def w2v_type_fully():
x = tf.placeholder(tf.float32, [None, 600], name = 'x')
dense_1 = tf.layers.dense(inputs=x, units=1000, activation=tf.nn.relu, name = 'dense_1')
dense_2 = tf.layers.dense(inputs=dense_1, units=150, activation=tf.nn.tanh, name = 'dense_2')
W = tf.Variable(tf.zeros([150, 1]), name = 'W')
b = tf.Variable(tf.zeros([1]), name = 'b')
y = tf.sigmoid(tf.matmul(dense_2, W) + b, name = 'y')
y_ = tf.placeholder(tf.float32, [None, 1], name = 'y_')
loss = tf.reduce_mean((y - y_) * (y - y_), name = 'loss')
train_step = tf.train.AdamOptimizer(0.003).minimize(loss, name = 'train_step')
sess = tf.InteractiveSession()
tf.global_variables_initializer().run()
figer = data_utility.figer_data()
test_xs, test_ys = figer.get_test_data(0, 10000)
saver = tf.train.Saver()
for i in range(0,10000):
batch_xs, batch_ys = figer.next_batch()
_, print_loss, print_y = sess.run([train_step, loss, y], feed_dict={x: batch_xs, y_: batch_ys})
zero_R_count = 0.0
NN_count = 0.0
for j in range(0, len(print_y)):
if batch_ys[j][0] == 0.0:
zero_R_count += 1.0
if abs(print_y[j] - batch_ys[j][0]) < 0.5:
NN_count += 1.0
zero_R_acc = zero_R_count / float(len(print_y))
NN_acc = NN_count / float(len(print_y))
with open('loss_record.txt', 'a') as out_file:
out_file.write('epoch = ' + str(i) + '\t lost = ' + str(print_loss) + \
'\t zero_R_acc = ' + str(zero_R_acc) + '\t NN_acc = ' + str(NN_acc) + '\n')
print ('epoch = ' + str(i) + '\t lost = ' + str(print_loss) + \
'\t zero_R_acc = ' + str(zero_R_acc) + '\t NN_acc = ' + str(NN_acc) + '\n')
print_loss, print_y = sess.run([loss, y], feed_dict={x: test_xs, y_: test_ys})
zero_R_count = 0.0
NN_count = 0.0
for j in range(0, len(print_y)):
if test_ys[j][0] == 0.0:
zero_R_count += 1.0
if abs(print_y[j] - test_ys[j][0]) < 0.5:
NN_count += 1.0
zero_R_acc = zero_R_count / float(len(print_y))
NN_acc = NN_count / float(len(print_y))
if i%10 == 0:
with open('loss_record.txt', 'a') as out_file:
out_file.write('test: epoch = ' + str(i) + '\t lost = ' + str(print_loss) + \
'\t zero_R_acc = ' + str(zero_R_acc) + '\t NN_acc = ' + str(NN_acc) + '\n')
if i % 100 == 0:
model_name = 'model/model_' + str(i) + '.ckpt'
saver.save(sess, model_name)
def w2v_type_L1_norm():
x = tf.placeholder(tf.float32, [None, 600], name='x')
x1 = tf.slice(x, [0, 0], [-1, 300])
x2 = tf.slice(x, [0, 300], [-1, -1])
x1_norm = tf.reshape(tf.norm(x1, ord=1, axis=1), [-1, 1])
x2_norm = tf.reshape(tf.norm(x2, ord=1, axis=1), [-1, 1])
x_with_norm = tf.concat([x, x1_norm, x2_norm], 1)
dense_1 = tf.layers.dense(inputs=x_with_norm,
units=150,
activation=tf.nn.tanh,
name='dense_1')
example_w = tf.placeholder(tf.float32, [None, 1], name='example_w')
W = tf.Variable(tf.zeros([150, 1]), name='W')
b = tf.Variable(tf.zeros([1]), name='b')
y_logit = tf.add(tf.matmul(dense_1, W), b, name='y_logit')
y = tf.sigmoid(y_logit, name='y')
y_ = tf.placeholder(tf.float32, [None, 1], name='y_')
# loss = tf.reduce_mean((y - y_) * (y - y_), name = 'loss')
loss = tf.reduce_mean(
tf.losses.sigmoid_cross_entropy(y_, y_logit, example_w))
train_step = tf.train.GradientDescentOptimizer(0.33).minimize(
loss, name='train_step')
# sess = tf.InteractiveSession()
# tf.global_variables_initializer().run()
figer = data_utility.figer_data()
