In [1]:
import tensorflow as tf
import numpy as np
import gensim
import sys
import datetime,time
from six.moves import cPickle as pickle
In [2]:
def log_time_model(s):
now = datetime.datetime.now()
print "[%s][%s.%s][%s]" %(time.strftime('%Y-%m-%d %H-%M-%S', time.localtime(time.time())), now.second, now. microsecond, s)
In [4]:
def reformat(dataset, labels):
dataset = dataset.reshape(
(-1, WordVec_Length, Word_Length_Default * 2,num_channels)).astype(np.float32)
return dataset
In [5]:
def reAdd(dataset,Filter_Width):
# give pairs's distance,then distance = Filter_Length * File_Width
# and return real input data x = [None, WordVec_Length, Word_Length_Default * 2 + Filter_Length,num_channels]
b = np.zeros([Filter_Length,Filter_Width,num_channels])
return np.insert(dataset, Word_Length_Default, values=b, axis=1)
In [6]:
def accuracy(predictions, labels):
return (100.0 * np.sum(np.argmax(predictions, 1) == np.argmax(labels, 1)) / predictions.shape[0])
In [7]:
# define wordvec parameter
WordVec_Length = 50
WordVec_Depth = 1
# define batch size
BatchSize = 2
# define filter parameter
Filter_Length = WordVec_Length
Filter_Width = 3
num_channels = WordVec_Depth
### define filter depth, then equal next layer filter_curmat_depth
Filter_Depth = 9
# one question 's max word length , is equal with lenth of sentence. >50 的 sentence
Word_Length_Default = 50
# dim(y)
num_labels = 2
x = tf.placeholder(tf.float32,[BatchSize, Word_Length_Default * 2 + Filter_Width, WordVec_Length,num_channels],
name='x-input')
# dimension[0] is batch size,dim[1] and dim[2] means pairs's size,dim[3] is depth of pairs
y = tf.placeholder(tf.float32, shape=(BatchSize, num_labels),name='y-input')
valid_x = tf.constant(valid_dataset)
valid_label = tf.constant(valid_labels)
test_x = tf.constant(test_dataset)
log_time_model("x.shape()" + str(x.get_shape()))
In [8]:
def inference(input_tensor):
with tf.variable_scope('layer1-conv1'):
conv1_filter_weights = tf.get_variable('weight',
[Filter_Width,Filter_Length,num_channels,Filter_Depth],
initializer=tf.truncated_normal_initializer(stddev=0.1))
log_time_model("conv1_filter_weights = "+str(conv1_filter_weights.get_shape()))
biases1 = tf.get_variable('biases',[Filter_Depth],initializer=tf.constant_initializer(0.1))
log_time_model("biases1 = "+str(biases1.get_shape()))
conv1 = tf.nn.conv2d(x, conv1_filter_weights, strides = [1,1,1,1],padding='VALID')
log_time_model("conv1 = "+str(conv1.get_shape()))
bias1 = tf.nn.bias_add(conv1, biases1)
log_time_model("bias1 = "+str(bias1.get_shape()))
actived_conv1 = tf.nn.relu(bias1)
log_time_model("actived_conv1 = "+str(actived_conv1.get_shape()))
s = (Word_Length_Default * 2 + Filter_Width) - Filter_Width +1
# after conv , dim(2)'s size of conv
log_time_model(s)
# pool1 对一个句子做pool
pool1 = tf.nn.max_pool(actived_conv1[:,0:Word_Length_Default - Filter_Width +1,:,:],
ksize=[1,1,Word_Length_Default - Filter_Width +1,1],strides=[1,1,1,1],padding='VALID')
log_time_model("begin: 0:"+str(Word_Length_Default - Filter_Width +1))
## pool2 对第2个句子做pool
pool2 = tf.nn.max_pool(actived_conv1[:,Word_Length_Default +Filter_Width:s,:,:],
