Request for the command for reproducing longhistory square result

[HaoranZhangumich]

The current command I run for reproducing the paper LDP with PTP is

python train.py --config-dir=experiment_configs/longhist --config-name=transformer_longhist_emb hydra.run.dir=data/outputs/longhist_ldp_emb_s20 training.debug=false logging.name=longhist_ldp_emb_s20 subsample_frames=20 global_obs=20 global_horizon=36 policy.use_embed_if_present=true task.dataset.use_embed_if_present=true task.dataset.dataset_path=data/longhistsquare100/demos.hdf5 task.dataset_path=data/longhistsquare100/demos.hdf5 task.env_runner.dataset_path=data/longhistsquare100/demos.hdf

But after 500 epoch the result is really bad although I got a really low error in training and valuation

wandb:                             train/mean_score 0.0
wandb:                       train/sim_max_reward_0 0
wandb:                       train/sim_max_reward_1 0
wandb:                       train/sim_max_reward_2 0
wandb:                       train/sim_max_reward_3 0
wandb:                       train/sim_max_reward_4 0
wandb:                       train/sim_max_reward_5 0
wandb:                       train_action_mse_error 2e-05
wandb:                                   train_loss 0.00282
wandb:                                     val_loss 0.00152