def test_default_preprocessor(self):
rngs = PRNGSequence(13)
box = gym.spaces.Box(low=0, high=1, shape=(2, 3))
dsc = gym.spaces.Discrete(7)
mbn = gym.spaces.MultiBinary(11)
mds = gym.spaces.MultiDiscrete(nvec=[3, 5])
tup = gym.spaces.Tuple((box, dsc, mbn, mds))
dct = gym.spaces.Dict({'box': box, 'dsc': dsc, 'mbn': mbn, 'mds': mds})
self.assertArrayShape(default_preprocessor(box)(next(rngs), box.sample()), (1, 2, 3))
self.assertArrayShape(default_preprocessor(dsc)(next(rngs), dsc.sample()), (1, 7))
self.assertArrayShape(default_preprocessor(mbn)(next(rngs), mbn.sample()), (1, 11))
self.assertArrayShape(default_preprocessor(mds)(next(rngs), mds.sample())[0], (1, 3))
self.assertArrayShape(default_preprocessor(mds)(next(rngs), mds.sample())[1], (1, 5))
self.assertArrayShape(default_preprocessor(tup)(next(rngs), tup.sample())[0], (1, 2, 3))
self.assertArrayShape(default_preprocessor(tup)(next(rngs), tup.sample())[1], (1, 7))
self.assertArrayShape(default_preprocessor(tup)(next(rngs), tup.sample())[2], (1, 11))
self.assertArrayShape(default_preprocessor(tup)(next(rngs), tup.sample())[3][0], (1, 3))
self.assertArrayShape(default_preprocessor(tup)(next(rngs), tup.sample())[3][1], (1, 5))
self.assertArrayShape(default_preprocessor(dct)(next(rngs), dct.sample())['box'], (1, 2, 3))
self.assertArrayShape(default_preprocessor(dct)(next(rngs), dct.sample())['dsc'], (1, 7))
self.assertArrayShape(default_preprocessor(dct)(next(rngs), dct.sample())['mbn'], (1, 11))
self.assertArrayShape(default_preprocessor(dct)(next(rngs), dct.sample())['mds'][0], (1, 3))
self.assertArrayShape(default_preprocessor(dct)(next(rngs), dct.sample())['mds'][1], (1, 5))
def test_add_noise():
rng = PRNGSequence(0)
assert np.isclose(add_noise(0.0, next(rng), 0.0), 0.0)
assert np.isclose(add_noise(0.0, next(rng), 0.0, -1.0, 1.0), 0.0)
assert np.isclose(add_noise(0.0, next(rng), 100.0, -1.0, 1.0, 0.0, 0.0),
0.0)
assert -1.0 <= add_noise(0.0, next(rng), 100.0, -1.0, 1.0) <= 1.0
assert -20.0 <= add_noise(0.0, next(rng), 100.0, -20.0, 20.0) <= 20.0
def test_argmax_random_tiebreaking(self):
rngs = PRNGSequence(13)
vec = jnp.ones(shape=(5,))
mat = jnp.ones(shape=(3, 5))
self.assertEqual(argmax(next(rngs), vec), 2) # not zero
self.assertArrayAlmostEqual(argmax(next(rngs), mat), [1, 1, 3])
def test_tree_sample(self):
rngs = PRNGSequence(42)
tn = get_transition_batch(self.env_discrete, batch_size=5)
tn_sample = tree_sample(tn, next(rngs), n=3)
assert tn_sample.batch_size == 3
tn_sample = tree_sample(tn, next(rngs), n=7, replace=True)
assert tn_sample.batch_size == 7
msg = r"Cannot take a larger sample than population when 'replace=False'"
with self.assertRaisesRegex(ValueError, msg):
tree_sample(tn, next(rngs), n=7, replace=False)
def __init__(
self,
state_dim,
action_dim,
max_action,
lr=3e-4,
discount=0.99,
tau=0.005,
policy_noise=0.2,
expl_noise=0.1,
noise_clip=0.5,
