twm
twm copied to clipboard
jrobine
Transformer-based World Models
Hi, Thank you for open-sourcing your work! I have a question regarding reproduciblity. I tried running the code on the Atari 100k benchmark and I did not manage to reproduce...
Hello, I have a question regarding the code in the trainer.py file, specifically lines 251-252 in the # pretrain ac section: ``` budget = config['pretrain_budget'] * (1 - config['pretrain_obs_p'] +...
When rolling out policy [here](https://github.com/jrobine/twm/blob/e6a8e599864f5539b166ee9a8d173ece8f3fba84/twm/trainer.py#L78). The nested function `policy(index)` ALWAYS assumes dreamer is None (i.e. never going to the else section).
When compute_dynamics_loss is called in [world_model.py](https://github.com/jrobine/twm/blob/e6a8e599864f5539b166ee9a8d173ece8f3fba84/twm/world_model.py#L69), it is fed in an argument called `preds` returned from dyn_model.predict(). For `z` prediction, preds have keys `z_dist` and `z_hat_probs` and there is no...
