for i in range(1000 * FLAGS.NUM_OF_EPOCH):
# i = 0
if i % Num_of_Iterlation == 1000:
epoch += 1
FLAGS.lr_value *= FLAGS.lr_decay
################################################################
# Training batch OPTIMIZE
################################################################
index = utils.sampler(LABEL_POS=TRAIN_LABEL_POS,
LABEL_NEG=TRAIN_LABEL_NEG,
BATCH_SIZE=FLAGS.BATCH_SIZE)
batch_input, batch_label = utils.generate_batch_jaso(INDEX=index,
MODEL=model,
DOC=TRAIN_DOC,
LABEL=TRAIN_LABEL,
MAXLEN=FLAGS.MAXLEN,
SESS=sess,
ATTENTION=False)
seq_length, _ = utils.length(batch_input)
_ = sess.run(
[model.optm],
feed_dict={
model.X: batch_input,
model.Y: batch_label,
model.LEARNING_RATE: FLAGS.lr_value,
model.SEQ: seq_length,
model.Dropout_Rate1: FLAGS.Dropout_Rate1,
model.Dropout_Rate2: FLAGS.Dropout_Rate2,
model.TRAIN_PH: True
def main():
args = parse_args()
if args is None:
exit()
# test file
filename = args.inputdir + args.input
# init model config
#TRAIN_DOC, TRAIN_LABEL, TEST_DOC, TEST_LABEL, LABEL_IDX = data_load.digi_data_load()
TRAIN_DOC, TRAIN_LABEL, TEST_DOC, TEST_LABEL, LABEL_IDX = data_load.testcase_shuffle_data_load(
)
#data_load.testcase_add_data_load()
class_num = TRAIN_LABEL.shape[1]
FLAGS.NUM_OF_CLASS = class_num
JM = utils.lookup_JM(FLAGS.INPUT_WIDTH, FLAGS.INPUT_DEPTH)
# Start Session
sess = tf.Session()
print("Session Ready!")
model = MODEL(sess=sess, JM=JM, FLAGS=FLAGS)
# Initialization
sess.run(tf.global_variables_initializer())
model.JM.init_table(sess)
# Restore parameter
saver = tf.train.Saver()
saver.restore(sess, "./Saver/{}/{}.ckpt".format(FLAGS.WRITER,
FLAGS.WRITER))
if args.printtest == 'True':
index = np.array(range(0, len(TEST_DOC)))
batch_input, batch_label = utils.generate_batch_jaso(
INDEX=index,
MODEL=model,
DOC=TEST_DOC,
LABEL=TEST_LABEL,
MAXLEN=FLAGS.INPUT_WIDTH,
SESS=sess)
proba, ts_loss, ts_acc, ts_merged = sess.run(
[model.y_proba, model.cross_entropy, model.accuracy, model.merge],
feed_dict={
model.X: batch_input,
model.Y: batch_label,
model.LEARNING_RATE: FLAGS.lr_value,
model.TRAIN_PH: False
})
pred_idx = np.apply_along_axis(np.argmax, 1, proba)
real_idx = np.apply_along_axis(np.argmax, 1, batch_label)
pos_idx = np.where(np.equal(pred_idx, real_idx) == True)[0]
neg_idx = np.where(np.equal(pred_idx, real_idx) == False)[0]
print('[ TEST ]')
desc = """
size:{}, correct:{}, wrong:{}, acc:{}, f1_score:{}, ts_loss:{}, ts_acc:{}
""".format(index.shape[0], pos_idx.shape[0], neg_idx.shape[0],
round(pos_idx.shape[0] / index.shape[0] * 100, 3),
round(f1_score(real_idx, pred_idx, average='weighted'), 4),
ts_loss, ts_acc)
print(desc)
for idx in pos_idx:
print('Positive Case:\t', TEST_DOC[index[idx]], '\t->\t',
LABEL_IDX[np.argmax(proba[idx])],
'({0:.2f})\t'.format(round(max(proba[idx]), 3)),
LABEL_IDX[np.argmax(batch_label[idx])])
for idx in neg_idx:
print(
'Negative Case:\t', TEST_DOC[index[idx]], '\t->\t',
LABEL_IDX[np.argmax(proba[idx])],
'({0:.2f})\t'.format(round(max(proba[idx]), 3)),
LABEL_IDX[np.argmax(batch_label[idx])],
'({0:.2f})\t'.format(proba[idx][np.argmax(batch_label[idx])]))
print()
if args.input == 'a.txt':
testfile = open(filename)
for line in testfile:
line = line.rstrip('\n\r').lower()
jaso_splitted = jmu.jaso_split([line], MAXLEN=FLAGS.INPUT_WIDTH)
batch_input = sess.run(model.jaso_Onehot,
