'bidirectional_rnn' question

gahu1125 commented 7 years ago

I'm using Tensorflow 1.0 and I got an error message as below while running model.py.

AttributeError: module 'tensorflow.python.ops.nn' has no attribute 'bidirectional_rnn'

I looked up this problem in StackOverflow, and users told me that:

In TensorFlow 1.0, you have the choice of two bidirectional RNN functions: tf.nn.bidirectional_dynamic_rnn() tf.contrib.rnn.static_bidirectional_rnn()

Any idea which bidirectional RNN should I change to?

utkrist commented 7 years ago

@gahu1125 It really depends on your use case. If you don't know your nsteps parameter in advance, you can set it None when defining input or output placeholder and use the dynamic one but if you know your max_nstetps in advance and have sufficient gpu memory you can even use the static verison.

I'm posting an snippet from my code for your reference:

def multi_layer_birnn_static(config, input, seq_len, dropout):
    nhidden = config.nb_hidden
    ntags   = config.out_dim
    nsteps  = config.nb_steps
    nlayers = config.nb_layers
    cell    = rnn_cell(config.cell_type)

    # input shape: (batch_size, nsteps, in_dim)
    # Unstack to get a list of 'n_steps' tensors of shape (batch_size, n_input)
    input = tf.unstack(input, nsteps, 1)

    def _single_cell():
        _cell = cell(num_units=nhidden, state_is_tuple=True)
        _cell = tf.contrib.rnn.DropoutWrapper(_cell, output_keep_prob=dropout)
        return _cell

    fw_cell = tf.contrib.rnn.MultiRNNCell(cells=[_single_cell() for _ in range(nlayers)], state_is_tuple = True)
    bw_cell = tf.contrib.rnn.MultiRNNCell(cells=[_single_cell() for _ in range(nlayers)], state_is_tuple = True)

    output, _, _ = tf.contrib.rnn.static_bidirectional_rnn(fw_cell, bw_cell, input ,dtype=tf.float32)
    output = tf.stack(output, 1)
    return output

def multi_layer_birnn_dynamic(config, input, seq_len, dropout):
    nhidden = config.nb_hidden
    ntags   = config.out_dim
    nsteps  = config.nb_steps
    nlayers = config.nb_layers
    cell    = rnn_cell(config.cell_type)

   # permute n_steps and batch_size
   input = tf.transpose(input, [1, 0, 2]) 

    def _single_cell():
        _cell = cell(num_units=nhidden, state_is_tuple=True)
        _cell = tf.contrib.rnn.DropoutWrapper(_cell, output_keep_prob=dropout)
        return _cell

    fw_cell = tf.contrib.rnn.MultiRNNCell(cells=[_single_cell() for _ in range(nlayers)], state_is_tuple = True)
    bw_cell = tf.contrib.rnn.MultiRNNCell(cells=[_single_cell() for _ in range(nlayers)], state_is_tuple = True)

    outputs, states = tf.nn.bidirectional_dynamic_rnn(
        cell_fw=fw_cell, 
        cell_bw=bw_cell,
        dtype=tf.float32,
        inputs=input,
        time_major=True, 
        sequence_length=seq_len)
    out_fw, out_bw = outputs
    output = tf.concat([out_fw, out_bw], axis=-1)
    output = tf.transpose(output, [1, 0 ,2])
    return output