How is it that I am getting different results when I perform batch evaluation of a neural network in tensor flow as opposed to evaluation one by one?

Question

I built a regression convolutional neural network based on this tutorial in the tensorflow website.
When I evaluate the result for multiple images at once, I'm getting different results from when I evaluate each of those images one by one.

To be more concrete, for three identical sample images, I get [ 729027.5625 729027.5625 729027.5625] when I calculate the output for them as a batch, but I get [ 729026.4375] [ 729026.4375] [ 729026.4375] when I calculate the outputs for the images one by one.

Any idea why this is the case? The input to my neural network is defined as below:

x = tf.placeholder(tf.float32, shape=[None, 784])

And when performing batch evaluation, I input a 2 dimensional numpy array of images (shape = (100, 784))

EDIT: Please see the below MWE

mwe_db.npy

from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
import tensorflow as tf
import numpy as np

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

def max_pool_2x2(x):
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')


sess = tf.Session()

x = tf.placeholder(tf.float32, shape=[None, 784])
y_ = tf.placeholder(tf.float32, shape=[None, 10])
x_image = tf.reshape(x, [-1,28,28,1])

W_conv1 = weight_variable([5, 5, 1, 32])
b_conv1 = bias_variable([32])
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)

W_conv2 = weight_variable([5, 5, 32, 64])
b_conv2 = bias_variable([64])
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)

W_fc1 = weight_variable([7 * 7 * 64, 1024])
b_fc1 = bias_variable([1024])
h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

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

W_fc2 = weight_variable([1024, 10])
b_fc2 = bias_variable([10])
y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2

cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_conv, y_))
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
sess.run(tf.initialize_all_variables())
for i in range(20000):
  batch = mnist.train.next_batch(50)
  if i%100 == 0:
    train_accuracy = accuracy.eval(feed_dict={
        x:batch[0], y_: batch[1], keep_prob: 1.0}, session=sess)
    print("step %d, training accuracy %g"%(i, train_accuracy))
  train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5}, session=sess)

print("test accuracy %g"%accuracy.eval(feed_dict={
    x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0}, session=sess))

####################
# THE ANOMALY
####################

print "----------------------"
print "THE ANOMALY"
print "----------------------"

mwe_db = np.load("mwe_db.npy")
num_rows = mwe_db.shape[0]

batch_y = y_conv.eval(feed_dict={x: mwe_db[0:2, :], keep_prob: 1.0}, session=sess)
print batch_y

print "----------------------"

for i in xrange(2):
    single_y = y_conv.eval(feed_dict={x: mwe_db[i:i+1, :], keep_prob: 1.0}, session=sess)
    print single_y

Answer 1

So, I do not have a definite answer, but tf.matmul is to blame. The minimal working example:

import numpy as np
np.random.seed(123)
import tensorflow as tf

tf.set_random_seed(123)

# To enforce determinism
sess = tf.Session(config=tf.ConfigProto(
        inter_op_parallelism_threads=1,
        intra_op_parallelism_threads=1,
        device_count={'GPU': 0}
))

init = np.random.rand(2,50)
x1=tf.Variable(init, dtype=tf.float32)
x2=tf.Variable(init[:1], dtype=tf.float32)

W=tf.Variable(np.random.rand(50,50), dtype=tf.float32)

sess.run(tf.initialize_all_variables())

r1 = tf.matmul(x1, W)
r2 = tf.matmul(x2, W)
v1, v2 = sess.run([r1, r2])
print(v1[0] - v2)