policy_freq=2,
seed=0,
):
self.rng = PRNGSequence(seed)
actor_input_dim = [((1, state_dim), jnp.float32)]
init_rng = next(self.rng)
actor = build_td3_actor_model(actor_input_dim, action_dim, max_action,
init_rng)
self.actor_target = build_td3_actor_model(actor_input_dim, action_dim,
max_action, init_rng)
actor_optimizer = optim.Adam(learning_rate=lr).create(actor)
self.actor_optimizer = jax.device_put(actor_optimizer)
init_rng = next(self.rng)
critic_input_dim = [
((1, state_dim), jnp.float32),
((1, action_dim), jnp.float32),
]
critic = build_td3_critic_model(critic_input_dim, init_rng)
self.critic_target = build_td3_critic_model(critic_input_dim, init_rng)
critic_optimizer = optim.Adam(learning_rate=lr).create(critic)
self.critic_optimizer = jax.device_put(critic_optimizer)
self.max_action = max_action
self.discount = discount
self.tau = tau
self.policy_noise = policy_noise
self.expl_noise = expl_noise
self.noise_clip = noise_clip
self.policy_freq = policy_freq
self.total_it = 0
def __init__(
self,
state_dim: int,
action_dim: int,
max_action: float,
lr: float = 3e-4,
discount: float = 0.99,
tau: float = 0.005,
policy_noise: float = 0.2,
expl_noise: float = 0.1,
noise_clip: float = 0.5,
policy_freq: int = 2,
seed: int = 0,
):
self.rng = PRNGSequence(seed)
actor_input_dim = (1, state_dim)
init_rng = next(self.rng)
actor_params = build_td3_actor_model(
actor_input_dim, action_dim, max_action, init_rng
)
self.actor_target_params = build_td3_actor_model(
actor_input_dim, action_dim, max_action, init_rng
)
actor_optimizer = optim.Adam(learning_rate=lr).create(actor_params)
self.actor_optimizer = jax.device_put(actor_optimizer)
init_rng = next(self.rng)
critic_input_dim = [(1, state_dim), (1, action_dim)]
critic_params = build_td3_critic_model(critic_input_dim, init_rng)
self.critic_target_params = build_td3_critic_model(critic_input_dim, init_rng)
critic_optimizer = optim.Adam(learning_rate=lr).create(critic_params)
self.critic_optimizer = jax.device_put(critic_optimizer)
self.max_action = max_action
self.discount = discount
self.tau = tau
self.policy_noise = policy_noise
self.expl_noise = expl_noise
self.noise_clip = noise_clip
self.policy_freq = policy_freq
self.action_dim = action_dim
self.total_it = 0
def __init__(
self,
features: tp.Iterable[h5py.Group],
labels: h5py.Group,
batch_size: int,
shuffle_rng=0,
):
def get_features(group):
if hasattr(group, "train"):
return group["train"]
return group
super().__init__((get_features(g) for g in features), labels,
batch_size)
self._rng = PRNGSequence(shuffle_rng)
def __init__(
self,
state_dim: int,
action_dim: int,
max_action: float,
discount: float = 0.99,
tau: float = 0.005,
policy_freq: int = 2,
lr: float = 3e-4,
entropy_tune: bool = True,
seed: int = 0,
):
self.rng = PRNGSequence(seed)
actor_input_dim = (1, state_dim)
actor_params = build_gaussian_policy_model(actor_input_dim, action_dim,
