class agent(): def __init__(self,is_training): # Place Holder self.word = tf.placeholder(dtype=tf.int32, shape=[None, FLAGS.max_length], name='input_word') self.pos1 = tf.placeholder(dtype=tf.int32, shape=[None, FLAGS.max_length], name='input_pos1') self.pos2 = tf.placeholder(dtype=tf.int32, shape=[None, FLAGS.max_length], name='input_pos2') self.length = tf.placeholder(dtype=tf.int32, shape=[None], name='input_length') self.mask = tf.placeholder(dtype=tf.int32, shape=[None, FLAGS.max_length], name='input_mask') self.label = tf.placeholder(dtype=tf.int32, shape=[None], name='label') self.label_for_select = tf.placeholder(dtype=tf.int32, shape=[None], name='label_for_select') self.scope = tf.placeholder(dtype=tf.int32, shape=[FLAGS.batch_size + 1], name='scope') self.weights = tf.placeholder(dtype=tf.float32, shape=[None]) self.data_word_vec = np.load(os.path.join(FLAGS.export_path, 'vec.npy')) self.reward_holder = tf.placeholder(dtype = tf.float32, shape=[None],name='reward') self.action_holder = tf.placeholder(dtype = tf.int32, shape=[None],name='action') # Network self.embedding = Embedding(is_training, self.data_word_vec, self.word, self.pos1, self.pos2) self.encoder = Encoder(is_training, FLAGS.drop_prob) self.selector = Selector(FLAGS.num_classes, is_training, FLAGS.drop_prob) self.classifier = Classifier(is_training, self.label, self.weights) #compute self.word_embedding = self.embedding.word_embedding() self.pos_embedding = self.embedding.pos_embedding() self.embedding = self.embedding.concat_embedding(self.word_embedding, self.pos_embedding) self.x = self.encoder.cnn(self.embedding, FLAGS.hidden_size, self.mask, activation=tf.nn.relu) self.logit, self.repre = self.selector.no_bag(self.x) self.outputvalue=self.classifier.outputvalue(self.logit) self.output = self.classifier.output(self.logit) self.softmax_outputs=tf.nn.softmax(self.logit) self.action_onehot=tf.one_hot(indices=self.action_holder,depth=FLAGS.num_classes,dtype=tf.float32) self.step=tf.multiply(self.softmax_outputs,self.action_onehot) self.action_outputvalue=tf.reduce_sum(self.step,axis=1) self.temp=tf.log(self.action_outputvalue)*self.reward_holder self.loss=-tf.reduce_mean(tf.log(self.action_outputvalue)*self.reward_holder)#reduce_mean里面的是188个句子的loss self.loss_pre = self.classifier.softmax_cross_entropy(self.logit) self.optimizer_pre = tf.train.GradientDescentOptimizer(0.5) self.grads_and_vars = self.optimizer_pre.compute_gradients(self.loss_pre) self.train_op = self.optimizer_pre.apply_gradients(self.grads_and_vars) self.optimizer = tf.train.AdamOptimizer(0.0001) self.train_op_rl=self.optimizer.minimize(self.loss) self.tvars = tf.trainable_variables()
class PCNN():
def __init__(self, is_training):
#Place Holder
self.word = tf.placeholder(dtype=tf.int32,
shape=[None, FLAGS.max_length],
name='input_word')
self.pos1 = tf.placeholder(dtype=tf.int32,
shape=[None, FLAGS.max_length],
name='input_pos1')
self.pos2 = tf.placeholder(dtype=tf.int32,
shape=[None, FLAGS.max_length],
name='input_pos2')
self.length = tf.placeholder(dtype=tf.int32,
shape=[None],
name='input_length')
self.mask = tf.placeholder(dtype=tf.int32,
shape=[None, FLAGS.max_length],
name='input_mask')
self.scope = tf.placeholder(dtype=tf.int32,
shape=[FLAGS.batch_size + 1],
name='scope')
self.bag_label = tf.placeholder(dtype=tf.int32,
shape=[None],
name='bag_label')
self.sentence_label = tf.placeholder(dtype=tf.int32,
shape=[None],
name='sentence_label')
self.label_weights = tf.placeholder(dtype=tf.float32,
shape=[FLAGS.batch_size])
self.data_word_vec = np.load(os.path.join(FLAGS.export_path,
'vec.npy'))
#Network
self.embedding = Embedding(is_training, self.data_word_vec, self.word,
self.pos1, self.pos2)
self.encoder = Encoder(is_training, FLAGS.drop_prob)
self.selector = Selector(FLAGS.num_classes, is_training,
FLAGS.drop_prob)
self.classifier = Classifier(is_training, self.bag_label,
self.label_weights)
#compute
self.word_embedding = self.embedding.word_embedding()
self.pos_embedding = self.embedding.pos_embedding()
self.embedding = self.embedding.concat_embedding(
self.word_embedding, self.pos_embedding)
self.x = self.encoder.pcnn(self.embedding,
FLAGS.hidden_size,
self.mask,
activation=tf.nn.relu)
self.logit, self.repre = self.selector.attention(
self.x, self.scope, self.sentence_label)
#用于判断ds和selected哪一个好,与优化无关
self.label_onehot = tf.one_hot(indices=self.bag_label,
depth=FLAGS.num_classes,
dtype=tf.float32)
self.bag_loss_temp = tf.nn.softmax_cross_entropy_with_logits(
labels=self.label_onehot, logits=self.logit)
self.bag_loss = tf.reshape(self.bag_loss_temp, [1, -1])
self.loss_mean = tf.reduce_mean(self.bag_loss)
#计算reward
self.softmax_output = tf.nn.softmax(self.logit)
self.reward = tf.log(
tf.reduce_sum(self.label_onehot * self.softmax_output, axis=1))
#self.loss_mine = -tf.reduce_mean(self.reward, axis=0)这个就是loss一样的，只是没有加上权重
#计算梯度下降
self.loss = self.classifier.softmax_cross_entropy(self.logit)
#self.loss_one = self.classifier.softmax_cross_entropy(self.logit_one)
self.output = self.classifier.output(self.logit)
#self.output_one = self.classifier.output(self.logit_one)
self.outputvalue = self.classifier.outputvalue(self.logit)
self.test_output = tf.argmax(self.logit, 1) #输出什么关系
self.test_outputvalue = tf.reduce_max(self.logit, axis=1) #输出关系的概率
# Optimizer
self.global_step = tf.Variable(0, name='global_step', trainable=False)
tf.summary.scalar('learning_rate', FLAGS.learning_rate)
self.optimizer = tf.train.GradientDescentOptimizer(FLAGS.learning_rate)
self.grads_and_vars = self.optimizer.compute_gradients(self.loss)
self.train_op = self.optimizer.apply_gradients(
self.grads_and_vars, global_step=self.global_step)