saver = tf.train.Saver()
config = tf.ConfigProto(device_count={'GPU': 1})
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
for i in range(0, 40 * figer.total_batch_num):
batch_xs, batch_ys, batch_ws = figer.next_shuffle_train_batch()
_, print_loss = sess.run([train_step, loss],
feed_dict={
x: batch_xs,
y_: batch_ys,
example_w: batch_ws
})
with open('loss_record.txt', 'a') as out_file:
out_file.write('train: epoch= {} loss= {}\n' \
.format(i, print_loss))
# print ('train: epoch= {} loss= {}\n' \
# .format(i, print_loss))
if i % figer.total_batch_num == 10:
get_validation_data_info(sess, figer, i, loss, x, y, y_,
example_w)
get_train_data_info(sess, figer, i, loss, x, y, y_, example_w)
get_test_data_info(sess, figer, i, loss, x, y, y_, example_w)
model_name = 'model/model_' + str(i) + '.ckpt'
saver.save(sess, model_name)
def context_type_RNN():
context = tf.placeholder(tf.float32, [None, None, 300], name = 'context')
cell = tf.contrib.rnn.BasicLSTMCell(300)
batch_size = tf.shape(context)[1]
initial_state = cell.zero_state(batch_size, tf.float32)
sequence_length = tf.placeholder(tf.int32, [None], name = 'sequence_length')
outputs, _ = tf.nn.dynamic_rnn(cell, context, sequence_length = sequence_length, initial_state=initial_state, time_major=True, dtype=tf.float32)
final_outputs = select_rnn(outputs, sequence_length - 1)
type_emb = tf.placeholder(tf.float32, [None, 300], name = 'type_emb')
x = tf.concat([final_outputs, type_emb], 1, name = 'x')
W = tf.Variable(tf.zeros([600, 1]), name = 'W')
b = tf.Variable(tf.zeros([1]), name = 'b')
y_logit = tf.add(tf.matmul(x, W), b, name = 'y_logit')
y = tf.sigmoid(y_logit, name = 'y')
y_ = tf.placeholder(tf.float32, [None, 1], name = 'y_')
example_w = tf.placeholder(tf.float32, [None, 1], name = 'example_w')
loss = tf.reduce_mean(tf.losses.sigmoid_cross_entropy(y_, y_logit, example_w), name = 'loss')
# train_step = tf.train.AdamOptimizer(0.003).minimize(loss, name = 'train_step')
train_step = tf.train.GradientDescentOptimizer(0.33).minimize(loss, name = 'train_step')
figer = data_utility.figer_data()
saver = tf.train.Saver()
config = tf.ConfigProto(
device_count = {'GPU': 1}
)
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
for i in range(0, 40*figer.total_batch_num):
batch_cs, batch_cs_lengths, batch_ls, batch_ys, batch_ws = figer.next_shuffle_train_context_batch()
_, print_loss = sess.run([train_step, loss], feed_dict={context: batch_cs, sequence_length: batch_cs_lengths, type_emb: batch_ls, y_: batch_ys, example_w: batch_ws})
with open('loss_record.txt', 'a') as out_file:
out_file.write('train: epoch= {} loss= {}\n' \
.format(i, print_loss))
if i % figer.total_batch_num == 10:
get_train_data_info(sess, figer, i, loss, y, context, sequence_length, type_emb, y_, example_w)
get_validation_data_info(sess, figer, i, loss, y, context, sequence_length, type_emb, y_, example_w)
get_test_data_info(sess, figer, i, loss, y, context, sequence_length, type_emb, y_, example_w)
model_name = 'model/model_' + str(i) + '.ckpt'
saver.save(sess, model_name)
def w2v_type_linear():
x = tf.placeholder(tf.float32, [None, 600], name = 'x')
example_w = tf.placeholder(tf.float32, [None, 1], name = 'example_w')