max_action, next(self.rng))
actor_optimizer = optim.Adam(learning_rate=lr).create(actor_params)
self.actor_optimizer = jax.device_put(actor_optimizer)
init_rng = next(self.rng)
critic_input_dim = [(1, state_dim), (1, action_dim)]
critic_params = build_double_critic_model(critic_input_dim, init_rng)
self.critic_target_params = build_double_critic_model(
critic_input_dim, init_rng)
critic_optimizer = optim.Adam(learning_rate=lr).create(critic_params)
self.critic_optimizer = jax.device_put(critic_optimizer)
self.entropy_tune = entropy_tune
log_alpha_params = build_constant_model(-3.5, next(self.rng))
log_alpha_optimizer = optim.Adam(
learning_rate=lr).create(log_alpha_params)
self.log_alpha_optimizer = jax.device_put(log_alpha_optimizer)
self.target_entropy = -action_dim
self.max_action = max_action
self.discount = discount
self.tau = tau
self.policy_freq = policy_freq
self.action_dim = action_dim
self.total_it = 0
def __init__(
self,
language_dim: int,
vision_dim: Tuple[int, int],
num_embeddings: int,
embedding_dim: int,
memory_hidden_dim: int,
tokenizer: Tokenizer,
discount: float = 0.9,
hidden_size: int = 512,
lr: float = 1e-5,
policy_eps: float = 0.1,
baseline_eps: float = 0.5,
entropy_eps: float = 1e-5,
reconstruction_eps: float = 1.0,
seed: int = 0,
):
self.rng = PRNGSequence(seed)
input_dims = [
(1, language_dim),
(1, *vision_dim),
]
agent_params = build_fast_slow_agent_model(
input_dims,
memory_hidden_dim,
embedding_dim,
num_embeddings,
embedding_dim,
next(self.rng),
)
agent_optimizer = optim.Adam(learning_rate=lr).create(agent_params)
self.agent_optimizer = jax.device_put(agent_optimizer)
self.num_embeddings = num_embeddings
self.embedding_dim = embedding_dim
self.hidden_size = hidden_size
self.discount = discount
self.policy_eps = policy_eps
self.baseline_eps = baseline_eps
self.entropy_eps = entropy_eps
self.reconstruction_eps = reconstruction_eps
self.tokenizer = tokenizer
self.total_it = 0
def test_argmax_consistent(self):
rngs = PRNGSequence(13)
vec = jax.random.normal(next(rngs), shape=(5, ))
mat = jax.random.normal(next(rngs), shape=(3, 5))
ten = jax.random.normal(next(rngs), shape=(3, 5, 7))
self.assertEqual(argmax(next(rngs), vec), jnp.argmax(vec, axis=-1))
self.assertArrayAlmostEqual(argmax(next(rngs), mat),
jnp.argmax(mat, axis=-1))
self.assertArrayAlmostEqual(argmax(next(rngs), mat, axis=0),
jnp.argmax(mat, axis=0))
self.assertArrayAlmostEqual(argmax(next(rngs), ten),
jnp.argmax(ten, axis=-1))
self.assertArrayAlmostEqual(argmax(next(rngs), ten, axis=0),
jnp.argmax(ten, axis=0))
self.assertArrayAlmostEqual(argmax(next(rngs), ten, axis=1),
jnp.argmax(ten, axis=1))
def main():
loss_obj = hk.transform(loss_fn, apply_rng=True)
# Initial parameter values are typically random. In JAX you need a key in order
# to generate random numbers and so Haiku requires you to pass one in.
rng = PRNGSequence(42)
