# Run tf_gbds
$ python run_model.py --model_dir='new_model' (Directory where the model is saved) \
--train_data_dir='/directory/of/training/data' (Directory of training dataset file) \
--val_data_dir='/directory/of/validation/data' (Directory of validation dataset file) \
--synthetic_data=False (Is the model trained on synthetic data?) \
--save_posterior=True (Will posterior samples be retrieved after training?) \
--load_saved_model=False (Is the model restored from an existing checkpoint?)
--saved_model_dir='/directory/you/save/checkpoint' (Directory where the model to be restored is saved) \

--game_name="penaltykick" (Name of the game)
--n_agents=2 (Number of agents in the model)
--agent_name="goalie,shooter" (Name of each agent (separated by ,))
--agent_col="0;1,2" (Columns of dataset corresponding to each agent (separated by ; and ,))
--obs_dim=3 (Dimension of observation)
--extra_conds=Flase (Are extra conditions included in the dataset?)
--extra_dim=0 (Dimension of extra conditions)
--ctrl_obs=False (Are observed control signals included in the dataset?)
--add_accel=False (Is acceleration included in game state?")

--GMM_K=8 (Number of components in GMM) \
--gen_n_layers=3 (Number of layers in neural networks (generative model)) \
--gen_hidden_dim=64 (Number of hidden units in each dense layer of neural networks (generative model)) \
--rec_lag=10 (Number of previous timepoints included as input to recognition model) \
--rec_n_layers=3 (Number of layers in neural networks (recognition model)) \
--rec_hidden_dim=32 (Number of hidden units in each dense layer of neural networks (recognition model)) \

--sigma_init=-7. (Initial value of goal state variance) \
--sigma_trainable=False (Is sigma trainable?) \
--sigma_pen=1e3 (Penalty on large sigma) \
--g_lb=-1. (Goal state lower boundary) \
--g_ub=1. (Goal state upper boundary) \
--g_bounds_pen=None (Penalty for goal states escaping boundaries) \

--eps_init=-11. (Initial value of control signal variance) \
--eps_trainable=False (Is epsilon trainable?) \
--eps_pen=1e5 (Penalty on large epsilon) \
--latent_u=False (Is the true control signal modeled as latent variable?)
--clip=False (Is the observed control signal censored?) \
--clip_lb=-1. (Control signal censoring lower bound) \
--clip_ub=1. (Control signal censoring upper bound) \
--clip_tol=1e-5 (Tolerance of signal censoring) \
--clip_pen=1e8 (Penalty on control signal censoring)

--seed=1234 (Random seed) \
--opt='Adam' (Gradient descent optimizer) \
--lr=1e-3 (Initial learning rate) \
--n_epochs=500 (Number of iterations algorithm runs through the training set) \
--B=1 (Size of mini-batches) \
--n_samp=1 (Number of samples drawn for gradient estimation) \
--n_post_samp=30 (Number of samples from posterior distributions to draw and save) \
--max_ckpt=10 (Maximum number of checkpoints to keep in the directory) \
--freq_ckpt=5 (Frequency of saving checkpoints to the directory)
--freq_val_loss=1 (Frequency of computing validation set loss)