W = tf.Variable(tf.zeros([600, 1]), name = 'W')
b = tf.Variable(tf.zeros([1]), name = 'b')
y_logit = tf.add(tf.matmul(x, W), b, name = 'y_logit')
y = tf.sigmoid(y_logit, name = 'y')
y_ = tf.placeholder(tf.float32, [None, 1], name = 'y_')
# loss = tf.reduce_mean((y - y_) * (y - y_), name = 'loss')
loss = tf.reduce_mean(tf.losses.sigmoid_cross_entropy(y_, y_logit, example_w))
train_step = tf.train.GradientDescentOptimizer(0.33).minimize(loss, name = 'train_step')
# sess = tf.InteractiveSession()
# tf.global_variables_initializer().run()
figer = data_utility.figer_data()
saver = tf.train.Saver()
config = tf.ConfigProto(
device_count = {'GPU': 0}
)
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
for i in range(0, 20*figer.total_batch_num):
batch_xs, batch_ys, batch_ws = figer.next_shuffle_train_batch()
_, print_loss = sess.run([train_step, loss], feed_dict={x: batch_xs, y_: batch_ys, example_w: batch_ws})
with open('loss_record.txt', 'a') as out_file:
out_file.write('train: epoch= {} loss= {}\n' \
.format(i, print_loss))
# print ('train: epoch= {} loss= {}\n' \
# .format(i, print_loss))
if i % 100 == 20:
get_train_data_info(sess, figer, i, loss, x, y, y_, example_w)
get_test_data_info(sess, figer, i)
if i % 1000 == 20:
model_name = 'model/model_' + str(i) + '.ckpt'
saver.save(sess, model_name)
def context_type_CNN():
context = tf.placeholder(tf.float32, [None, 50, 300], name = 'context')
pos_emb = tf.placeholder(tf.float32, [None, 50, 50], name = 'lengths_emb')
x = tf.concat([context, pos_emb], 2, name = 'x')
conv1 = tf.layers.conv1d(inputs=x, filters=300, kernel_size=[3], padding="same", activation=tf.nn.relu)
pool1 = tf.layers.max_pooling1d(inputs=conv1, pool_size=[50], strides=1, name = 'pool1')
pool_last = tf.reshape(pool1, [-1,300])
type_emb = tf.placeholder(tf.float32, [None, 300], name = 'type_emb')
concat = tf.concat([pool_last, type_emb], 1, name = 'concat')
y_ = tf.placeholder(tf.float32, [None, 1], name = 'y_')
W = tf.Variable(tf.zeros([600, 1]), name = 'W')
b = tf.Variable(tf.zeros([1]), name = 'b')
y_logit = tf.add(tf.matmul(concat, W), b, name = 'y_logit')
y = tf.sigmoid(y_logit, name = 'y')
example_w = tf.placeholder(tf.float32, [None, 1], name = 'example_w')
loss = tf.reduce_mean(tf.losses.sigmoid_cross_entropy(y_, y_logit, example_w), name = 'loss')
train_step = tf.train.GradientDescentOptimizer(0.33).minimize(loss, name = 'train_step')
figer = data_utility.figer_data()
saver = tf.train.Saver()
config = tf.ConfigProto(
device_count = {'GPU': 1}
)
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
for i in range(0, 40*figer.total_batch_num):
batch_cs, batch_pos, batch_ls, batch_ys, batch_ws = figer.next_shuffle_train_context_batch_CNN()
_, print_loss = sess.run([train_step, loss], feed_dict={context: batch_cs, pos_emb: batch_pos, type_emb: batch_ls, y_: batch_ys, example_w: batch_ws})
with open('loss_record.txt', 'a') as out_file:
out_file.write('train: epoch= {} loss= {}\n' \
.format(i, print_loss))
if i % figer.total_batch_num == 10:
get_train_data_info_CNN(sess, figer, i, loss, y, context, pos_emb, type_emb, y_, example_w)
get_validation_data_info_CNN(sess, figer, i, loss, y, context, pos_emb, type_emb, y_, example_w)