# `init` runs your function, as such we need an example input. Typically you can
# pass "dummy" inputs (e.g. ones of the same shape and dtype) since initialization
# is not usually data dependent.
shape = [([1000], float)]
adam = optim.Adam(learning_rate=0.1)
partial = Net.partial()
_, params = partial.init_by_shape(next(rng), shape)
net = nn.Model(partial, params)
optimizer = jax.device_put(adam.create(net))
_, params = partial.init_by_shape(next(rng), shape) # HERE
net = net.replace(params=params)
optimizer = jax.device_put(adam.create(net)) # HERE
def __init__(
self,
num_agent_steps,
state_space,
action_space,
seed,
max_grad_norm,
gamma,
):
np.random.seed(seed)
self.rng = PRNGSequence(seed)
self.agent_step = 0
self.episode_step = 0
self.learning_step = 0
self.num_agent_steps = num_agent_steps
self.state_space = state_space
self.action_space = action_space
self.gamma = gamma
self.max_grad_norm = max_grad_norm
self.discrete_action = False if type(action_space) == Box else True
def __init__(
self,
batch_size,
buffer_size,
discount,
env_id,
eval_freq,
eval_episodes,
learning_rate,
load_path,
max_time_steps,
actor_freq,
save_freq,
save_model,
seed,
start_time_steps,
tau,
train_steps,
render,
use_tune,
prefix=None,
env=None,
):
self.prefix = prefix
seed = int(seed)
policy = "SAC"
if env_id == "levels":
env_id = None
self.use_tune = use_tune
self.max_time_steps = int(max_time_steps) if max_time_steps else None
self.start_time_steps = int(start_time_steps)
self.train_steps = int(train_steps)
self.batch_size = int(batch_size)
self.buffer_size = int(buffer_size)
self.eval_freq = eval_freq
def make_env():
if env is None:
return Environment.wrap(gym.make(env_id) if env_id else Env(100))
return env
def eval_policy():
eval_env = make_env()
eval_env.seed(seed)
avg_reward = 0.0
it = (
itertools.count() if render else tqdm(range(eval_episodes), desc="eval")
)
for _ in it:
eval_time_step = eval_env.reset()
while not eval_time_step.last():
if render:
eval_env.render()
action = policy.select_action(eval_time_step.observation)
eval_time_step = eval_env.step(action)
avg_reward += eval_time_step.reward
avg_reward /= eval_episodes
self.report(eval_reward=avg_reward)
return avg_reward
self.eval_policy = eval_policy
self.report(policy=policy)
self.report(env=env_id)
self.report(seed=seed)
if save_model and not os.path.exists("./models"):
os.makedirs("./models")
self.env = env = make_env()
assert isinstance(env, Environment)
# Set seeds
np.random.seed(seed)
state_shape = env.observation_spec().shape
action_dim = env.action_spec().shape[0]
max_action = env.max_action()
# Initialize policy
self.policy = policy = SAC.SAC( # TODO
state_shape=state_shape,
action_dim=action_dim,
max_action=max_action,
save_freq=save_freq,
discount=discount,
lr=learning_rate,
actor_freq=actor_freq,
tau=tau,
)
self.rng = PRNGSequence(seed)
def test_preprocess_state():
rng = PRNGSequence(0)
state = np.random.randint(0, 256, size=(64, 64, 3)).astype(np.uint8)
state = preprocess_state(state, next(rng))
assert (-0.5 <= state).all() and (state <= 0.5).all()
def __init__(
self,
state_dim,
action_dim,
max_action,
discount=0.99,
lr=3e-4,
eps_q=0.1,
eps_mu=0.1,
eps_sig=1e-4,
temp_steps=10,
target_freq=250,
seed=0,
):
self.rng = PRNGSequence(seed)
init_rng = next(self.rng)
actor_input_dim = [((1, state_dim), jnp.float32)]
actor = build_gaussian_policy_model(
actor_input_dim, action_dim, max_action, init_rng
)
self.actor_target = build_gaussian_policy_model(
actor_input_dim, action_dim, max_action, init_rng
)
actor_optimizer = optim.Adam(learning_rate=lr).create(actor)
