Tensorflow LSTM Gate weights

Question

Hello I have a question about Tensorflow. I have some LSTM models trained and I can access the weights and biases of the synaptic connections however I can't seem to access the input, new input, output and forget gate weights of the LSTM cell. I can get the gate tensors out but when I try to .eval() them in a Session I get errors. I'm using the class BasicLSTMCell found in tensorflow/python/ops/rnn_cell.py for my network

`

class BasicLSTMCell(RNNCell):
  """Basic LSTM recurrent network cell.

  The implementation is based on: http://arxiv.org/abs/1409.2329.

  We add forget_bias (default: 1) to the biases of the forget gate in order to
  reduce the scale of forgetting in the beginning of the training.

  It does not allow cell clipping, a projection layer, and does not
  use peep-hole connections: it is the basic baseline.

  For advanced models, please use the full LSTMCell that follows.
  """

  def __init__(self, num_units, forget_bias=1.0, input_size=None,
               state_is_tuple=True, activation=tanh):
    """Initialize the basic LSTM cell.

    Args:
      num_units: int, The number of units in the LSTM cell.
      forget_bias: float, The bias added to forget gates (see above).
      input_size: Deprecated and unused.
      state_is_tuple: If True, accepted and returned states are 2-tuples of
        the `c_state` and `m_state`.  If False, they are concatenated
        along the column axis.  The latter behavior will soon be deprecated.
      activation: Activation function of the inner states.
    """
    if not state_is_tuple:
      logging.warn("%s: Using a concatenated state is slower and will soon be "
                   "deprecated.  Use state_is_tuple=True.", self)
    if input_size is not None:
      logging.warn("%s: The input_size parameter is deprecated.", self)
    self._num_units = num_units
    self._forget_bias = forget_bias
    self._state_is_tuple = state_is_tuple
    self._activation = activation

  @property
  def state_size(self):
    return (LSTMStateTuple(self._num_units, self._num_units)
            if self._state_is_tuple else 2 * self._num_units)

  @property
  def output_size(self):
    return self._num_units

  def __call__(self, inputs, state, scope=None):
    """Long short-term memory cell (LSTM)."""
    with vs.variable_scope(scope or type(self).__name__):  # "BasicLSTMCell"
      # Parameters of gates are concatenated into one multiply for efficiency.
      if self._state_is_tuple:
        c, h = state
      else:
        c, h = array_ops.split(1, 2, state)
      concat = _linear([inputs, h], 4 * self._num_units, True)

      # i = input_gate, j = new_input, f = forget_gate, o = output_gate
      i, j, f, o = array_ops.split(1, 4, concat)

      new_c = (c * sigmoid(f + self._forget_bias) + sigmoid(i) *
               self._activation(j))
      new_h = self._activation(new_c) * sigmoid(o)

      if self._state_is_tuple:
        new_state = LSTMStateTuple(new_c, new_h)
      else:
        new_state = array_ops.concat(1, [new_c, new_h])
      return new_h, new_state


def _get_concat_variable(name, shape, dtype, num_shards):
  """Get a sharded variable concatenated into one tensor."""
  sharded_variable = _get_sharded_variable(name, shape, dtype, num_shards)
  if len(sharded_variable) == 1:
    return sharded_variable[0]

  concat_name = name + "/concat"
  concat_full_name = vs.get_variable_scope().name + "/" + concat_name + ":0"
  for value in ops.get_collection(ops.GraphKeys.CONCATENATED_VARIABLES):
    if value.name == concat_full_name:
      return value

  concat_variable = array_ops.concat(0, sharded_variable, name=concat_name)
  ops.add_to_collection(ops.GraphKeys.CONCATENATED_VARIABLES,
                        concat_variable)
  return concat_variable


def _get_sharded_variable(name, shape, dtype, num_shards):
  """Get a list of sharded variables with the given dtype."""
  if num_shards > shape[0]:
    raise ValueError("Too many shards: shape=%s, num_shards=%d" %
                     (shape, num_shards))
  unit_shard_size = int(math.floor(shape[0] / num_shards))
  remaining_rows = shape[0] - unit_shard_size * num_shards

  shards = []
  for i in range(num_shards):
    current_size = unit_shard_size
    if i < remaining_rows:
      current_size += 1
    shards.append(vs.get_variable(name + "_%d" % i, [current_size] + shape[1:],
                                  dtype=dtype))
  return shards

`

I can see the i, j, f, o gates being used in the def call however when I tf.print them I get tensors out, and when I try to .eval() them in a Session I get errors. I also tried tf.getVariable but was not able to extract the weight matrices. My question: is there a way to evaluate the i, j ,f and o gate weights/matrices?

