Myujth

I'd like to enter the tensor to the intial_state of an LSTM as a placeholder so that later on I assign it's value myself during the learning process.

I wrote the following:

import tensorflow as tf
class PGNetwork:
def __init__(self, inpsize, name='PGNetwork'):
 tf.reset_default_graph()
 self.inpsize = inpsize

 with tf.variable_scope(name):
 # We create the placeholders
 self.inputs_vec = tf.placeholder(tf.float32, [None, 
 self.inpsize], name="inputs_vec")
 self.in_state = tf.placeholder(tf.float32, [1, 16], name="in_state")

 self.lstm_layer = tf.contrib.rnn.BasicLSTMCell(16,forget_bias=1)
# self.in_state = self.lstm_layer.zero_state(1, dtype=tf.float32) # By uncommenting this line the error is no longer there.
 self.out_rnn, self.rnn_state = tf.nn.dynamic_rnn(self.lstm_layer, 
 tf.expand_dims(self.inputs_vec, 1), initial_state=self.in_state)

 self.output = tf.layers.dense(inputs = self.out_rnn, 
 kernel_initializer=tf.contrib.layers.xavier_initializer(),
 units = 5, 
 activation=None, name="output")

 self.action_distribution = tf.nn.softmax(self.output, name="softmax")

PGNetwork = PGNetwork(8)

I get this error :
Tensor objects are not iterable when eager execution is not enabled. To iterate over this tensor use tf.map_fn.

Which is actually there because of the initial_state in tf.nn.dynamic_rnn() since it is not accepted to be a placeholder.

Is it possible to convert the placeholder to something accepted by the dynamic_rnn????