TensorFlow: loss jumps up after restoring RNN net

Question

Environment info

• Operating System: Windows 7 64-bit
• Tensorflow installed from pre-built pip (no CUDA): 1.0.1
• Python 3.5.2 64-bit

Answer 1

The problem is solved. It had nothing to do with RNN nor TensorFlow.

I changed

chars = list(set(data))

to

chars = sorted(set(data))

and now it works.

This is because python uses a random hash function to build the set, and every time python restarted, 'chars' had a different ordering.