{model.X_Onehot: jaso_splitted})
y_proba = sess.run(model.y_proba,
feed_dict={
model.X: batch_input,
model.TRAIN_PH: False
})
#print(batch_input.shape, y_proba.shape)
label = LABEL_IDX[np.argmax(y_proba[0])]
if round(max(y_proba[0]), 3) > 0.8:
print(line, '\t->\t', label, round(max(y_proba[0]), 3))
if args.input == 'tmp.txt':
tmp_df = pd.read_csv(filename, sep='\t', header=None)
tmp_df['class'] = tmp_df[1].apply(lambda x: json2intent(x))
sentences = tmp_df[0].tolist()
labels = []
probas = []
for sent in sentences:
jaso_splitted = jmu.jaso_split([sent], MAXLEN=FLAGS.INPUT_WIDTH)
batch_input = sess.run(model.jaso_Onehot,
{model.X_Onehot: jaso_splitted})
y_proba = sess.run(model.y_proba,
feed_dict={
model.X: batch_input,
model.TRAIN_PH: False
})
# print(batch_input.shape, y_proba.shape)
label = LABEL_IDX[np.argmax(y_proba[0])]
labels.append(label)
probas.append(round(max(y_proba[0]), 3))
tmp_df['pred'] = labels
tmp_df['proba'] = probas
for index, row in tmp_df.iterrows():
# if row['class'] is None:
# continue
if row['class'] is not None:
continue
if round(row['proba'], 3) > 0.8:
print(row[0], '\t->\t', row['pred'], round(row['proba'], 3),
'\t', row['class'])
if args.input == 'tmp_not_under.txt':
tmp_df = pd.read_csv(filename, sep='\t', header=None)
tmp_df['class'] = tmp_df[1].apply(lambda x: json2intent(x))
sentences = tmp_df[0].tolist()
labels = []
probas = []
for sent in sentences:
jaso_splitted = jmu.jaso_split([sent], MAXLEN=FLAGS.INPUT_WIDTH)
batch_input = sess.run(model.jaso_Onehot,
{model.X_Onehot: jaso_splitted})
y_proba = sess.run(model.y_proba,
feed_dict={
model.X: batch_input,
model.TRAIN_PH: False
})
# print(batch_input.shape, y_proba.shape)
label = LABEL_IDX[np.argmax(y_proba[0])]
labels.append(label)
probas.append(round(max(y_proba[0]), 3))
tmp_df['pred'] = labels
tmp_df['proba'] = probas
for index, row in tmp_df.iterrows():
if round(row['proba'], 3) > 0.1:
print(row[1])
print(row[0], '\t->\t', row['pred'], round(row['proba'], 3),
'\t')
print()
################################################################################
# 한 epoch 당 iteration 횟수
Num_of_Iterlation = np.shape(TRAIN_DOC)[0] // FLAGS.BATCH_SIZE
epoch = 0
for i in range(Num_of_Iterlation * FLAGS.NUM_OF_EPOCH):
# i=0
if i % Num_of_Iterlation == 0:
epoch += 1
FLAGS.lr_value *= FLAGS.lr_decay
################################################################
# Training batch OPTIMIZE
################################################################
#index = utils.sampler(LABEL_POS=TRAIN_LABEL_POS, LABEL_NEG=TRAIN_LABEL_NEG, BATCH_SIZE=FLAGS.BATCH_SIZE)
index = np.random.choice(range(0, len(TRAIN_DOC)), FLAGS.BATCH_SIZE)
batch_input, batch_label = utils.generate_batch_jaso(INDEX=index, MODEL=model, DOC=TRAIN_DOC,
LABEL=TRAIN_LABEL, MAXLEN=FLAGS.INPUT_WIDTH, SESS=sess)
_ = sess.run([model.optm],
feed_dict={model.X: batch_input,
model.Y: batch_label,
model.LEARNING_RATE: FLAGS.lr_value,
model.TRAIN_PH: True})
################################################################
# Calculate Train & Test Loss, Accuracy & Summary / Print
################################################################
if i % 10 == 0:
################################################################
# Train batch LOSS CHECK
################################################################