Python rnn.bidirectional_rnn方法代码示例

# 需要导入模块: from tensorflow.models.rnn import rnn [as 别名]
# 或者: from tensorflow.models.rnn.rnn import bidirectional_rnn [as 别名]
def __init__(self, config):
        sent_len = config.sent_len
        batch_size = config.batch_size
        vocab_size = config.vocab_size
        embed_size = config.embed_size
        num_layers = config.num_layers
        state_size = config.state_size
        keep_prob = config.keep_prob

        self.input_data = tf.placeholder(tf.int32, [batch_size, sent_len])
        self.lengths = tf.placeholder(tf.int64, [batch_size])
        self.targets = tf.placeholder(tf.float32, [batch_size, 1])

        # Get embedding layer which requires CPU
        with tf.device("/cpu:0"):
            embeding = tf.get_variable("embeding", [vocab_size, embed_size])
            inputs = tf.nn.embedding_lookup(embeding, self.input_data)

        #LSTM 1 -> Encode the characters of every tok into a fixed dense representation
        with tf.variable_scope("rnn1", reuse=None):
            cell = rnn_cell.LSTMCell(state_size, input_size=embed_size, initializer=tf.contrib.layers.xavier_initializer())
            back_cell = rnn_cell.LSTMCell(state_size, input_size=embed_size, initializer=tf.contrib.layers.xavier_initializer())
            cell = rnn_cell.DropoutWrapper(
              cell, input_keep_prob=keep_prob,
                         output_keep_prob=keep_prob)
            back_cell = rnn_cell.DropoutWrapper(
              back_cell, input_keep_prob=keep_prob,
                              output_keep_prob=keep_prob) 
            cell = rnn_cell.MultiRNNCell([cell] * num_layers)
            backcell = rnn_cell.MultiRNNCell([back_cell] * num_layers)
            
            rnn_splits = [tf.squeeze(input_, [1]) for input_ in tf.split(1, sent_len, inputs)]

            # Run the bidirectional rnn
            outputs, last_fw_state, last_bw_state = rnn.bidirectional_rnn(
                                                        cell, backcell, rnn_splits,
                                                        sequence_length=self.lengths,
                                                        dtype=tf.float32)
        
        sent_out = tf.concat(1, [last_fw_state, last_bw_state])
        #sent_out = outputs[-1]
        #sent_out = tf.add_n(outputs)
        output_size = state_size*4

        with tf.variable_scope("linear", reuse=None):
            w = tf.get_variable("w", [output_size, 1])
            b = tf.get_variable("b", [1], initializer=tf.constant_initializer(0.0))
            raw_logits = tf.matmul(sent_out, w) + b 
        self.probabilities = tf.sigmoid(raw_logits)
        self.cost = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(raw_logits, self.targets))

        #Calculate gradients and propagate
        #Aggregation method 2 is really important for rnn per the tensorflow issues list
        tvars = tf.trainable_variables()
        self.lr = tf.Variable(0.0, trainable=False) #Assign to overwrite
        optimizer = tf.train.AdamOptimizer()
        grads, _vars = zip(*optimizer.compute_gradients(self.cost, tvars, aggregation_method=2))
        grads, self.grad_norm = tf.clip_by_global_norm(grads,
                                      config.max_grad_norm)
        self.train_op = optimizer.apply_gradients(zip(grads, _vars)) 

# 需要导入模块: from tensorflow.models.rnn import rnn [as 别名]
# 或者: from tensorflow.models.rnn.rnn import bidirectional_rnn [as 别名]
def BiRNN(x, weights, biases):

    # Prepare data shape to match `bidirectional_rnn` function requirements
    # Current data input shape: (batch_size, n_steps, n_input)
    # Permuting batch_size and n_steps
    x = tf.transpose(x, [1, 0, 2])
    # Reshape to (n_steps*batch_size, n_input)
    x = tf.reshape(x, [-1, n_input])
    # Split to get a list of 'n_steps' tensors of shape (batch_size, n_hidden)
    x = tf.split(0, n_steps, x)

    # Define lstm cells with tensorflow
    # Forward direction cell
    lstm_fw_cell = rnn_cell.BasicLSTMCell(n_hidden, forget_bias=1.0)
    # Backward direction cell
    lstm_bw_cell = rnn_cell.BasicLSTMCell(n_hidden, forget_bias=1.0)

    # Get lstm cell output
    '''
    ????,????????????lstm_cell???????????bidirectional_rnn??????,?????????feed???back
    '''
    outputs = rnn.bidirectional_rnn(lstm_fw_cell, lstm_bw_cell, x,
                                    dtype=tf.float32)

    # Linear activation, using rnn inner loop last output
    return tf.matmul(outputs[-1], weights['out']) + biases['out']