small-rl-binary/config.txt

# Parameters for Actor:
# ==============================================================================
Actor.activation = 'leaky_relu'
Actor.cont_dist_kind = 'normal'
Actor.d_hidden = 300
Actor.dropout_p = 0.0
Actor.gmm_modes = 5
Actor.log_std_high = 2.0
Actor.log_std_low = -5.0
Actor.n_layers = 2

# Parameters for Agent:
# ==============================================================================
Agent.fake_filter = False
Agent.gamma = 0.999
Agent.num_critics = 4
Agent.num_critics_td = 2
Agent.offline_coeff = 1.0
Agent.online_coeff = 0.0
Agent.popart = True
Agent.reward_multiplier = 10.0
Agent.tau = 0.003
Agent.use_multigamma = True
Agent.use_target_actor = True

# Parameters for Experiment:
# ==============================================================================
Experiment.batches_per_update = 1
Experiment.critic_loss_weight = 10.0
Experiment.env_mode = 'async'
Experiment.force_reset_train_envs_every = None
Experiment.grad_clip = 1.0
Experiment.has_replay_buffer_rights = True
Experiment.l2_coeff = 0.001
Experiment.learning_rate = 0.0001
Experiment.local_time_optimizer = False
Experiment.lr_warmup_steps = 500
Experiment.mixed_precision = 'no'
Experiment.padded_sampling = 'none'
Experiment.save_trajs_as = 'npz'
Experiment.stagger_traj_file_lengths = True
Experiment.wandb_group_name = None

# Parameters for FlashAttention:
# ==============================================================================
FlashAttention.window_size = (-1, -1)

# Parameters for MetamonTstepEncoder:
# ==============================================================================
MetamonTstepEncoder.d_model = 100
MetamonTstepEncoder.extra_emb_dim = 18
MetamonTstepEncoder.n_heads = 5
MetamonTstepEncoder.n_layers = 3
MetamonTstepEncoder.scratch_tokens = 4
MetamonTstepEncoder.token_mask_aug = False

# Parameters for Multigammas:
# ==============================================================================
Multigammas.continuous = [0.1, 0.9, 0.95, 0.97, 0.99, 0.995]
Multigammas.discrete = [0.1, 0.9, 0.95, 0.97, 0.99, 0.995]

# Parameters for MultiModalEmbedding:
# ==============================================================================
MultiModalEmbedding.dropout = 0.05
MultiModalEmbedding.numerical_tokens = 6

# Parameters for NCritics:
# ==============================================================================
NCritics.activation = 'leaky_relu'
NCritics.d_hidden = 300
NCritics.dropout_p = 0.0
NCritics.n_layers = 2

# Parameters for PopArtLayer:
# ==============================================================================
PopArtLayer.beta = 0.0005
PopArtLayer.init_nu = 100.0

# Parameters for TformerTrajEncoder:
# ==============================================================================
TformerTrajEncoder.activation = 'leaky_relu'
TformerTrajEncoder.causal = True
TformerTrajEncoder.d_ff = 2048
TformerTrajEncoder.d_model = 512
TformerTrajEncoder.dropout_attn = 0.0
TformerTrajEncoder.dropout_emb = 0.05
TformerTrajEncoder.dropout_ff = 0.05
TformerTrajEncoder.dropout_qkv = 0.0
TformerTrajEncoder.head_scaling = True
TformerTrajEncoder.n_heads = 8
TformerTrajEncoder.n_layers = 3
TformerTrajEncoder.norm = 'layer'
TformerTrajEncoder.normformer_norms = True
TformerTrajEncoder.sigma_reparam = True

# Parameters for TimestepTransformer:
# ==============================================================================
# None.

# Parameters for TokenEmbedding:
# ==============================================================================
# None.

# Parameters for TransformerTurnEmbedding:
# ==============================================================================
TransformerTurnEmbedding.dropout = 0.05