def test():
print(" >> Loading preprocessing data...", "\n")
parameters, data_info = load_preprocessing()
print(" >> Loading Test Dataset...", "\n")
TF_IDF_Feature_Matrix, target_idx_list = data_info.read_student_feature(parameters['dataset_testset'])
test_idx, test_target_idx = data_info.make_data(target_idx_list, TF_IDF_Feature_Matrix, is_test = True)
test_data = data_helpers.batch_construction(test_target_idx, test_idx)
session_conf = tf.ConfigProto()
session_conf.gpu_options.allow_growth = True
with tf.Session(config=session_conf) as sess:
Model = load_model(sess, parameters, data_info)
test_input_indices, test_target_indices, test_target_origin =\
data_helpers.get_minibatch(dataset=test_data,\
minibatch_seq=np.arange(len(test_data)),\
is_test = True)
feed_dict = {
Model.X: test_input_indices,\
Model.Y: test_target_indices
}
test_logits = sess.run([Model.softmax_output], feed_dict=feed_dict)
# Save the output of softmax layer
np.savetxt(fname=parameters['output_path'], X=test_logits[0], \
fmt='%.10f', delimiter = '\t')
np.savetxt(fname='answer.txt', X=test_target_indices, fmt='%d')
print(" >> End of Test...")
print(" >> Check 'output.txt' and 'answer.txt' file...")
print("")
def valid_check(current_session, valid_data, Model):
valid_input_indices, valid_target_indices = data_helpers.get_minibatch( \
dataset=valid_data, \
minibatch_seq=np.arange(len(valid_data)))
feed_dict = {
Model.X: valid_input_indices, \
Model.Y: valid_target_indices
}
valid_pred, valid_loss, valid_accuracy = current_session.run([Model.prediction, Model.loss, Model.accuracy], \
feed_dict=feed_dict)
return valid_pred, valid_loss, valid_accuracy
def train():
print(" >> Loading preprocessing information...", "\n")
parameters, data_info = load_preprocessing()
print(" >> Loading Train Data...", "\n")
train_data = data_info.train_data
valid_data = data_info.valid_data
valid_loss_history = []
valid_acc_history = []
train_loss_history = []
train_acc_history = []
bad_counter = 0
previous_min_valid_loss = 1000
session_conf = tf.ConfigProto()
session_conf.gpu_options.allow_growth = True
with tf.Session(config=session_conf) as sess:
Model = create_model(sess, parameters, data_info)
if Model.global_epoch_step.eval() + 1 > parameters['n_epoch']:
print(" >> Current Epoch: {}, Max Epoch: {}".format(
Model.global_epoch_step.eval(), parameters['n_epoch']))
print(" >> End of Training....")
#exit(-1)
return
for epoch_idx in range(parameters['n_epoch']):
try:
batches = data_helpers.batch_iter(parameters, train_data)
for minibatch in batches:
#print (minibatch)
input_indices, target_indices = data_helpers.get_minibatch(\
dataset = train_data,\
minibatch_seq = minibatch)
feed_dict = {
Model.X : input_indices,\
Model.Y : target_indices}
# Training model......
_, global_step, minibatch_loss, minibatch_accuracy = sess.run(\
[Model.train_op, Model.global_step, Model.loss, Model.accuracy], feed_dict)
# Validation Check
if (global_step + 1) % parameters['evaluation_every'] == 0:
#Validation Set
valid_pred, valid_loss, valid_accuracy = valid_check(current_session = sess,\
valid_data = valid_data,\
Model = Model)
