k1 = 19
num_filters = [6, 14]
dev = 300
batch_size = 50
n_epochs = 30
num_hidden = 100
sentence_length = 37
num_class = 6
lr = 0.01
evaluate_every = 100
checkpoint_every = 100
num_checkpoints = 5
# Load data
print("Loading data...")
x_, y_, vocabulary, vocabulary_inv, test_size = dataUtils.load_data()
#x_:长度为5952的np.array。（包含5452个训练集和500个测试集）其中每个句子都是padding成长度为37的list（padding的索引为0）
#y_:长度为5952的np.array。每一个都是长度为6的onehot编码表示其类别属性
#vocabulary：长度为8789的字典，说明语料库中一共包含8789各单词。key是单词，value是索引
#vocabulary_inv：长度为8789的list，是按照单词出现次数进行排列。依次为：<PAD?>,\\?,the,what,is,of,in,a....
#test_size:500,测试集大小
# Randomly shuffle data
x, x_test = x_[:-test_size], x_[-test_size:]
y, y_test = y_[:-test_size], y_[-test_size:]
shuffle_indices = np.random.permutation(np.arange(len(y)))
x_shuffled = x[shuffle_indices]
y_shuffled = y[shuffle_indices]
x_train, x_dev = x_shuffled[:-dev], x_shuffled[-dev:]
y_train, y_dev = y_shuffled[:-dev], y_shuffled[-dev:]
saver.restore(sess, checkpoint_prefix + '-' + str(best_at_step))
print('Finish training. On test set:')
acc, loss = dev_step(self.x_test, self.y_test, writer=None)
print(acc, loss)
if __name__ == "__main__":
for num_aug in range(0, 17):
aug_split(num_aug=num_aug)
train_path = './sougou/train_aug' + str(num_aug) + '.txt'
test_path = './sougou/test_aug' + str(num_aug) + '.txt'
dev_percent = 0.05
# Load data
print("Loading data...")
x_, y_, vocabulary, vocabulary_inv, train_size, test_size, sent_length, num_class = dataUtils.load_data(
train_path, test_path)
x, x_test = x_[:-test_size], x_[-test_size:]
y, y_test = y_[:-test_size], y_[-test_size:]
shuffle_indices = np.random.permutation(np.arange(len(y)))
x_shuffled = x[shuffle_indices]
y_shuffled = y[shuffle_indices]
x_train, x_dev = x_shuffled[:int(-dev_percent * train_size
)], x_shuffled[int(-dev_percent *
train_size):]
y_train, y_dev = y_shuffled[:int(-dev_percent * train_size
)], y_shuffled[int(-dev_percent *
train_size):]
print("Train/Dev/Test split: {:d}/{:d}/{:d}".format(
def train(fliter_size, stride, embeding_size):
embedding = np.load(r"G:\MR\embedding_matrix_{}.npy".format(embeding_size))
sent = tf.placeholder(tf.int64, [None, sentence_length])
y = tf.placeholder(tf.float64, [None, num_class])
seq_len = tf.placeholder(tf.int32, [None])
dropout_blstm_prob = tf.placeholder(tf.float32, name="dropout")
dropout_word_prob = tf.placeholder(tf.float32, name="dropout")
model = Blstm_att(batch_size, sentence_length, embeding_size,
embeding_size, num_label, embedding, seq_len,
regularizer_rate, dropout_blstm_prob, dropout_word_prob)
out_trian = model.Blstm_att(sent, fliter_size, stride, True)
global_step = tf.Variable(0, name="global_step", trainable=False)
with tf.name_scope("cost"):
reg_loss = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
cost_train = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits=out_trian,
labels=y))
loss_train = cost_train + regularizer_rate * sum(reg_loss)
with tf.name_scope("acc"):
Acc = tf.reduce_mean(
tf.cast(tf.equal(tf.argmax(y, 1), tf.argmax(out_trian, 1)),
tf.float32))
optimizer = tf.train.AdadeltaOptimizer(1.0)
tvars = tf.trainable_variables()
