Spaces:
Running
Running
File size: 8,631 Bytes
5146e76 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 | import jax
from agents.mlp_actor_critic_agent import MLPActorCriticPolicy, ActorWithDoubleCriticPolicy, \
ActorWithConditionalCriticPolicy, PseudoActorWithDoubleCriticPolicy, \
PseudoActorWithConditionalCriticPolicy
from agents.rnn_actor_critic_agent import RNNActorCriticPolicy
from agents.s5_actor_critic_agent import S5ActorCriticPolicy
from agents.liam_agent import LIAMPolicy, initialize_liam_encoder_decoder
from agents.meliba_agent import MeLIBAPolicy, initialize_meliba_encoder_decoder
def initialize_s5_agent(config, env, rng):
"""Initialize an S5 agent with the given config.
Args:
config: dict, config for the agent
env: gymnasium environment
rng: jax.random.PRNGKey, random key for initialization
Returns:
policy: S5ActorCriticPolicy, the policy object
params: dict, initial parameters for the agent
"""
# Create the S5 policy with direct parameters
policy = S5ActorCriticPolicy(
action_dim=env.action_space(env.agents[0]).n,
obs_dim=config.get("POLICY_INPUT_DIM", env.observation_space(env.agents[0]).shape[0]),
d_model=config.get("S5_D_MODEL", 128),
ssm_size=config.get("S5_SSM_SIZE", 128),
# d_model=config.get("S5_D_MODEL", 16),
# ssm_size=config.get("S5_SSM_SIZE", 16),
ssm_n_layers=config.get("S5_N_LAYERS", 2),
blocks=config.get("S5_BLOCKS", 1),
fc_hidden_dim=config.get("S5_ACTOR_CRITIC_HIDDEN_DIM", 1024),
fc_n_layers=config.get("FC_N_LAYERS", 3),
# fc_hidden_dim=config.get("S5_ACTOR_CRITIC_HIDDEN_DIM", 64),
# fc_n_layers=config.get("FC_N_LAYERS", 2),
s5_activation=config.get("S5_ACTIVATION", "full_glu"),
s5_do_norm=config.get("S5_DO_NORM", True),
s5_prenorm=config.get("S5_PRENORM", True),
s5_do_gtrxl_norm=config.get("S5_DO_GTRXL_NORM", True),
)
rng, init_rng = jax.random.split(rng)
init_params = policy.init_params(init_rng)
return policy, init_params
def initialize_rnn_agent(config, env, rng):
"""Initialize an RNN agent with the given config.
Args:
config: dict, config for the agent
env: gymnasium environment
rng: jax.random.PRNGKey, random key for initialization
Returns:
policy: RNNActorCriticPolicy, the policy object
params: dict, initial parameters for the agent
"""
# Create the RNN policy
policy = RNNActorCriticPolicy(
action_dim=env.action_space(env.agents[0]).n,
obs_dim=config.get("POLICY_INPUT_DIM", env.observation_space(env.agents[0]).shape[0]),
activation=config.get("ACTIVATION", "tanh"),
fc_hidden_dim=config.get("FC_HIDDEN_DIM", 64),
gru_hidden_dim=config.get("GRU_HIDDEN_DIM", 64),
)
rng, init_rng = jax.random.split(rng)
init_params = policy.init_params(init_rng)
return policy, init_params
def initialize_mlp_agent(config, env, rng):
"""
Initialize an MLP agent with the given config.
"""
policy = MLPActorCriticPolicy(
action_dim=env.action_space(env.agents[0]).n,
obs_dim=config.get("POLICY_INPUT_DIM", env.observation_space(env.agents[0]).shape[0]),
activation=config.get("ACTIVATION", "tanh"),
fc_hidden_dim=config.get("FC_HIDDEN_DIM", 64),
)
rng, init_rng = jax.random.split(rng)
init_params = policy.init_params(init_rng)
return policy, init_params
def initialize_actor_with_double_critic(config, env, rng):
"""Initialize an actor with double critic with the given config."""
policy = ActorWithDoubleCriticPolicy(
action_dim=env.action_space(env.agents[0]).n,
obs_dim=config.get("POLICY_INPUT_DIM", env.observation_space(env.agents[0]).shape[0]),
activation=config.get("ACTIVATION", "tanh"),
fc_hidden_dim=config.get("FC_HIDDEN_DIM", 64),
)
rng, init_rng = jax.random.split(rng)
init_params = policy.init_params(init_rng)
return policy, init_params
def initialize_pseudo_actor_with_double_critic(config, env, rng):
"""Initialize a pseudo actor with double critic with the given config."""
