def train(learning_rate, batch_size, epoch, out_len, process_data):
word_placeholder = tf.placeholder(tf.int32, [None, process_data.max_len])
postag_placeholder = tf.placeholder(tf.int32, [None, process_data.max_len])
out_placeholder = tf.placeholder(tf.float32, [None, out_len])
cnn_model = cnn_p.cnn_model(word_placeholder, postag_placeholder, \
np.shape(process_data.word_embedding), np.shape(process_data.postag_embedding), \
process_data.word_embedding, process_data.postag_embedding, out_len)
error = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels = out_placeholder, logits = cnn_model))
train_step = tf.train.AdamOptimizer(learning_rate).minimize(error)
train_data_iter = process_data.generate_batch(batch_size, process_data.train_data)
saver = tf.train.Saver(max_to_keep = 10)
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
if RESTORE:
saver.restore(sess, tf.train.latest_checkpoint('./cnn_model'))
step = 0
for ep in range(epoch):
print('epoch:{}'.format(ep))
try:
data, label = train_data_iter.__next__()
while data:
step += 1
sess.run(train_step, feed_dict = {word_placeholder:data['word_embedding'], postag_placeholder:data['postag'], out_placeholder:label})
_error = sess.run(error, feed_dict = {word_placeholder:data['word_embedding'], postag_placeholder:data['postag'], out_placeholder:label})
data, label = train_data_iter.__next__()
if step % 100 == 0:
print(step)
saver.save(sess, './cnn_model/cnn.ckpt'+str(_error), global_step = step, write_meta_graph=False)
print('step:{}, error:{}'.format(step, _error))
except Exception as e:
print(e)
train_data_iter = process_data.generate_batch(batch_size, process_data.train_data)
def train(learning_rate, batch_size, epoch, process_data):
word_placeholder = tf.placeholder(tf.int32, [None, process_data.max_len])
postag_placeholder = tf.placeholder(tf.int32, [None, process_data.max_len])
p_placeholder = tf.placeholder(tf.int32, [None])
sequence_lengths = tf.placeholder(tf.int32, [None])
batch_length = tf.placeholder(tf.int32)
out_placeholder = tf.placeholder(tf.int32, [None, process_data.max_len])
ps_model = ps_lstmcrf.ps_lstmcrf_model(word_placeholder, postag_placeholder, p_placeholder, np.shape(process_data.word_embedding), np.shape(process_data.postag_embedding), \
np.shape(process_data.p_embedding), process_data.word_embedding, process_data.postag_embedding, \
process_data.p_embedding, sequence_lengths, batch_length)
log_likelihood, transition_params = tf.contrib.crf.crf_log_likelihood(
ps_model, out_placeholder, sequence_lengths)
error = tf.reduce_mean(-log_likelihood)
train_step = tf.train.AdamOptimizer(learning_rate).minimize(error)
train_data_iter = process_data.generate_batch(batch_size, process_data.train_data, \
features = ['word_embedding', 'postag', 'p', 'sequence_lengths'], label_type = 's')
saver = tf.train.Saver(max_to_keep=20)
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
if RESTORE:
saver.restore(sess,
tf.train.latest_checkpoint('./lstmcrf_ps_model'))
step = 0
for ep in range(epoch):
print('epoch:{}'.format(ep))
try:
data, label = train_data_iter.__next__()
while data:
step += 1
sess.run(train_step, feed_dict = {word_placeholder:data['word_embedding'], postag_placeholder:data['postag'], \
p_placeholder:data['p'], sequence_lengths:data['sequence_lengths'], batch_length:len(data['sequence_lengths']), out_placeholder:label})
_error = sess.run(error, feed_dict = {word_placeholder:data['word_embedding'], postag_placeholder:data['postag'], \
p_placeholder:data['p'], sequence_lengths:data['sequence_lengths'], batch_length:len(data['sequence_lengths']), out_placeholder:label})
data, label = train_data_iter.__next__()
if step % 100 == 0:
print(step)
if step % 2200 == 0:
print(step)
saver.save(sess,
'./lstmcrf_ps_model/lstmcrf_ps.ckpt' +
str(_error),
global_step=step,
write_meta_graph=False)
print('step:{}, error:{}'.format(step, _error))
except Exception as e:
print(e)
train_data_iter = process_data.generate_batch(
batch_size,
process_data.train_data,
features=[
'word_embedding', 'postag', 'p', 'sequence_lengths'
],
label_type='s')
def evaluate(process_data):
pred_correct_num = 0
pred_num = 0
real_num = 0
def stats(predict_output, label):
nonlocal pred_correct_num, pred_num, real_num
real_num += sum(list(map(lambda x : sum(x), label)))
for i, row in enumerate(predict_output):