grads, _ = tf.clip_by_global_norm(tf.gradients(loss_train, tvars),
grad_clip)
grads_and_vars = tuple(zip(grads, tvars))
train_op = optimizer.apply_gradients(grads_and_vars,
global_step=global_step)
with tf.Session() as sess:
train_seq_len = np.ones(batch_size) * sentence_length
sess.run(tf.global_variables_initializer())
#=========save_the_model======
checkpoint_dir = os.path.abspath(
os.path.join(model_path, "checkpoints"))
checkpoint_prefix = os.path.join(checkpoint_dir, "model")
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
saver = tf.train.Saver()
#=============================
def dev(x_batch, y_batch):
feed_dict = {
sent: x_batch,
y: y_batch,
seq_len: train_seq_len,
dropout_blstm_prob: 1.0,
dropout_word_prob: 1.0
}
step, loss, acc = sess.run([global_step, cost_train, Acc],
feed_dict)
return loss, acc
def train(x_batch, y_batch):
feed_dict = {
sent: x_batch,
y: y_batch,
seq_len: train_seq_len,
dropout_blstm_prob: 0.5,
dropout_word_prob: 0.5
}
_, step, loss, acc = sess.run(
[train_op, global_step, loss_train, Acc], feed_dict)
print("Train Step{},loss {:g},acc {:g} ".format(step, loss, acc))
return loss, acc
max_acc = 0
valid = []
for i in range(n_epochs):
training_step, test_npochs, x_train, y_train, x_val, y_val = dataUtils.load_data(
batch_size, "G:\MR", embeding_size)
for i in range(training_step):
current_step = tf.train.global_step(sess, global_step)
start = i * batch_size
end = start + batch_size
batch_train = x_train[start:end]
train_label = y_train[start:end]
train(batch_train, train_label)
if current_step % 500 == 0:
dev_loss = 0
dev_acc = 0
for i in range(test_npochs):
start = i * batch_dev
end = start + batch_dev
batch_val = x_val[start:end]
val_label = y_val[start:end]
m, n = dev(batch_val, val_label)
dev_loss = m + dev_loss
dev_acc = n + dev_acc
print("\nValid Step{},loss {:g},acc {:g} \n".format(
current_step, dev_loss / test_npochs,
dev_acc / test_npochs))
if dev_acc / test_npochs > max_acc:
max_acc = dev_acc / test_npochs
print("the max acc: {:g}".format(max_acc))
path = saver.save(sess,
checkpoint_prefix,
global_step=current_step)
valid.append("Valid Step:{},acc:{:g},loss:{:g}".format(
current_step, dev_acc / test_npochs,
dev_loss / test_npochs))
np.save("MR_{}_.npy".format(embeding_size), valid)
ws = [7, 5]
top_k = 4
num_filters = [6, 14]
dev = 300
batch_size = 512
n_epochs = 50
num_hidden = 100
sentence_length = 52
lr = 1e-3
evaluate_every = 100
checkpoint_every = 100
num_checkpoints = 5
# Load data
print("Loading data...")
x_, y_, vocabulary, vocabulary_inv, test_size, dev_size = dataUtils.load_data(
num_class)
print(y_)
# Randomly shuffle data
x_train, x_dev, x_test = x_[:-(test_size + dev_size)], x_[-(
test_size + dev_size):-test_size], x_[-test_size:]
y_train, y_dev, y_test = y_[:-(test_size + dev_size)], y_[-(
test_size + dev_size):-test_size], y_[-test_size:]
print("Train/Dev/Test split: {:d}/{:d}/{:d}".format(len(y_train), len(y_dev),
len(y_test)))
#--------------------------------------------------------------------------------------#
def init_weights(shape, name):
return tf.Variable(tf.truncated_normal(shape, stddev=0.01), name=name)