(1) ]
input_sentences[0] = tokenized_seq
seq_input = readFile.word_to_idx(input_sentences, word_to_index_input_g )
sorted_seq_input = [ None for _ in range(minibatch_size_g) ]
sorted_seq_input[0] = seq_input[0]
input_len = len(seq_input[0])
for i in range(minibatch_size_g-1):
for j in range(input_len):
sorted_seq_input[i+1] = [-1]
input_finally = []
input_finally.append(sorted_seq_input)
return input_finally
#########################################################################################################################
def gen_one_hot(input_len, input_seq):
one_hot = readFile.seq_to_1hot(n_time_step_input_g, input_seq, "predict", 1, 1)
one_hot[0] = one_hot[0].astype(dtype)
print("one_hot : ", one_hot)
return one_hot
def get_idx(argmax, num_of_word):
idx_list = argmax[ : num_of_word, 0]
return idx_list
#########################################################################################################################
def predict():
input_sentence = raw_input("Input the English Sentence You Want to Translate into Spanish : ")
input_seq = gen_processed_seq(input_sentence)
print("input_seq[0][0] : ",input_seq[0][0])
num_of_word = len(input_seq[0][0])
one_hot = gen_one_hot(n_time_step_input_g, input_seq)
argmax = seq2seq.predict(one_hot[0] , n_time_step_input_g )
print("argmax_fin shape : ", argmax.shape)
print("argmax_fin : ", argmax)
idx_list_np = get_idx(argmax, num_of_word)
idx_list_py = idx_list_np.tolist()
print("index_to_word_target_g : ",index_to_word_target_g)
print("index_to_word_target_g[6] :", index_to_word_target_g[6])
result = readFile.idx_to_word(idx_list_py, index_to_word_target_g)
translated = ""
for elem in result :
translated += elem
translated += " "
print("translated : " , translated)
print("Translation End")
#########################################################################################################################
def gen_global_var(word_to_index_input, word_to_index_targrt, voca_dim, si, st, index_to_word_target):
global word_to_index_input_g
global word_to_index_targrt_g
global voca_dim_g
global minibatch_size_g
global n_time_step_input_g
global n_timestep_target_g
global index_to_word_target_g
word_to_index_input_g = word_to_index_input
word_to_index_targrt_g = word_to_index_targrt
voca_dim_g = voca_dim + 2
minibatch_size_g = si[0].shape[1]
n_time_step_input_g = si[0].shape[0]
n_timestep_target_g = st[0].shape[0]
index_to_word_target_g = index_to_word_target
return
#########################################################################################################################
def menu(si, st):
None
#########################################################################################################################
def gen_object():
return None
#########################################################################################################################
if __name__ == "__main__":
si, st, maxlen_input, minibatch_size, voca_dim, word_to_index_input, word_to_index_targrt, index_to_word_target = readFile.preprocessing()
gen_global_var(word_to_index_input, word_to_index_targrt, voca_dim, si, st, index_to_word_target)
seq2seq = Seq2Seq(n_time_step_input_g, voca_dim_g )
seq2seq.compile(loss_func=categorical_crossentropy, optimizer=adadelta)
while(True):
print("select a menu")
print("1. Training")
print("2. Translate specific English sentence into Spanish.")
val = input("selection : ")
if val == 1:
train(si, st)
elif val == 2:
predict()
和readfile.py是
import numpy as np
import itertools
import nltk
import sys
import os
from nltk.tokenize import sent_tokenize
import codecs
unknown_token = 'UNKNOWN_TOKEN'
start_token = '_S'
end_token = '__E'
num_of_seq = 0
input_path = "./europarl-v7.es-en.en"
target_path = "./europarl-v7.es-en.es"
minibatch_unit = 100
voca_dim = 3000
SEQ_NUM_LIMIT = 1000
##########################################################################################
def file_tokenize(file):
f = codecs.open( file, "r", "utf-8" )
tokenized_seq = []
sentences = []
total_sentence_num = 0
# sequence tokenize
for i,line in enumerate(f):
print("tokenized Sentence No." , i)
