def DevLoss(dev_x, dev_y, word_index_dict_size, dropout_keep_prob,
embedding_dim, filter_sizes, num_filters, l2_reg_lambda, cpus):
x, y = data.LoadDevData(dev_path)
x_size = len(y)
dev_loss, dev_acc = BatchLoss(dev_x, dev_y, word_index_dict_size,
dropout_keep_prob, embedding_dim,
filter_sizes, num_filters, x_size,
l2_reg_lambda, cpus)
return dev_loss, dev_acc
def DevLoss(text_length, dev_x, dev_y, word_index_dict_size, dropout_keep_prob,
embedding_dim, filter_sizes, num_filters, l2_reg_lambda, cpus):
x, y = data.LoadDevData(dev_path)
x_size = len(y)
tf.get_variable_scope().reuse_variables()
dev_loss, dev_acc = BatchLoss(text_length, dev_x, dev_y,
word_index_dict_size, dropout_keep_prob,
embedding_dim, filter_sizes, num_filters,
x_size, l2_reg_lambda, cpus)
return dev_loss, dev_acc
def Train(train_path, dev_path, \
word_index_dict_size, text_length, embedding_dim, \
filter_sizes, num_filters, l2_reg_lambda, batch_size, \
log_device_placement, model_dir, dropout_keep_prob_value, \
num_epochs, evaluate_every, checkpoint_every, gpus, cpus):
g1 = tf.Graph()
valid_max_accuracy = 0
with g1.as_default(), tf.device("/gpu:" + str(gpus[0])):
config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=log_device_placement)
config.gpu_options.allow_growth = True
config.gpu_options.allocator_type = 'BFC'
sess = tf.Session(config=config)
global_step = tf.get_variable('global_step', [], \
initializer=tf.constant_initializer(0), trainable=False)
opt = tf.train.AdamOptimizer(1e-3)
#opt=tf.train.GradientDescentOptimizer(1e-3)
input_x = tf.placeholder(tf.int32, [None, text_length], name="input_x")
input_y = tf.placeholder(tf.float32, [None, 9], name="input_y")
# For online predicting.
dev_x = tf.placeholder(tf.int32, [None, text_length], name="dev_x")
dev_y = tf.placeholder(tf.float32, [None, 9], name="dev_y")
dropout_keep_prob = tf.placeholder(tf.float32,
name="dropout_keep_prob")
train_op, average_loss, average_accuracy = MultiGPUCalcLossAccu(
text_length, gpus, input_x, input_y, batch_size,
word_index_dict_size, dropout_keep_prob, embedding_dim,
filter_sizes, num_filters, l2_reg_lambda, opt, global_step, cpus)
dev_loss, dev_acc = DevLoss(text_length, dev_x, dev_y,
word_index_dict_size, dropout_keep_prob,
embedding_dim, filter_sizes, num_filters,
l2_reg_lambda, cpus)
# Session
saver = tf.train.Saver(tf.global_variables(), max_to_keep=2)
ckpt = tf.train.get_checkpoint_state(model_dir)
if reload_model and ckpt and ckpt.model_checkpoint_path:
#_saver=tf.train.Saver(tf.global_variables()
saver.restore(sess, ckpt.model_checkpoint_path)
print "++++++++++++++++++++++****************"
print sess.run(global_step)
else:
init = tf.global_variables_initializer()
sess.run(init)
#config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=log_device_placement)
#config.gpu_options.allow_growth = True
#config.gpu_options.allocator_type = 'BFC'
#sess = tf.Session(config=config)
'''
sess = tf.Session(config=tf.ConfigProto(
allow_soft_placement=True,
log_device_placement=log_device_placement))
'''
#sess.run(init)
tf.train.start_queue_runners(sess=sess)
# Output directory for models and summaries
out_dir = os.path.abspath(os.path.join(os.path.curdir, model_dir))
print("Writing to {}\n".format(out_dir))
# Summaries for loss and accuracy
loss_summary = tf.summary.scalar("loss", average_loss)
