交互式会话eval和tf.random_uniform无法产生预期的输出

Question

I have a problem with the following code piece, the obs_pattern , obs_seq and obs_seq_s do not produce the expected behavior. I tried with TensorFlow 1.2.1. I suspect there is something wrong.

import tensorflow as tf

sess = tf.InteractiveSession()
seq_length = 5
num_bits = 4
obs_pattern_shape = [num_bits, seq_length]
obs_pattern = tf.cast(
    tf.random_uniform(obs_pattern_shape, minval=0, maxval=2, seed=1234, dtype=tf.int32), 
    tf.float32)
print(obs_pattern.eval())
seq_length_zeros = tf.zeros([1, seq_length])
obs_seq = tf.concat([obs_pattern, seq_length_zeros], axis=0)
print(obs_seq.eval())
add_vec = tf.one_hot([num_bits], (num_bits + 1), on_value = 1.0, off_value=0.0, axis=0)
obs_seq_s = tf.concat([obs_seq, add_vec], axis=1)
print(obs_seq_s.eval())

sess.close()

obs_pattern

[[ 1.  1.  1.  0.  0.]
[ 0.  0.  0.  1.  0.]
[ 0.  1.  1.  0.  0.]
[ 0.  1.  0.  0.  0.]]

obs_seq

[[ 1.  1.  0.  1.  1.]
[ 0.  1.  1.  0.  1.]
[ 1.  1.  1.  0.  0.]
[ 1.  1.  0.  1.  0.]
[ 0.  0.  0.  0.  0.]]

obs_seq_s

[[ 1.  0.  0.  0.  0.  0.]
[ 0.  0.  1.  0.  0.  0.]
[ 0.  1.  0.  0.  0.  0.]
[ 0.  1.  0.  0.  0.  0.]
[ 0.  0.  0.  0.  0.  1.]]

EDIT: Based on the comment below I changed obs_pattern and it behaves as I'd thought

import numpy as np
arr = np.random.randint(2, size=(num_bits, seq_length))
obs_pattern = tf.convert_to_tensor(arr, dtype=tf.float32)

Answer 1

It is expected each time you run sess on a tensor formed with obs_pattern it reruns a new random uniform. To get the behavior you expect generate ONCE a numpy array with numpy then feed it to a placeholder of the same shape : obs_pattern. To clarify you used a seed but you ran the op a number of times so you get the first iterates of the sequence.