Function make_ddpg_actor

advanced_source/coding_ddpg.py:726–766

(
    transform_state_dict,
    device="cpu",
)

def make_ddpg_actor(
    transform_state_dict,
    device="cpu",
):
    proof_environment = make_transformed_env(make_env())
    proof_environment.transform[2].init_stats(3)
    proof_environment.transform[2].load_state_dict(transform_state_dict)

    out_features = proof_environment.action_spec.shape[-1]

    actor_net = DdpgMlpActor(
        action_dim=out_features,
    )

    in_keys = ["observation_vector"]
    out_keys = ["param"]

    actor = TensorDictModule(
        actor_net,
        in_keys=in_keys,
        out_keys=out_keys,
    )

    actor = ProbabilisticActor(
        actor,
        distribution_class=TanhDelta,
        in_keys=["param"],
        spec=CompositeSpec(action=proof_environment.action_spec),
    ).to(device)

    q_net = DdpgMlpQNet()

    in_keys = in_keys + ["action"]
    qnet = ValueOperator(
        in_keys=in_keys,
        module=q_net,
    ).to(device)

    # initialize lazy modules
    qnet(actor(proof_environment.reset().to(device)))
    return actor, qnet


actor, qnet = make_ddpg_actor(