Unable to produce the result for dmlab30 language environment

Hi, I tried to run the shallow IMPA,LA model on dmlab30 language model, but besides _language_select_described_object_, both _language_select_located_object_ and _language_execute_random_task_ are stuck in the 0 reward. I'm not sure if there's any hyper parameters needed to tune or more frames are needed to train.


I only changes the larger model to the shallow model in the source code. Number of actors is 48, batch size is 32, all other parameters follow the default value.