Errors when used in bidirectional_dynamic_rnn about indrnn

[batzner]

Hi @makai281, thanks for reaching out. I tried to reproduce your code, but it didn't raise an error. Could you provide a minimal code example that raises an error?

Here is my code:

import tensorflow as tf

from ind_rnn_cell import IndRNNCell

sess = tf.Session()

time_steps = 100
hidden_size = 10
batch_size = 32
inputs = tf.random_uniform([batch_size, time_steps, 20])

recurrent_max = pow(2, 1.0 / time_steps)
fw_rnn_cell = IndRNNCell(hidden_size, recurrent_max_abs=recurrent_max)
bw_rnn_cell = IndRNNCell(hidden_size, recurrent_max_abs=recurrent_max)
bi_states, _ = tf.nn.bidirectional_dynamic_rnn(
    fw_rnn_cell,
    bw_rnn_cell,
    inputs,
    dtype=tf.float32
)

sess.run(tf.global_variables_initializer())
output_fw, output_bw = sess.run(bi_states)
print(output_fw.shape, output_bw.shape)
# prints (32, 100, 10) (32, 100, 10)

Thanks!

[makai281]

sorry for late reply. I seems there are some sub-standard usages in my code, but when I moved the build function to call , the error disappeared. So strange.