# strip() method to remove the newline character at the end of the input line.
tokenized_seq = nltk.word_tokenize( line.strip() )
tokenized_seq.insert(0, start_token)
tokenized_seq.append(end_token)
sentences.append(tokenized_seq)
total_sentence_num += 1;
if(total_sentence_num == SEQ_NUM_LIMIT):
break
return sentences,total_sentence_num
##########################################################################################
# Count the word frequencies
def cntWordFreq(sentences):
word_freq = nltk.FreqDist(itertools.chain(*sentences))
return word_freq
##########################################################################################
# Get the most common words and build index_to_word and word_to_index vectors
def build_WordToIdx_IdxtoWord(word_freq):
vocab = word_freq.most_common(voca_dim-1)
index_to_word = [x[0] for x in vocab]
index_to_word.append(unknown_token)
word_to_index = dict([(w,i) for i,w in enumerate(index_to_word)])
return index_to_word, word_to_index
##########################################################################################
# change word to index
def word_to_idx(sequences, word_to_index ) :
for i, sent in enumerate(sequences):
sequences[i] = [w if w in word_to_index else unknown_token for w in sent]
sequences[i] = [word_to_index[w] if w in word_to_index else -1 for w in sequences[i]]
return sequences
##########################################################################################
def idx_to_word(seq, index_to_word):
for i, sent in enumerate(seq):
seq[i] = index_to_word[sent]
#seq[i] = [index_to_word[sent] if sent in index_to_word else '?' ]
return seq
##########################################################################################
def sortByLen(seqs_input, seqs_target) :
# check maximum sentence length
max_len_input = 0
max_len_target = 0
for sentence in seqs_input :
tmp = len(sentence)
if max_len_input < tmp:
max_len_input = tmp
for sentence in seqs_target :
tmp = len(sentence)
if max_len_target < tmp:
max_len_target = tmp
seqs_sorted_input = [ [] for _ in range(max_len_input+1) ]
seqs_sorted_target = [ [] for _ in range(max_len_input+1) ]
i = 0
for sentence_input, sentence_target in zip(seqs_input, seqs_target) :
sentence_len = len(sentence_input)
seqs_sorted_input[sentence_len].append(sentence_input)
seqs_sorted_target[sentence_len].append(sentence_target)
i+=1
return seqs_sorted_input, seqs_sorted_target, max_len_input, max_len_target
##########################################################################################
def find_maxlen(sentence_group):
max_seq_len = 0
for seq in sentence_group :
if len(seq) > max_seq_len :
max_seq_len = len(seq)
return max_seq_len
##########################################################################################
def sort_by_timestep(sentence_group):
same_len_seq = np.asarray(sentence_group)
same_len_seq = apply_to_m1(same_len_seq)
sorted_seq = same_len_seq.transpose()
return sorted_seq
##########################################################################################
def seq_to_1hot(max_len, sorted_sentences, type, minibatch_unit, num_of_seq):
one_hot = [None for _ in range( len(sorted_sentences) )]
for i, sentence_group in enumerate(sorted_sentences):
if sentence_group and len(sentence_group[0]) != 0 :
max_seq_len = find_maxlen(sentence_group)
row = max_seq_len * minibatch_unit
one_hot[i] = np.zeros( (row, voca_dim + 2) )
time_step_seq = sort_by_timestep(sentence_group)
j = 0
for word_idx in np.nditer( time_step_seq ) :
if word_idx != -1:
one_hot[i][j][word_idx] = 1
j+=1
one_hot[i] = np.reshape(one_hot[i], ( max_seq_len, -1, voca_dim+2) )
return one_hot
##########################################################################################
def apply_to_m1(lst, dtype=np.int64):
inner_max_len = max(map(len, lst))
result = np.zeros( [len(lst), inner_max_len], dtype )
result[:] = -1
for i, row in enumerate(lst):
for j, val in enumerate(row):
result[i][j] = val
return result
##########################################################################################
def seq_group_by_mini_batch_size(minibatch_unit, sorted_seq, num_of_seq):
idx = 0
cnt = 0
minibatch_seq = [ [] for _ in range( (num_of_seq/minibatch_unit)+1) ]
for seqs in sorted_seq :
if seqs :
for seq in seqs :
if seq:
minibatch_seq[idx].append(seq)
cnt+=1
if minibatch_unit == cnt:
cnt = 0
idx+= 1
for i, seq in enumerate (minibatch_seq):
if seq == []:
minibatch_seq = minibatch_seq[: i- 1]
break
return minibatch_seq
##########################################################################################
def preprocessing():
global num_of_seq
global minibatch_unit
global input_path
global target_path
print("Start Preprocessing")
sentences_input, total_sentence_num = file_tokenize(input_path)
sentences_target, total_sentence_num_target = file_tokenize(target_path)
print("FINISHED : file_tokenize ")
word_freq_input = cntWordFreq(sentences_input)
word_freq_target = cntWordFreq(sentences_target)
print("FINISHED : cntWordFreq ")
index_to_word_input, word_to_index_input = build_WordToIdx_IdxtoWord(word_freq_input)
index_to_word_target, word_to_index_targrt = build_WordToIdx_IdxtoWord(word_freq_target)
print("FINISHED : build_WordToIdx_IdxtoWord ")
seqs_input = word_to_idx(sentences_input, word_to_index_input)
seqs_target = word_to_idx(sentences_target, word_to_index_targrt)
print("FINISHED : word_to_idx ")
seqs_sorted_input, seqs_sorted_target, maxlen_input, maxlen_target = sortByLen(seqs_input, seqs_target)
print("FINISHED : sortByLen ")
for seqs in seqs_input:
if seqs:
for seq in seqs:
if seq:
num_of_seq+=1
seq_by_mini_batch_size_input = seq_group_by_mini_batch_size(minibatch_unit, seqs_sorted_input, num_of_seq)
seq_by_mini_batch_size_target = seq_group_by_mini_batch_size(minibatch_unit, seqs_sorted_target, num_of_seq)
print("FINISHED : seq_group_by_mini_batch_size ")
_1hot_input = seq_to_1hot(maxlen_input, seq_by_mini_batch_size_input, "input",minibatch_unit, num_of_seq)
_1hot_target = seq_to_1hot(maxlen_target, seq_by_mini_batch_size_target, "target",minibatch_unit, num_of_seq)
print("FINISHED : seq_to_1hot ")
if minibatch_unit > total_sentence_num:
minibatch_unit = total_sentence_num
print("exit preprocessing")
return _1hot_input, _1hot_target, maxlen_input, minibatch_unit, voca_dim, word_to_index_input, word_to_index_targrt, index_to_word_target