def test_config_init(self, train_config):
c = train_config
config = HDDPG.generate_config({})
config["frame_config"]["models"] = [
"Actor", "Actor", "Critic", "Critic"
]
config["frame_config"][
"model_kwargs"] = [{
"state_dim": c.observe_dim,
"action_dim": c.action_dim,
"action_range": c.action_range,
}] * 2 + [{
"state_dim": c.observe_dim,
"action_dim": c.action_dim
}] * 2
hddpg = HDDPG.init_from_config(config)
old_state = state = t.zeros([1, c.observe_dim], dtype=t.float32)
action = t.zeros([1, c.action_dim], dtype=t.float32)
hddpg.store_episode([{
"state": {
"state": old_state
},
"action": {
"action": action
},
"next_state": {
"state": state
},
"reward": 0,
"terminal": False,
} for _ in range(3)])
hddpg.update()
def hddpg(self, train_config, device, dtype):
c = train_config
actor = smw(
Actor(c.observe_dim, c.action_dim,
c.action_range).type(dtype).to(device),
device,
device,
)
actor_t = smw(
Actor(c.observe_dim, c.action_dim,
c.action_range).type(dtype).to(device),
device,
device,
)
critic = smw(
Critic(c.observe_dim, c.action_dim).type(dtype).to(device), device,
device)
critic_t = smw(
Critic(c.observe_dim, c.action_dim).type(dtype).to(device), device,
device)
hddpg = HDDPG(
actor,
actor_t,
critic,
critic_t,
t.optim.Adam,
nn.MSELoss(reduction="sum"),
replay_device="cpu",
replay_size=c.replay_size,
)
return hddpg
def hddpg_vis(self, train_config, device, dtype, tmpdir):
# not used for training, only used for testing apis
c = train_config
tmp_dir = tmpdir.make_numbered_dir()
actor = smw(
Actor(c.observe_dim, c.action_dim,
c.action_range).type(dtype).to(device),
device,
device,
)
actor_t = smw(
Actor(c.observe_dim, c.action_dim,
c.action_range).type(dtype).to(device),
device,
device,
)
critic = smw(
Critic(c.observe_dim, c.action_dim).type(dtype).to(device), device,
device)
critic_t = smw(
Critic(c.observe_dim, c.action_dim).type(dtype).to(device), device,
device)
hddpg = HDDPG(
actor,
actor_t,
critic,
critic_t,
t.optim.Adam,
nn.MSELoss(reduction="sum"),
replay_device="cpu",
replay_size=c.replay_size,
visualize=True,
visualize_dir=str(tmp_dir),
)
return hddpg
def hddpg_train(self, train_config, gpu):
c = train_config
# cpu is faster for testing full training.
actor = smw(Actor(c.observe_dim, c.action_dim, c.action_range), "cpu",
"cpu")
actor_t = smw(Actor(c.observe_dim, c.action_dim, c.action_range),
"cpu", "cpu")
critic = smw(Critic(c.observe_dim, c.action_dim), "cpu", "cpu")
critic_t = smw(Critic(c.observe_dim, c.action_dim), "cpu", "cpu")
hddpg = HDDPG(actor,
actor_t,
critic,
critic_t,
t.optim.Adam,
nn.MSELoss(reduction='sum'),
replay_device="cpu",
replay_size=c.replay_size)
return hddpg
def hddpg(self, train_config):
c = train_config
actor = smw(
Actor(c.observe_dim, c.action_dim, c.action_range).to(c.device),
c.device, c.device)
actor_t = smw(
Actor(c.observe_dim, c.action_dim, c.action_range).to(c.device),
c.device, c.device)
critic = smw(
Critic(c.observe_dim, c.action_dim).to(c.device), c.device,
c.device)
critic_t = smw(
Critic(c.observe_dim, c.action_dim).to(c.device), c.device,
c.device)
hddpg = HDDPG(actor,
actor_t,
critic,
critic_t,
t.optim.Adam,
nn.MSELoss(reduction='sum'),
replay_device=c.device,
replay_size=c.replay_size)
return hddpg