Preserving tensor values between session runs

Question

Consider the following example:

import tensorflow as tf
import math
import numpy as np

INPUTS = 10
HIDDEN_1 = 20
BATCH_SIZE = 3


def create_graph(inputs):
    with tf.name_scope('h1'):
        weights = tf.Variable(
        tf.truncated_normal([INPUTS, HIDDEN_1],
                            stddev=1.0 / math.sqrt(float(INPUTS))),
        name='weights')
        biases = tf.Variable(tf.zeros([HIDDEN_1]),
                         name='biases')
        state = tf.Variable(tf.zeros([HIDDEN_1]), name='inner_state')
        state = tf.Print(state, [state], message=" this is state before: ")
        state = 0.9*state + 0.1*( tf.matmul(inputs, weights) + biases )
        state = tf.Print(state, [state], message=" this is state after: ")
        output = tf.nn.relu(state)
    return output

def data_iter():
    while True:
        idxs = np.random.rand(BATCH_SIZE, INPUTS)
        yield idxs

with tf.Graph().as_default():
    inputs = tf.placeholder(tf.float32, shape=(BATCH_SIZE, INPUTS))
    output = create_graph(inputs)

    sess = tf.Session()
    # Run the Op to initialize the variables.
    init = tf.initialize_all_variables()
    sess.run(init)
    iter_ = data_iter()
    for i in xrange(0, 2):
        print ("iteration: ",i)
        input_data = iter_.next()
        out = sess.run(output, feed_dict={ inputs: input_data})

I was hoping to have the tensor state to preserve its intermediate and to change slowly on each iteration. However, What I am seeing is that on each sess.run invocation, the state begins on zero values:

('iteration: ', 0)
I tensorflow/core/kernels/logging_ops.cc:79]  this is state before: [0 0 0...]
I tensorflow/core/kernels/logging_ops.cc:79]  this is state after: [0.007762237 0.044753391 0.049343754...]
('iteration: ', 1)
I tensorflow/core/kernels/logging_ops.cc:79]  this is state before: [0 0 0...]
I tensorflow/core/kernels/logging_ops.cc:79]  this is state after: [0.040079735 0.074709542 0.078258425...]

I would appreciate any clarification how to address this behavior

Edit

after commenting the tf.Print lines and replacing the incorrect assignment with

state = state.assign(0.9*state + 0.1*( tf.matmul(inputs, weights) + biases ))

I get these errors:

Traceback (most recent call last):
  File "cycles_in_graphs.py", line 33, in <module>
    output = create_graph(inputs)
  File "cycles_in_graphs.py", line 21, in create_graph
    state = state.assign(0.9*state + 0.1*( tf.matmul(inputs, weights) + biases ))
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/variables.py", line 453, in assign
    return state_ops.assign(self._variable, value, use_locking=use_locking)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/gen_state_ops.py", line 40, in assign
    use_locking=use_locking, name=name)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/op_def_library.py", line 655, in apply_op
    op_def=op_def)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 2156, in create_op
    set_shapes_for_outputs(ret)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 1612, in set_shapes_for_outputs
    shapes = shape_func(op)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/state_ops.py", line 197, in _AssignShape
    return [op.inputs[0].get_shape().merge_with(op.inputs[1].get_shape())]
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/tensor_shape.py", line 554, in merge_with
    (self, other))
ValueError: Shapes (20,) and (3, 20) are not compatible

score 1 · Answer 1 · answered Jun 01 '16 at 13:32

1

When you write state = 0.9 * state + 0.1 * (tf.matmul(inputs, weights) + biases), you do not update the value of the variable state.
You only compute the value of 0.9 * state + 0.1 * ..., but the value of the Variable stays the same.

To update your tf.Variable, you should use the function assign or assign_add on your variable state:

state = state.assign(0.9 * state + 0.1 * (tf.matmul(inputs, weights) + biases))

Everything is explained in the TensorFlow tutorial on Variables.

answered Jun 01 '16 at 13:32

Olivier Moindrot

27,908
11
92
91

thanks. I get this error " state = state.assign(0.9*state + 0.1*( tf.matmul(inputs, weights) + biases )) AttributeError: 'Tensor' object has no attribute 'assign'" – diffeomorphism Jun 01 '16 at 19:40
Yes, remove the line `state = tf.Print(state)` just before. Only `tf.Variable` have attribute `assign()` – Olivier Moindrot Jun 01 '16 at 19:57
thanks! I updated the assignment expression, but I get some incompatible shape errors, I updated the OP with the errors – diffeomorphism Jun 01 '16 at 20:07
The error is quite clear: state has shape `[20]`, while `tf.matmul(inputs, weights) + biases` has shape `[3, 20]`. – Olivier Moindrot Jun 01 '16 at 22:34

Preserving tensor values between session runs

1 Answers1

Linked