ksize=[1,s-Word_Length_Default - Filter_Width,1 ,1],strides=[1,1,1,1],padding='VALID')
log_time_model("begin:"+str(Word_Length_Default +Filter_Width+1)+":"+str(s))
log_time_model("pool1.get_shape() = "+ str(pool1.get_shape()))
log_time_model("pool2.get_shape() = "+ str(pool2.get_shape()))
simM = tf.get_variable('M',[Filter_Depth,Filter_Depth],initializer=tf.truncated_normal_initializer(stddev=0.1))
# 定义相似度矩阵 simM
log_time_model("simM = "+str(simM.get_shape()))
pool1 = tf.reshape(pool1,[BatchSize,Filter_Depth])
log_time_model("pool1.get_shape() = "+ str(pool1.get_shape()))
simmat = tf.matmul(pool1,simM)
log_time_model("simmat1 shape = "+str(simmat.get_shape()))
simmat = tf.reshape(simmat,[BatchSize,1,1,Filter_Depth])
log_time_model("simmat reshape = "+str(simmat.get_shape()))
log_time_model("pool2.get_shape() = "+ str(pool2.get_shape()))
simmat = tf.matmul(simmat,pool2,transpose_b=True)
pool2 = tf.reshape(pool2,[BatchSize,Filter_Depth])
simmat = tf.reshape(simmat,[BatchSize,1])
log_time_model("pool1.get_shape() = "+ str(pool1.get_shape()))
log_time_model(" pool2 shape = "+str(pool2.get_shape()))
log_time_model(" simmat shape = "+str(simmat.get_shape()))
out = tf.concat([pool1,simmat,pool2],1)
log_time_model(" out shape = "+str(out.get_shape()))
# NEXT STEP: MAXPOOL 定义 两个不需要训练的变量，maxpool ()
# 定义一个需要训练的矩阵M
# 两个变量矩阵相乘
# 拉直拼接
# 全连接
# softmax
fulljoin_num = 200
with tf.variable_scope('layer2-fulllink'):
weights = tf.get_variable('weight',[out.get_shape().as_list()[1],fulljoin_num],
initializer=tf.truncated_normal_initializer(stddev=0.1))
biases = tf.get_variable('biases',[fulljoin_num],initializer=tf.constant_initializer(0.1))
hidden = tf.nn.relu(tf.matmul(out, weights) + biases)
weights = tf.Variable(tf.truncated_normal([fulljoin_num, num_labels], stddev=0.1))
biases = tf.Variable(tf.constant(0.1, shape=[num_labels]))
out =tf.matmul(hidden, weights) + biases
log_time_model(" out shape = "+str(out.get_shape()))
return out
out = inference(x)
# Training computation.
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=out, labels=y))
# Optimizer.
optimizer = tf.train.GradientDescentOptimizer(0.05).minimize(loss)
# Predictions for the training, validation, and test data.
train_prediction = tf.nn.softmax(out)
# valid_prediction = tf.nn.softmax(inference(valid_x))
# test_prediction = tf.nn.softmax(inference(test_x))
In [ ]:
In [9]:
# with tf.Session() as sess:
# print sess.run(a.initializer)
# print dir(a.get_shape()[1].value)
# print a.get_shape()[1].value
In [10]:
# read dataset
def openFile():
pickle_file = 'dataset.pickle'
with open(pickle_file, 'rb') as f:
data = pickle.load(f)
##第一个句子的样本矩阵为np.ndarray((samplesize, 300, 50), dtype=np.float32))
train_dataset1 = data['train_dataset1']
##第二个句子的样本矩阵为np.ndarray((samplesize, 300, 50), dtype=np.float32))
train_dataset2 = data['train_dataset2']
##类别标识np.ndarray((samplesize, dtype = np.int32))
train_labels = data['train_labels']
##验证集
valid_dataset1 = data['valid_dataset1']
valid_dataset2 = data['valid_dataset2']
valid_labels = data['valid_labels']
##测试集
test_dataset1 = data['test_dataset1']
test_dataset2 = data['test_dataset2']