Answer 1

First, to clear some confusion: i, j, f and o tensors are not weight matrices; they are intermediate calculation steps that depend on particular LSTM cell input. All the weights of the LSTM cell are stored in variables self._kernel and self._bias, and in a constant self._forget_bias.

So, to answer both possible interpretations of your question, I'll show how print the values of the self._kernel and self._bias, and the values of i, j, f and o tensors at every step.

Suppose we have the following graph:

import numpy as np
import tensorflow as tf

timesteps = 7
num_input = 4
num_units = 3
x_val = np.random.normal(size=(1, timesteps, num_input))

lstm = tf.nn.rnn_cell.BasicLSTMCell(num_units = num_units)
X = tf.placeholder("float", [1, timesteps, num_input])
inputs = tf.unstack(X, timesteps, 1)
outputs, state = tf.contrib.rnn.static_rnn(lstm, inputs, dtype=tf.float32)

We can find the value of any tensor if we know its name. One way to find a tensor's name is to look at TensorBoard.

init = tf.global_variables_initializer()
graph = tf.get_default_graph()
with tf.Session(graph=graph) as sess:
    train_writer = tf.summary.FileWriter('./graph', sess.graph)
    sess.run(init)

Now we can start TensorBoard by the terminal command

tensorboard --logdir=graph --host=localhost 

and find that the operation which produces i, j, f, o tensors has name 'rnn/basic_lstm_cell/split', while kernel and bias are called 'rnn/basic_lstm_cell/kernel' and 'rnn/basic_lstm_cell/bias':

The tf.contrib.rnn.static_rnn function calls our basic lstm cell 7 times, once for every timestep. When Tensorflow is asked to create several operations under the same name, it adds suffixes to them, like this: rnn/basic_lstm_cell/split, rnn/basic_lstm_cell/split_1, ..., rnn/basic_lstm_cell/split_6. These are the names of our operations.

The name of a tensor in tensorflow consists of the name of the operation that produces the tensor, followed by a colon, followed by the index of the operation's output that produces this tensor. Kernel and bias ops have a single output, so the tensor names will be

kernel = graph.get_tensor_by_name("rnn/basic_lstm_cell/kernel:0")
bias = graph.get_tensor_by_name("rnn/basic_lstm_cell/bias:0")

The split operation produces four outputs: i, j, f and o, so these tensors' names will be:

i_list = []
j_list = []
f_list = []
o_list = []
for suffix in ["", "_1", "_2", "_3", "_4", "_5", "_6"]:   
    i_list.append(graph.get_tensor_by_name(
        "rnn/basic_lstm_cell/split{}:0".format(suffix)
    ))
    j_list.append(graph.get_tensor_by_name(
        "rnn/basic_lstm_cell/split{}:1".format(suffix)
    ))
    f_list.append(graph.get_tensor_by_name(
        "rnn/basic_lstm_cell/split{}:2".format(suffix)
    ))        
    o_list.append(graph.get_tensor_by_name(
        "rnn/basic_lstm_cell/split{}:3".format(suffix)
    ))

and now we can find the values of all tensors:

    with tf.Session(graph=graph) as sess:
        train_writer = tf.summary.FileWriter('./graph', sess.graph)
        sess.run(init)
        weights = sess.run([kernel, bias])
        print("Weights:\n", weights)
        i_values, j_values, f_values, o_values = sess.run([i_list, j_list, f_list, o_list], 
                                                          feed_dict = {X:x_val})
        print("i values:\n", i_values)
        print("j values:\n", j_values)
        print("f_values:\n", f_values)
        print("o_values:\n", o_values)

Alternatively, we could find the tensor names by looking at the list of all tensors in a graph, which can be produced by:

tensors_per_node = [node.values() for node in graph.get_operations()]
tensor_names = [tensor.name for tensors in tensors_per_node for tensor in tensors]
print(tensor_names)

Or, for a shorter list of all operations:

print([node.name for node in graph.get_operations()])

The third way is to read the source code and find which names are assigned to which tensors.