# else:
# _pred = np.concatenate((_pred, pred))
# _true = np.concatenate((_true, y_batch))
_accs += _acc
_costs += _cost
mean_acc = _accs / batch_num
mean_cost = _costs / batch_num
cm = confusion_matrix(y_true=_true, y_pred=_pred)
return mean_acc, mean_cost, cm
if __name__ == '__main__':
max_len = 150
# data_train, data_valid, data_test = data_helper.getDataSet()
data_train, data_test = data_helper.getDataSet()
# wordEmbedding = data_helper.getWordEmbedding(targetFileName='./data/embeddingMatrix.300d.pkl')
wordEmbedding = data_helper.getRandomWordEmbedding(vocabularySize=10002,
embeddingSize=300)
# model = TextBiLSTM(wordEmbedding, textLength=max_len, vocabulary_size=10002, embedding_size=128)
model = TextBiLSTM(wordEmbedding,
textLength=max_len,
vocabulary_size=27590,
embedding_size=300)
# model = TextBiLSTM(wordEmbedding, textLength=max_len, vocabulary_size=42488, embedding_size=128)
config = tf.ConfigProto()
sess = tf.Session(config=config)
init = tf.global_variables_initializer()
sess.run(init)
def TrainAndTest():
max_len = 150
data_train, data_valid, data_test = data_helper.getDataSet()
# wordEmbedding = data_helper.getWordEmbedding(targetFileName='./data/embeddingMatrix.300d.pkl')
wordEmbedding = data_helper.getRandomWordEmbedding(vocabularySize=10002,
embeddingSize=300)
model = BiLSTMForMultiTask(wordEmbedding,
textLength=max_len,
embedding_size=300)
config = tf.ConfigProto()
sess = tf.Session(config=config)
init = tf.global_variables_initializer()
sess.run(init)
model_save_path = './ckpt/bilstmForMulitTask/bilstm.ckpt'
tr_batch_size = 128
decay = 0.85
max_epoch = 5
max_max_epoch = 12
# max_epoch = 12
# max_max_epoch = 25
display_num = 5
tr_batch_num = int(data_train.y.shape[0] / tr_batch_size)
display_batch = int(tr_batch_num / display_num)
for epoch in xrange(max_max_epoch):
_lr = 1e-3
if epoch > max_epoch:
_lr = _lr * ((decay)**(epoch - max_epoch))
print 'EPOCH {}, lr={}'.format(epoch + 1, _lr)
start_time = time.time()
_accs = np.zeros(shape=[4], dtype=np.float32)
_costs = 0.0
show_accs = np.zeros(shape=[4], dtype=np.float32)
show_costs = 0.0
for batch in xrange(tr_batch_num):
fetches = [model.accuracy, model.cost, model.train_op]
X_q_batch, X_r_batch, y_batch = data_train.next_batch(
tr_batch_size)
feed_dict = {
model.X_q_inputs: X_q_batch,
model.X_r_inputs: X_r_batch,
model.y_inputs: y_batch,
model.batch_size: tr_batch_size,
model.dropout_keep_prob: 0.5,
model.lr: _lr,
}
_acc, _cost, _ = sess.run(fetches, feed_dict)
_accs += _acc
_costs += _cost
show_accs += _acc
show_costs += _cost
if (batch + 1) % display_batch == 0:
valid_acc, valid_cost = test_epoch(sess, model, data_valid)
disagree_agree_acc = show_accs[0] / display_batch
attacking_respectful_acc = show_accs[1] / display_batch
emotion_fact_acc = show_accs[2] / display_batch
nasty_nice_acc = show_accs[3] / display_batch
print '\ttraining acc={}, {}, {}, {}, cost={}; valid acc={}, {}, {}, {}, cost={} '.format(
disagree_agree_acc, attacking_respectful_acc,
emotion_fact_acc, nasty_nice_acc,
show_costs / display_batch, valid_acc[0], valid_acc[1],
valid_acc[2], valid_acc[3], valid_cost)
show_accs = np.zeros(shape=[4], dtype=np.float32)
show_costs = 0.0
mean_acc = _accs / tr_batch_num
mean_cost = _costs / tr_batch_num
if (epoch + 1) % 3 == 0:
save_path = model.saver.save(sess,
model_save_path,
global_step=(epoch + 1))
print 'The save path is ', save_path
print '\ttraining {}, acc={}, {}, {}, {}, cost={}'.format(
data_train.y.shape[0], mean_acc[0], mean_acc[1], mean_acc[2],
mean_acc[3], mean_cost)
print 'Epoch training {}, acc={}, {}, {}, {}, cost={}, speed={} s/epoch\n'.format(
data_train.y.shape[0], mean_acc[0], mean_acc[1], mean_acc[2],
mean_acc[3], mean_cost,
time.time() - start_time)
print '**TEST RESULT:'
test_acc, test_cost = test_epoch(sess, model, data_test)
print '**Test {}, acc={}, {}, {}, {}, cost={}'.format(
data_test.y.shape[0], test_acc[0], test_acc[1], test_acc[2],
test_acc[3], test_cost)