policy = PseudoActorWithDoubleCriticPolicy(
action_dim=env.action_space(env.agents[0]).n,
obs_dim=config.get("POLICY_INPUT_DIM", env.observation_space(env.agents[0]).shape[0]),
activation=config.get("ACTIVATION", "tanh"),
fc_hidden_dim=config.get("FC_HIDDEN_DIM", 64),
)
rng, init_rng = jax.random.split(rng)
init_params = policy.init_params(init_rng)
return policy, init_params
def initialize_actor_with_conditional_critic(config, env, rng):
"""Initialize an actor with conditional critic with the given config."""
policy = ActorWithConditionalCriticPolicy(
action_dim=env.action_space(env.agents[0]).n,
obs_dim=config.get("POLICY_INPUT_DIM", env.observation_space(env.agents[0]).shape[0]),
pop_size=config["POP_SIZE"],
activation=config.get("ACTIVATION", "tanh"),
fc_hidden_dim=config.get("FC_HIDDEN_DIM", 64),
)
rng, init_rng = jax.random.split(rng)
init_params = policy.init_params(init_rng)
return policy, init_params
def initialize_pseudo_actor_with_conditional_critic(config, env, rng):
"""Initialize a pseudo actor with conditional critic with the given config."""
policy = PseudoActorWithConditionalCriticPolicy(
action_dim=env.action_space(env.agents[0]).n,
obs_dim=config.get("POLICY_INPUT_DIM", env.observation_space(env.agents[0]).shape[0]),
pop_size=config["POP_SIZE"],
activation=config.get("ACTIVATION", "tanh"),
fc_hidden_dim=config.get("FC_HIDDEN_DIM", 64),
)
rng, init_rng = jax.random.split(rng)
init_params = policy.init_params(init_rng)
return policy, init_params
def initialize_liam_agent(config, env, rng):
"""Initialize the LIAM ego agent with the given config.
Args:
config: dict, config for the agent
env: gymnasium environment
rng: jax.random.PRNGKey, random key for initialization
Returns:
liam: LIAMPolicy, the policy object
params: tuple, initial parameters for the {encoder, decoder} and policy
"""
rng, init_encoder_decoder_rng, init_policy_rng = jax.random.split(rng, 3)
# Initialize the policy based on the specified type
if config["EGO_ACTOR_TYPE"] == "s5":
ego_policy, init_ego_params = initialize_s5_agent(config, env, init_policy_rng)
elif config["EGO_ACTOR_TYPE"] == "mlp":
ego_policy, init_ego_params = initialize_mlp_agent(config, env, init_policy_rng)
elif config["EGO_ACTOR_TYPE"] == "rnn":
ego_policy, init_ego_params = initialize_rnn_agent(config, env, init_policy_rng)
# Initialize the encoder and decoder for LIAM
encoder, decoder, init_encoder_decoder_params = initialize_liam_encoder_decoder(config, env, init_encoder_decoder_rng)
liam = LIAMPolicy(
policy=ego_policy,
encoder=encoder,
decoder=decoder
)
params = {'encoder': init_encoder_decoder_params['encoder'],
'decoder': init_encoder_decoder_params['decoder'],
'policy': init_ego_params}
return liam, params
def initialize_meliba_agent(config, env, rng):
"""Initialize the MeLIBA ego agent with the given config.
Args:
config: dict, config for the agent
env: gymnasium environment
rng: jax.random.PRNGKey, random key for initialization
Returns:
meliba: MeLIBAPolicy, the policy object
params: tuple, initial parameters for the {encoder, decoder} and policy
"""
rng, init_encoder_decoder_rng, init_policy_rng = jax.random.split(rng, 3)
# Initialize the policy based on the specified type
if config["EGO_ACTOR_TYPE"] == "s5":
ego_policy, init_ego_params = initialize_s5_agent(config, env, init_policy_rng)
elif config["EGO_ACTOR_TYPE"] == "mlp":
ego_policy, init_ego_params = initialize_mlp_agent(config, env, init_policy_rng)
elif config["EGO_ACTOR_TYPE"] == "rnn":
ego_policy, init_ego_params = initialize_rnn_agent(config, env, init_policy_rng)
# Initialize the encoder and decoder for LIAM
encoder, decoder, init_encoder_decoder_params = initialize_meliba_encoder_decoder(config, env, init_encoder_decoder_rng)
meliba = MeLIBAPolicy(
policy=ego_policy,
encoder=encoder,
decoder=decoder
)
params = {'encoder': init_encoder_decoder_params['encoder'],
'decoder': init_encoder_decoder_params['decoder'],
'policy': init_ego_params}
return meliba, params
|