Trouble sending np.zeros through to tf.placeholder

Question

I've been learning tensorflow recently and I'm having trouble sending the appropriately sized and typed numerical data through to the placeholder in tensorflow. This code has been adapted from https://github.com/SullyChen/Nvidia-Autopilot-TensorFlow.

My model file:

import tensorflow as tf
import scipy
from tensorflow.python.ops import rnn, rnn_cell


def weight_variable(shape):
    initial = tf.truncated_normal(shape, stddev=0.1)
    return tf.Variable(initial)

def bias_variable(shape):
    initial = tf.constant(0.1, shape=shape)
    return tf.Variable(initial)

def conv2d(x, W, stride):
    return tf.nn.conv2d(x, W, strides=[1, stride, stride, 1], padding='VALID')



class AutopilotRCNModel(object):
    """
    Use Recurrent Neural Networks in combination with Convolutional Neural
    Networks to predict steering wheel angle.
    """
    def __init__(self, n_hidden=50):

        # Create a state variable to be passed in through feed_dict
        self.c_state = tf.placeholder(tf.float32, shape=[1, n_hidden])
        self.h_state = tf.placeholder(tf.float32, shape=[1, n_hidden])

        # Do the same with the input x and the target variable y_
        self.x = tf.placeholder(tf.float32, shape=[None, 66, 200, 3])
        self.y_ = tf.placeholder(tf.float32, shape=[None, 1])

        x_image = self.x

        # First convolutional layer.
        W_conv1 = weight_variable([5, 5, 3, 24])
        b_conv1 = bias_variable([24])

        # Stride is 2.
        h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1, 2) + b_conv1)

        # Second convolutional layer.
        W_conv2 = weight_variable([5, 5, 24, 36])
        b_conv2 = bias_variable([36])

        # Stride still set to 2.
        h_conv2 = tf.nn.relu(conv2d(h_conv1, W_conv2, 2) + b_conv2)

        # Third convolutional layer.
        W_conv3 = weight_variable([5, 5, 36, 48])
        b_conv3 = bias_variable([48])

        # Stride still 2.
        h_conv3 = tf.nn.relu(conv2d(h_conv2, W_conv3, 2) + b_conv3)

        # Fourth convolutional layer.
        W_conv4 = weight_variable([3, 3, 48, 64])
        b_conv4 = bias_variable([64])

        # Stride now set to 1.
        h_conv4 = tf.nn.relu(conv2d(h_conv3, W_conv4, 1) + b_conv4)

        # Fifth convolutional layer.
        W_conv5 = weight_variable([3, 3, 64, 64])
        b_conv5 = bias_variable([64])

        # Stride of 1.
        h_conv5 = tf.nn.relu(conv2d(h_conv4, W_conv5, 1) + b_conv5)

        # Fully connected layer 1.
        W_fc1 = weight_variable([1152, 1164])
        b_fc1 = bias_variable([1164])

        # Requires flattening out the activations.
        h_conv5_flat = tf.reshape(h_conv5, [-1, 1152])
        h_fc1 = tf.nn.relu(tf.matmul(h_conv5_flat, W_fc1) + b_fc1)

        self.keep_prob = tf.placeholder(tf.float32)
        h_fc1_drop = tf.nn.dropout(h_fc1, self.keep_prob)

        # Fully connected layer 2.
        W_fc2 = weight_variable([1164, 100])
        b_fc2 = bias_variable([100])

        h_fc2 = tf.nn.relu(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)

        h_fc2_drop = tf.nn.dropout(h_fc2, self.keep_prob)

        print(h_fc2_drop.get_shape())


        # LSTM layer 1

        # Create weight matrix and bias to map from output of LSTM
        # to steering wheel angle in radians.

        # Input gate input weights, recurrent weights, and bias.
        W_i = weight_variable([100, n_hidden])
        U_i = weight_variable([n_hidden, n_hidden])
        b_i = bias_variable([n_hidden])

        # Forget gate input weights, recurrent weights, and bias.
        W_f = weight_variable([100,n_hidden])
        U_f = weight_variable([n_hidden, n_hidden])
        b_f = bias_variable([n_hidden])

        # Candidate gate input weights, recurrent weights, and bias.
        W_c = weight_variable([100, n_hidden])
        U_c = weight_variable([n_hidden,n_hidden])
        b_c = bias_variable([n_hidden])

        # Output gate input weights, recurrent weights and bias.
        W_o = weight_variable([100, n_hidden])
        U_o = weight_variable([n_hidden, n_hidden])
        b_o = bias_variable([1])

        V_o = weight_variable([n_hidden, n_hidden])

        ingate = tf.nn.sigmoid(tf.matmul(h_fc2_drop, W_i) + tf.matmul(self.h_state, U_i) + b_i)

        cgate = tf.nn.tanh(tf.matmul(h_fc2_drop, W_c) + tf.matmul(self.h_state, U_c) + b_c)

        fgate = tf.nn.sigmoid(tf.matmul(h_fc2_drop, W_f) + tf.matmul(self.h_state, U_f) + b_f)

        self.c_state = tf.mul(ingate, cgate) + tf.mul(fgate, self.c_state)

        h_rnn1 = tf.nn.sigmoid(tf.matmul(h_fc2_drop, W_o) + \
        tf.matmul(self.h_state, U_o) + tf.matmul(self.c_state, V_o) + b_o)

        self.h_state = tf.mul(h_rnn1, tf.nn.tanh(self.c_state))