for j, _ in enumerate(row):
if predict_output[i][j] > 0.5:
pred_num += 1
if label[i][j] == 1:
pred_correct_num += 1
word_placeholder = tf.placeholder(tf.int32, [None, process_data.max_len])
postag_placeholder = tf.placeholder(tf.int32, [None, process_data.max_len])
cnn_model = cnn_p.cnn_model(word_placeholder, postag_placeholder, \
np.shape(process_data.word_embedding), np.shape(process_data.postag_embedding), \
process_data.word_embedding, process_data.postag_embedding, out_len)
test_data_iter = process_data.generate_batch(batch_size, process_data.test_data)
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, './cnn_model/cnn.ckpt2.99603e-06-43100')#tf.train.latest_checkpoint('./cnn_model'))
try:
data, label = test_data_iter.__next__()
while data:
predict_output = sess.run(cnn_model, feed_dict = {word_placeholder:data['word_embedding'], \
postag_placeholder:data['postag']})
predict_output = sess.run(tf.nn.sigmoid(predict_output))
stats(predict_output, label)
data, label = test_data_iter.__next__()
except:
print('预测完毕')
precision = pred_correct_num / pred_num
recall = pred_correct_num / real_num
f1 = (2 * precision * recall)/(precision + recall)
print('pred_correct_num:{}, pred_num:{}, real_num:{}'.format(pred_correct_num, pred_num, real_num))
print('precision:{},recall:{},f1:{}'.format(precision, recall, f1))
return f1, precision, recall
def evaluate(process_data):
pred_correct_num = 0
pred_num = 0
real_num = 0
def stats(predict_output, label):
def convert_to_num_str(tags):
num_str = []
for i in range(len(tags)):
if tags[i] == 3: #表示B
start_flag = str(i)
j = i + 1
while j < len(tags) and tags[j] != 0 and tags[j] != 3:
j += 1
start_flag += str(j)
num_str.append(start_flag)
return num_str
nonlocal pred_correct_num, pred_num, real_num
for i in range(len(predict_output)):
pred_num_str = convert_to_num_str(predict_output[i])
lab_num_str = convert_to_num_str(label[i])
pred_num += len(pred_num_str)
real_num += len(lab_num_str)
for ele in pred_num_str:
if ele in lab_num_str:
pred_correct_num += 1
word_placeholder = tf.placeholder(tf.int32, [None, process_data.max_len])
postag_placeholder = tf.placeholder(tf.int32, [None, process_data.max_len])
p_placeholder = tf.placeholder(tf.int32, [None])
sequence_lengths = tf.placeholder(tf.int32, [None])
out_placeholder = tf.placeholder(tf.int32, [None, process_data.max_len])
batch_length = tf.placeholder(tf.int32)
ps_model = ps_lstmcrf.ps_lstmcrf_model(word_placeholder, postag_placeholder, p_placeholder, np.shape(process_data.word_embedding), np.shape(process_data.postag_embedding), \
np.shape(process_data.p_embedding), process_data.word_embedding, process_data.postag_embedding, \
process_data.p_embedding, sequence_lengths, batch_length)
log_likelihood, transition_params = tf.contrib.crf.crf_log_likelihood(
ps_model, out_placeholder, sequence_lengths)
decode_tags, best_score = tf.contrib.crf.crf_decode(
ps_model, transition_params, sequence_lengths)
test_data_iter = process_data.generate_batch(batch_size, process_data.test_data, \
features = ['word_embedding', 'postag', 'p', 'sequence_lengths'], label_type = 's')
saver = tf.train.Saver()
with tf.Session() as sess:
# saver.restore(sess, tf.train.latest_checkpoint('./lstmcrf_ps_model'))
saver.restore(
sess, './lstmcrf_ps_model/lstmcrf_ps.ckpt0.104050994-8700'
) #'./lstmcrf_ps_model/lstmcrf_ps.ckpt0.037134737-8200')#'./lstmcrf_ps_model/lstmcrf_ps.ckpt0.01772517-8000')#'./lstmcrf_ps_model/lstmcrf_ps.ckpt0.0014773-72100')
decode_tags, best_score = tf.contrib.crf.crf_decode(
ps_model, transition_params, sequence_lengths)
try:
data, label = test_data_iter.__next__()
while data:
_decode_tags = sess.run(decode_tags, feed_dict = {word_placeholder:data['word_embedding'], postag_placeholder:data['postag'], \
p_placeholder:data['p'], sequence_lengths:data['sequence_lengths'], batch_length:len(data['sequence_lengths'])})
stats(_decode_tags, label)
data, label = test_data_iter.__next__()
except Exception as e:
print('预测完毕')
precision = pred_correct_num / pred_num
recall = pred_correct_num / real_num
f1 = (2 * precision * recall) / (precision + recall)
print('pred_correct_num:{}, pred_num:{}, real_num:{}'.format(
pred_correct_num, pred_num, real_num))
print('precision:{},recall:{},f1:{}'.format(precision, recall, f1))
return f1, precision, recall