# 매 "evaluation_every" step마다 train의 결과를 저장!!!
train_loss_history.append(minibatch_loss)
train_acc_history.append(minibatch_accuracy)
valid_loss_history.append(valid_loss)
valid_acc_history.append(valid_accuracy)
print("")
print(" >> Global_Step # {} at {}-epoch".format(
global_step, Model.global_epoch_step.eval()))
print(
" - Train Loss (Validation_Loss) : {:,.2f} ({:,.2f})"
.format(minibatch_loss, valid_loss))
print(
" - Train Accuracy (Validation Accuracy) : {:,.2f} ({:.2f})"
.format(minibatch_accuracy, valid_accuracy))
print("")
if valid_loss <= previous_min_valid_loss:
bad_counter = 0
previous_min_valid_loss = valid_loss
#save the model checkpoint
checkpoint_path = os.path.join(
os.path.join(parameters['save_dir'],
parameters['model_name']), 'ckpt')
saver = tf.train.Saver()
saver.save(sess, checkpoint_path, global_step)
print(
" >> Saving the current model with loss {:,.2f} at {}"
.format(valid_loss, checkpoint_path))
print("")
else:
bad_counter += 1
# Early Stopping
if bad_counter > parameters['patience']:
print(" >> EARLY STOPPING with bad_counter {}".
format(bad_counter))
print(" >> Training Process Terminated....")
#exit(-1)
return
Model.global_epoch_step_op.eval() #Increment Global_epoch_step
except KeyboardInterrupt:
print(" >> Interrupted by user at {}-epoch, {}-global_step".
format(Model.global_epoch_step.eval(), global_step))
# save the model checkpoint
saver = tf.train.Saver()
checkpoint_path = os.path.join(parameters['save_dir'],
parameters['model_name'])
saver.save(sess, checkpoint_path, global_step)
print(" - Saving the model with {}-epoch in {}".format(
Model.global_epoch_step.eval(), checkpoint_path))
print(" - Training Process Terminated....")
#exit(-1)
return
print(" >> Save the last model...")
saver = tf.train.Saver()
checkpoint_path = os.path.join(parameters['save_dir'], 'MLP.ckpt')
saver.save(sess, checkpoint_path, global_step=global_step)
print(" >> End of Training...")
print("")
print("")
def train():
print(" >> Loading preprocessing information...", "\n")
parameters, data_info = load_preprocessing()
print(" >> Loading Train Data...", "\n")
train_data = data_info.train_data
train_loss_history = []
train_acc_history = []
session_conf = tf.ConfigProto()
session_conf.gpu_options.allow_growth = True
with tf.Session(config=session_conf) as sess:
Model = create_model(sess, parameters, data_info)
if Model.global_epoch_step.eval() + 1 > parameters['n_epoch']:
print(" >> Current Epoch: {}, Max Epoch: {}".format(
Model.global_epoch_step.eval(), parameters['n_epoch']))
print(" >> End of Training....")
exit(-1)
for epoch_idx in range(parameters['n_epoch']):
try:
batches = data_helpers.batch_iter(parameters, train_data)
for minibatch in batches:
#print (minibatch)
input_indices, target_indices = data_helpers.get_minibatch(\
dataset = train_data,\
minibatch_seq = minibatch)
feed_dict = {
Model.X : input_indices,\
Model.Y : target_indices}
# Training model......
_, global_step, minibatch_loss, minibatch_accuracy = sess.run(\
[Model.train_op, Model.global_step, Model.loss, Model.accuracy], feed_dict)
# Check Training Process
if (global_step + 1) % parameters['evaluation_every'] == 0:
# 매 "evaluation_every" step마다 train의 결과를 저장!!!
train_loss_history.append(minibatch_loss)
train_acc_history.append(minibatch_accuracy)
print(" >> Global_Step # {} at {}-epoch".format(
global_step, Model.global_epoch_step.eval()))
print(" - Train Loss : {:,.2f}".format(
minibatch_loss))
print(" - Train Accuracy: {:,.2f}".format(
minibatch_accuracy))
print("")
Model.global_epoch_step_op.eval() #Increment Global_epoch_step
except KeyboardInterrupt:
print(" - Training Process Terminated....")
exit(-1)
print(" >> Save the last model...")
saver = tf.train.Saver()
checkpoint_path = os.path.join(parameters['save_dir'], 'MLP.ckpt')
saver.save(sess, checkpoint_path, global_step=global_step)
print(" >> End of Training...")
print("")
print("")