acc_summary = tf.summary.scalar("accuracy", average_accuracy)
dev_loss_summary = tf.summary.scalar("loss", dev_loss)
dev_acc_summary = tf.summary.scalar("accuracy", dev_acc)
# Train Summaries
print "Train Summaries"
train_summary_op = tf.summary.merge([loss_summary, acc_summary])
train_summary_dir = os.path.join(out_dir, "summaries", "train")
train_summary_writer = tf.summary.FileWriter(train_summary_dir,
sess.graph)
# Dev summaries
print "Dev summaries"
dev_summary_op = tf.summary.merge([dev_loss_summary, dev_acc_summary])
dev_summary_dir = os.path.join(out_dir, "summaries", "dev")
dev_summary_writer = tf.summary.FileWriter(dev_summary_dir, sess.graph)
for epoch in range(num_epochs):
train_iter = data.BatchIter(train_path, batch_size)
for train_batch in train_iter:
x_batch, y_batch = zip(*train_batch)
train_feed_dict = {
input_x: x_batch,
input_y: y_batch,
dropout_keep_prob: dropout_keep_prob_value
}
"""
TrainBatch(train_feed_dict, sess, train_op, global_step, \
train_summary_op, average_loss, average_accuracy, \
train_summary_writer)
"""
TrainStep(sess, train_feed_dict, train_op, global_step,
train_summary_op, average_loss, average_accuracy,
train_summary_writer)
current_step = tf.train.global_step(sess, global_step)
if current_step % evaluate_every == 0:
print("\nEvaluation:")
x, y = data.LoadDevData(dev_path)
dev_feed_dict = {
dev_x: x,
dev_y: y,
dropout_keep_prob: 1.0
}
#print("1111")
valid_accuracy = DevStep(sess, dev_feed_dict,
dev_summary_op, dev_loss, dev_acc,
current_step, dev_summary_writer)
print("")
current_step = tf.train.global_step(sess, global_step)
if current_step % checkpoint_every == 0:
if valid_accuracy > valid_max_accuracy:
valid_max_accuracy = valid_accuracy
checkpoint_path = os.path.join(model_dir, 'model.ckpt')
saver.save(sess,
checkpoint_path,
global_step=current_step)
# Remove the old checkpoint files.
#RemovePBXStepsBefore(model_dir, 'cnn', 4, current_step, checkpoint_every)
# Save the latest checkpoint.
"""
#save old meta data.
checkpoint_path = os.path.join(model_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=current_step)
"""
#SavePB(sess, g1, os.path.join(model_dir, 'models'), 'cnn', \
# current_step=str(current_step))
#SavePB(sess, g1, os.path.join(model_dir, 'models'), 'cnn')
sess.close()
# Parameters
#tf.flags.DEFINE_string("test_data", "../gen_class_data/test_data/0617tag-3-20.test", "")
tf.flags.DEFINE_string(
"test_data",
"/mnt/workspace/yezhenxu/data/termlevel/len40/test/topic.s2id", "")
#tf.flags.DEFINE_string("test_data", "../gen_class_data/train-0614-20/train-0613-test.seg.id", "")
#tf.flags.DEFINE_string("model_dir", "./model_test0617-20/models/cnn.pb10000", "")
tf.flags.DEFINE_string(
"model_dir",
"/mnt/workspace/yezhenxu/model/term_model/len40/models/cnn.pb10000", "")
FLAGS = tf.flags.FLAGS
FLAGS._parse_flags()
x, y = data.LoadDevData(FLAGS.test_data)
print "**********************"
print x
print "*****************************"
print y
cat_num_ary = [0, 0, 0, 0, 0, 0, 0, 0, 0]
cat_right_num_ary = [0, 0, 0, 0, 0, 0, 0, 0, 0]
cat_max_ary = [0, 0, 0, 0, 0, 0, 0, 0, 0]
cat_min_ary = [1.1, 1.1, 1.1, 1.1, 1.1, 1.1, 1.1, 1.1, 1.1]
cat_total_ary = [0, 0, 0, 0, 0, 0, 0, 0, 0]
cat_recall_ary = [0, 0, 0, 0, 0, 0, 0, 0, 0]
cat_num_total = 0
for i in range(len(y)):
index = get_index(y[i])
if index < 0: