Tensorflow: Replacement for tf.nn.rnn_cell._linear(input, size, 0, scope)

Question

I am trying to get the SequenceGAN (https://github.com/LantaoYu/SeqGAN) from https://arxiv.org/pdf/1609.05473.pdf to run.
After fixing the obvious errors, like replacing pack with stack, it still doesn't run, since the highway-network part requires the tf.nn.rnn_cell._linear function:

# highway layer that borrowed from https://github.com/carpedm20/lstm-char-cnn-tensorflow
def highway(input_, size, layer_size=1, bias=-2, f=tf.nn.relu):
    """Highway Network (cf. http://arxiv.org/abs/1505.00387).

    t = sigmoid(Wy + b)
    z = t * g(Wy + b) + (1 - t) * y
    where g is nonlinearity, t is transform gate, and (1 - t) is carry gate.
    """
    output = input_
    for idx in range(layer_size):
        output = f(tf.nn.rnn_cell._linear(output, size, 0, scope='output_lin_%d' % idx)) #tf.contrib.layers.linear instad doesn't work either.
        transform_gate = tf.sigmoid(tf.nn.rnn_cell._linear(input_, size, 0, scope='transform_lin_%d' % idx) + bias)
        carry_gate = 1. - transform_gate

        output = transform_gate * output + carry_gate * input_

    return output

the tf.nn.rnn_cell._linear function doesn't appear to be there anymore in Tensorflow 1.0 or 0.12, and I have no clue what to replace it with. I can't find any new implementations of this, or any information on tensorflow's github or (unfortunately very sparse) documentation.

Does anybody know the new pendant of the function? Thanks a lot in advance!

Solution 1

I met this error while using SkFlow's TensorFlowDNNRegressor. The first time I saw the answer of ruoho ruots, I am a bit confused. But the next day I realized what he meant.

Here is what I do:

from tensorflow.python.ops import rnn_cell_impl

replace tf.nn.rnn_cell._linear with rnn_cell_impl._linear

Solution 2

With version 1.0, stuff has moved all around. I've had similar hunts updating tf.nn.rnn_cell.LSTMCell to tf.contrib.rnn.BasicLSTMCell.

For your case tf.nn.rnn_cell._linear now lives in tf.contrib.rnn.python.ops.core_rnn_cell_impl as well as the definition of the BasicRNNCell. Checking the BasicRNNCell docs and source code, we see at L113-L118 the use of _linear.

  def __call__(self, inputs, state, scope=None):
    """Most basic RNN: output = new_state = act(W * input + U * state + B)."""
    with _checked_scope(self, scope or "basic_rnn_cell", reuse=self._reuse):
      output = self._activation(
          _linear([inputs, state], self._num_units, True))
    return output, output

the _linear method is defined at line 854 as a:
Linear map: sum_i(args[i] * W[i]), where W[i] is a variable.

Good luck!

Solution 3

The answer of ruoho ruotsi is almost correct: Yet, the definition of linear is not located in tf.contrib.rnn.basicRNNCell, but in tf.contrib.rnn.python.ops.rnn_cell, or tf.contrib.rnn.python.ops.core_rnn_cell_impl, respectively.

You can find their source code here and here.

Solution 4

To solving this problem,we can define a linear() function.

def linear(input_, output_size, scope=None):
    '''
    Linear map: output[k] = sum_i(Matrix[k, i] * args[i] ) + Bias[k]
    Args:
        args: a tensor or a list of 2D, batch x n, Tensors.
    output_size: int, second dimension of W[i].
    scope: VariableScope for the created subgraph; defaults to "Linear".
    Returns:
    A 2D Tensor with shape [batch x output_size] equal to
    sum_i(args[i] * W[i]), where W[i]s are newly created matrices.
    Raises:
    ValueError: if some of the arguments has unspecified or wrong shape.
    '''

    shape = input_.get_shape().as_list()
    if len(shape) != 2:
        raise ValueError("Linear is expecting 2D arguments: %s" % str(shape))
    if not shape[1]:
        raise ValueError("Linear expects shape[1] of arguments: %s" % str(shape))
    input_size = shape[1]

    # Now the computation.
    with tf.variable_scope(scope or "SimpleLinear"):
        matrix = tf.get_variable("Matrix", [output_size, input_size], dtype=input_.dtype)
        bias_term = tf.get_variable("Bias", [output_size], dtype=input_.dtype)

    return tf.matmul(input_, tf.transpose(matrix)) + bias_term


def highway(input_, size, num_layers=1, bias=-2.0, f=tf.nn.relu, scope='Highway'):
    """Highway Network (cf. http://arxiv.org/abs/1505.00387).
    t = sigmoid(Wy + b)
    z = t * g(Wy + b) + (1 - t) * y
    where g is nonlinearity, t is transform gate, and (1 - t) is carry gate.
    """

    with tf.variable_scope(scope):
        for idx in range(num_layers):
            g = f(linear(input_, size, scope='highway_lin_%d' % idx))

            t = tf.sigmoid(linear(input_, size, scope='highway_gate_%d' % idx) + bias)

            output = t * g + (1. - t) * input_
            input_ = output

    return output

https://github.com/mkroutikov/tf-lstm-char-cnn/blob/7e899e6992cbf9a96e6d791e5d364eaaeec339a2/model.py

Solution 5

tensorflow.python.ops.rnn_cell_impl._linear now is at tensorflow.contrib.rnn.python.ops.core_rnn_cell._linear. And I prefer to use tf.layers.Dense to replace. for example, change

from tensorflow.contrib.rnn.python.ops import core_rnn_cell
core_rnn_cell._linear(states, length, bias=True)

to

tf.layers.Dense(units=length)(states)

I'm using tensorflow 1.6.

View more solutions

Comments

[Majid Alfifi]

I tried this solution on v 1.0 but got AttributeError: type object 'BasicRNNCell' has no attribute '_linear'

[dennlinger]

It is actually defined in another file. Please see my answer for the correct code location.

[ruoho ruotsi]

I referenced the tf.contrib.rnn.python.ops.core_rnn_cell_impl but you're right, I wasn't clear that the BasicRNNCell and linear both live in this file. github.com/tensorflow/tensorflow/blob/master/tensorflow/cont‌​rib/… Good catch. I'll update my Answer.

[Escachator]

Not in version 1.2

[dennlinger]

This was valid for versions 1.0 and 1.1 only. They (at that time) said they would move the functions to other places in a later release, which is probably what happened in 1.2

[Thava]

For tensorflow version 1.15 this is what worked for me!