        W_out = weight_variable([n_hidden,1])
        b_out = bias_variable([1])

        self.y = tf.mul(tf.atan(tf.matmul(h_rnn1, W_out) + b_out), 2)

        self.loss = tf.square(tf.sub(self.y, self.y_))

First tester code is here:

import model
import driving_data

import tensorflow as tf
import numpy as np

n_hidden = 65

rcn = model.AutopilotRCNModel(n_hidden=n_hidden)

xs, ys = driving_data.LoadTrainBatch(1)

c_state = np.zeros((1,n_hidden),dtype=np.float32)
h_state = np.zeros((1,n_hidden),dtype=np.float32)

sess = tf.InteractiveSession()

train_step = tf.train.AdamOptimizer().minimize(rcn.loss)

sess.run(tf.initialize_all_variables())

train_step.run(
    feed_dict = {
        rcn.x: xs,
        rcn.y_: ys,
        rcn.keep_prob: 0.8,
        rcn.c_state: c_state,
        rcn.h_state: h_state
        }
)

and I'm getting this error:

Traceback (most recent call last):
  File "test.py", line 28, in <module>
    rcn.h_state: h_state
  File "/home/thomas/.local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 1621, in run
    _run_using_default_session(self, feed_dict, self.graph, session)
  File "/home/thomas/.local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 3804, in _run_using_default_session
    session.run(operation, feed_dict)
  File "/home/thomas/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 717, in run
    run_metadata_ptr)
  File "/home/thomas/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 915, in _run
    feed_dict_string, options, run_metadata)
  File "/home/thomas/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 965, in _do_run
    target_list, options, run_metadata)
  File "/home/thomas/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 985, in _do_call
    raise type(e)(node_def, op, message)
tensorflow.python.framework.errors.InvalidArgumentError: You must feed a value for placeholder tensor 'Placeholder' with dtype float and shape [1,65]
     [[Node: Placeholder = Placeholder[dtype=DT_FLOAT, shape=[1,65], _device="/job:localhost/replica:0/task:0/gpu:0"]()]]
     [[Node: _recv_add_13_0 = _Recv[client_terminated=true, recv_device="/job:localhost/replica:0/task:0/cpu:0", send_device="/job:localhost/replica:0/task:0/cpu:0", send_device_incarnation=7747556615882019196, tensor_name="add_13:0", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/cpu:0"]()]]

Caused by op u'Placeholder', defined at:
  File "test.py", line 9, in <module>
    rcn = model.AutopilotRCNModel(n_hidden=n_hidden)
  File "/home/thomas/projects/lstm_autopilot/model.py", line 27, in __init__
    self.c_state = tf.placeholder(tf.float32, shape=[1, n_hidden])
  File "/home/thomas/.local/lib/python2.7/site-packages/tensorflow/python/ops/array_ops.py", line 1330, in placeholder
    name=name)
  File "/home/thomas/.local/lib/python2.7/site-packages/tensorflow/python/ops/gen_array_ops.py", line 1762, in _placeholder
    name=name)
  File "/home/thomas/.local/lib/python2.7/site-packages/tensorflow/python/framework/op_def_library.py", line 748, in apply_op
    op_def=op_def)
  File "/home/thomas/.local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 2388, in create_op
    original_op=self._default_original_op, op_def=op_def)
  File "/home/thomas/.local/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 1300, in __init__
    self._traceback = _extract_stack()

InvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'Placeholder' with dtype float and shape [1,65]
     [[Node: Placeholder = Placeholder[dtype=DT_FLOAT, shape=[1,65], _device="/job:localhost/replica:0/task:0/gpu:0"]()]]
     [[Node: _recv_add_13_0 = _Recv[client_terminated=true, recv_device="/job:localhost/replica:0/task:0/cpu:0", send_device="/job:localhost/replica:0/task:0/cpu:0", send_device_incarnation=7747556615882019196, tensor_name="add_13:0", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/cpu:0"]()]]

My actual numeric feed_dict data is the same shape and type as the placeholder requires. Not really sure what's wrong here. I would think np.zeros((1,n_hidden), dtype=np.float32) should feed the tf.placeholder(tf.float32,size=[1,n_hidden]) quite well.

Answer 1

My problem was that I was updating the placeholder inside the class, a definite no-no. I just changed

self.c_state = tf.mul(ingate, cgate) + tf.mul(fgate, self.c_state)

to

self.new_c_state = tf.mul(ingate, cgate) + tf.mul(fgate, self.c_state)

and

self.h_state = tf.mul(h_rnn1, tf.nn.tanh(self.c_state))

to

self.new_h_state = tf.mul(h_rnn1, tf.nn.tanh(self.new_c_state))

!!! It all makes sense. Placeholders are placeholders for data. Don't do update operations on them or things get all screwy.