get_test_data_info_CNN(sess, figer, i, loss, y, context, pos_emb, type_emb, y_, example_w)
model_name = 'model/model_' + str(i) + '.ckpt'
saver.save(sess, model_name)
def context_type():
context = tf.placeholder(tf.float32, [15, None, 300], name = 'context')
cell = tf.contrib.rnn.BasicLSTMCell(300)
batch_size = tf.shape(context)[1]
initial_state = cell.zero_state(batch_size, tf.float32)
outputs, _ = tf.nn.dynamic_rnn(cell, context, initial_state=initial_state, time_major=True, dtype=tf.float32)
final_outputs = outputs[-1]
type_emb = tf.placeholder(tf.float32, [None, 300], name = 'type_emb')
x = tf.concat([final_outputs, type_emb], 1, name = 'x')
dense_1 = tf.layers.dense(inputs=x, units=1000, activation=tf.nn.relu, name = 'dense_1')
dense_2 = tf.layers.dense(inputs=dense_1, units=150, activation=tf.nn.tanh, name = 'dense_2')
W = tf.Variable(tf.zeros([150, 1]), name = 'W')
b = tf.Variable(tf.zeros([1]), name = 'b')
y = tf.sigmoid(tf.matmul(dense_2, W) + b, name = 'y')
y_ = tf.placeholder(tf.float32, [None, 1], name = 'y_')
loss = tf.reduce_mean((y - y_) * (y - y_), name = 'loss')
# train_step = tf.train.AdamOptimizer(0.003).minimize(loss, name = 'train_step')
train_step = tf.train.GradientDescentOptimizer(0.5).minimize(loss, name = 'train_step')
config = tf.ConfigProto(
device_count = {'GPU': 0}
)
with tf.Session(config=config) as sess:
tf.global_variables_initializer().run()
figer = data_utility.figer_data()
test_cs, test_ls, test_ys = figer.get_context_test_data(0, 10000)
saver = tf.train.Saver()
for i in range(0,10000):
batch_cs, batch_ls, batch_ys = figer.next_context_batch()
_, print_loss, print_y = sess.run([train_step, loss, y], feed_dict={context: batch_cs, type_emb: batch_ls, y_: batch_ys})
zero_R_count = 0.0
NN_count = 0.0
for j in range(0, len(print_y)):
if batch_ys[j][0] == 0.0:
zero_R_count += 1.0
if abs(print_y[j] - batch_ys[j][0]) < 0.5:
NN_count += 1.0
zero_R_acc = zero_R_count / float(len(print_y))
NN_acc = NN_count / float(len(print_y))
with open('loss_record.txt', 'a') as out_file:
out_file.write('train: epoch = ' + str(i) + '\t lost = ' + str(print_loss) + \
'\t zero_R_acc = ' + str(zero_R_acc) + '\t NN_acc = ' + str(NN_acc) + '\n')
# print ('epoch = ' + str(i) + '\t lost = ' + str(print_loss) + \
# '\t zero_R_acc = ' + str(zero_R_acc) + '\t NN_acc = ' + str(NN_acc) + '\n')
if i%10 == 0:
print_loss, print_y = sess.run([loss, y], feed_dict={context: test_cs, type_emb: test_ls, y_: test_ys})
zero_R_count = 0.0
NN_count = 0.0
for j in range(0, len(print_y)):
if test_ys[j][0] == 0.0:
zero_R_count += 1.0
if abs(print_y[j] - test_ys[j][0]) < 0.5:
NN_count += 1.0
zero_R_acc = zero_R_count / float(len(print_y))
NN_acc = NN_count / float(len(print_y))
with open('loss_record.txt', 'a') as out_file:
out_file.write('test: epoch = ' + str(i) + '\t lost = ' + str(print_loss) + \
'\t zero_R_acc = ' + str(zero_R_acc) + '\t NN_acc = ' + str(NN_acc) + '\n')
# print ('test: epoch = ' + str(i) + '\t lost = ' + str(print_loss) + \
# '\t zero_R_acc = ' + str(zero_R_acc) + '\t NN_acc = ' + str(NN_acc) + '\n')
if i % 100 == 0:
model_name = 'model/model_' + str(i) + '.ckpt'
saver.save(sess, model_name)