'data.parquet' contains detailed training/evaluation scores for all the runs in the experiment section.
To load the file, make sure pandas is installed and run:

> import pandas as pd
> df = pd.read_parquet('./data.parquet')

Keys:

'env': environment
'torso_multiplier': multiplier of image encoder (m in the paper)
'seq_type': mlp (frame-stack) or lstm
'burn_in': length of the burn-in sequence (for recurrent agents)
'z_dim': size of the latent space |Z|
'lucky': whether to include \hat{B} in the DAE objective
'n_step': backup horizon
'store_rnn_state': whether to store the RNN state in the replay buffer (used in the confounding experiment)
'recurrent_actor': recurrent sampling for the behavior policy (used in the confounding experiment)
'score_eval': evaluation scores
'score_train': training scores
'entropy_prior': entropy of the prior
'loss_reward': reconstruction loss for the reward
