Closed CSSE-NULL closed 4 years ago
hmm. it looks like the decoder_inputs
is empty. when does this error happen? does it happen deterministically?
@hhexiy yes, it happened deterministically. It happended when I had trained my manager and imporved it with RL, and I had trained with both modular approach and end-to-end approach. the only strange thing at that period was that I couldn't use RL on the end-to-end one with parameters --agents hybrid hybrid
as suggested, because when I use the suggested parameter, I got the error as follows:
Traceback (most recent call last):
File "reinforce.py", line 60, in <module>
model = system.env.model
AttributeError: 'HybridSystem' object has no attribute 'env'
I had replaced that parameter with --agents pt-nerual pt-nerual
and that works, I don't if that matters. If this is the point, could you tell me how to fix it?
Hi sorry that was a doc error... I've updated it. Training RL agents should always specify --agents pt-neural pt-neural
since it actually refers to the policy model. Otherwise we cannot backprop through the policy.
When I try to chat with the bot in the web interface, I got the error below:
I don't know the reason for it, and my command is:
Other informations:
Please let me know if there is any solutions for this problem, thanks.