I recently tried running the dreamer implementation in torch rl. But the reconstruction loss (in loss_model_reco.csv), kl divergence loss (in loss_model_kl.csv), the world model loss (in loss_world_model.csv), reward loss (in loss_model_reward.csv), and the world model gradients (in grad_world_model) are all nans. Is this normal? It would be great if there were plots available for the torch rl implementation to compare with. @vmoens
Looking into this thanks for reporting!
Can you open an issue on the repo?
Sure, will open an issue. Thank you for looking into this.