Skip to content

Instantly share code, notes, and snippets.

@danijar danijar/
Last active Apr 17, 2019

What would you like to do?
Gated Recurrent Unit with Layer norm and Xavier initializer
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
import tensorflow as tf
class GRU(tf.contrib.rnn.RNNCell):
def __init__(
self, size, activation=tf.tanh, reuse=None,
super(GRU, self).__init__(_reuse=reuse)
self._size = size
self._activation = activation
self._normalizer = normalizer
self._initializer = initializer
def state_size(self):
return self._size
def output_size(self):
return self._size
def call(self, input_, state):
update, reset = tf.split(self._forward(
'update_reset', [state, input_], 2 * self._size, tf.nn.sigmoid,
bias_initializer=tf.constant_initializer(-1.)), 2, 1)
candidate = self._forward(
'candidate', [reset * state, input_], self._size, self._activation)
state = (1 - update) * state + update * candidate
return state, state
def _forward(self, name, inputs, size, activation, **kwargs):
with tf.variable_scope(name):
return _forward(
inputs, size, activation, normalizer=self._normalizer,
weight_initializer=self._initializer, **kwargs)
def _forward(
inputs, size, activation, normalizer=tf.contrib.layers.layer_norm,
if not isinstance(inputs, (tuple, list)):
inputs = (inputs,)
shapes = []
outputs = []
# Map each input to individually normalize their outputs.
for index, input_ in enumerate(inputs):
shapes.append(input_.shape[1: -1].as_list())
input_ = tf.contrib.layers.flatten(input_)
weight = tf.get_variable(
'weight_{}'.format(index + 1), (int(input_.shape[1]), size),
tf.float32, weight_initializer)
output = tf.matmul(input_, weight)
if normalizer:
output = normalizer(output)
output = tf.reduce_mean(outputs, 0)
# Add bias after normalization.
bias = tf.get_variable(
'weight', (size,), tf.float32, bias_initializer)
output += bias
# Activation function.
if activation:
output = activation(output)
# Restore shape dimensions that are consistent among inputs.
min_dim = min(len(shape[1:]) for shape in shapes)
dim_shapes = [[shape[dim] for shape in shapes] for dim in range(min_dim)]
matching_dims = ''.join('NY'[len(set(x)) == 1] for x in dim_shapes) + 'N'
agreement = matching_dims.index('N')
remaining = sum([agreement:]) for shape in shapes)
if agreement:
batch_size = output.shape[0].value or -1
shape = [batch_size] + shapes[:agreement] + [remaining]
output = tf.reshape(output, shape)
return output

This comment has been minimized.

Copy link

commented Sep 22, 2017

Nice work on this one.


This comment has been minimized.

Copy link

commented Nov 7, 2017

I am trying to use this GRU cell and i am getting below error:

File "", line 94, in _forward
output = tf.reshape(output, shape)
File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/ops/", line 3938, in reshape
"Reshape", tensor=tensor, shape=shape, name=name)
File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/", line 513, in _apply_op_helper
raise err
File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/", line 510, in _apply_op_helper
File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/", line 926, in internal_convert_to_tensor
ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)
File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/", line 229, in _constant_tensor_conversion_function
return constant(v, dtype=dtype, name=name)
File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/", line 208, in constant
value, dtype=dtype, shape=shape, verify_shape=verify_shape))
File "/usr/local/lib/python3.5/dist-packages/tensorflow/python/framework/", line 472, in make_tensor_proto
"supported type." % (type(values), values))
TypeError: Failed to convert object of type <class 'list'> to Tensor. Contents: [None, -1]. Consider casting elements to a supported type.

Variables before error are these:
dim = 0
shapes = [[], []]
output = Tensor("rnn/while/rnn/multi_rnn_cell/cell_0/cell_0/gru/update_reset/Sigmoid:0", shape=(?, 400), dtype=float32)
shape = [None, -1]

i am trying to test this using the sequence classification code from here, which has hidden layer length of 200.


This comment has been minimized.

Copy link
Owner Author

commented Jan 8, 2018

Thanks for reporting and sorry for finding this comment so late. The code used to only work with static batch size. I just updated it to also work when the batch size is unknown.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
You can’t perform that action at this time.