GATA-GTF rl training error
subhajitchaudhury opened this issue · 3 comments
Hello, During GATA-GTF rl training with train_rl_with_ground_truth_discrete_belief.py, when trying to process line 1282, I am getting an error on the np.stack due to prev_graph_hidden_state being a list of list of triplets with variable sizes. Any ideas what might be going wrong? Thanks.
For eg: [len(x) for x in prev_graph_hidden_state] gives [39, 49, 49, 49, 35, 49, 48, 45, 39, 39, 45, 49, 42, 41, 43, 51, 51, 43, 46, 45, 48, 45, 41, 48, 45, 45, 40, 39, 41, 48, 45, 41, 39, 49, 49, 42, 42, 46, 35, 52, 41, 41, 46, 49, 45, 48, 48, 44, 43, 40, 40, 42, 42, 49, 42, 40, 42, 40, 40, 51, 41] just before the np.stack command.
Hi @subhajitchaudhury , thanks for asking. This seems to be an overlook when we extract this repo from another development repo. A quick fix is to change THIS LINE to be False. The same thing happens in GATA-GTP config as well. Thanks for reporting and we will fix this.
I have merged a PR that should have fixed this, please let me know if anything else you find weird. Thanks!
Thanks @xingdi-eric-yuan for the prompt action. The PR solved the issue and the rl training is now running. I am closing this issue.