self.actor_optimizer = jax.device_put(actor_optimizer)
init_rng = next(self.rng)
critic_input_dim = [
((1, state_dim), jnp.float32),
((1, action_dim), jnp.float32),
]
critic = build_double_critic_model(critic_input_dim, init_rng)
self.critic_target = build_double_critic_model(critic_input_dim, init_rng)
critic_optimizer = optim.Adam(learning_rate=lr).create(critic)
self.critic_optimizer = jax.device_put(critic_optimizer)
temp = build_constant_model(1.0, next(self.rng))
temp_optimizer = optim.Adam(learning_rate=lr).create(temp)
self.temp_optimizer = jax.device_put(temp_optimizer)
mu_lagrange = build_constant_model(1.0, next(self.rng))
mu_lagrange_optimizer = optim.Adam(learning_rate=lr).create(mu_lagrange)
self.mu_lagrange_optimizer = jax.device_put(mu_lagrange_optimizer)
sig_lagrange = build_constant_model(100.0, next(self.rng))
sig_lagrange_optimizer = optim.Adam(learning_rate=lr).create(sig_lagrange)
self.sig_lagrange_optimizer = jax.device_put(sig_lagrange_optimizer)
self.eps_q = eps_q
self.eps_mu = eps_mu
self.eps_sig = eps_sig
self.temp_steps = temp_steps
self.max_action = max_action
self.discount = discount
self.target_freq = target_freq
self.action_dim = action_dim
self.total_it = 0
def __init__(
self,
state_dim: int,
action_dim: int,
max_action: float,
discount: float = 0.99,
lr: float = 3e-4,
eps_eta: float = 0.1,
eps_mu: float = 5e-4,
eps_sig: float = 1e-5,
target_freq: int = 250,
seed: int = 0,
):
self.rng = PRNGSequence(seed)
init_rng = next(self.rng)
actor_input_dim = (1, state_dim)
actor_params = build_gaussian_policy_model(
actor_input_dim, action_dim, max_action, init_rng
)
self.actor_target_params = build_gaussian_policy_model(
actor_input_dim, action_dim, max_action, init_rng
)
actor_optimizer = optim.Adam(learning_rate=lr).create(actor_params)
self.actor_optimizer = jax.device_put(actor_optimizer)
init_rng = next(self.rng)
critic_input_dim = [(1, state_dim), (1, action_dim)]
critic_params = build_double_critic_model(critic_input_dim, init_rng)
self.critic_target_params = build_double_critic_model(
critic_input_dim, init_rng
)
critic_optimizer = optim.Adam(learning_rate=lr).create(critic_params)
self.critic_optimizer = jax.device_put(critic_optimizer)
mu_lagrange_params = build_constant_model(
1.0, absolute=True, init_rng=next(self.rng)
)
mu_lagrange_optimizer = optim.Adam(learning_rate=lr).create(mu_lagrange_params)
self.mu_lagrange_optimizer = jax.device_put(mu_lagrange_optimizer)
sig_lagrange_params = build_constant_model(
100.0, absolute=True, init_rng=next(self.rng)
)
sig_lagrange_optimizer = optim.Adam(learning_rate=lr).create(
sig_lagrange_params
)
self.sig_lagrange_optimizer = jax.device_put(sig_lagrange_optimizer)
self.temp = 1.0
self.eps_eta = eps_eta
self.eps_mu = eps_mu
self.eps_sig = eps_sig
self.max_action = max_action
self.discount = discount
self.target_freq = target_freq
self.state_dim = state_dim
self.action_dim = action_dim
self.total_it = 0
def seed(self, seed=None):
self.rng = PRNGSequence(seed)
def __init__(
self,
state_shape,
action_dim,
max_action,
save_freq,
discount=0.99,
tau=0.005,
actor_freq=2,
lr=3e-4,
entropy_tune=False,
seed=0,
):
self.rng = PRNGSequence(seed)
actor_input_dim = [((1, *state_shape), jnp.float32)]
critic_input_dim = self.critic_input_dim = [
((1, *state_shape), jnp.float32),
((1, action_dim), jnp.float32),
]
self.actor = None
self.critic = None
self.log_alpha = None
self.entropy_tune = entropy_tune
self.target_entropy = -action_dim
self.adam = Optimizers(
actor=optim.Adam(learning_rate=lr),
critic=optim.Adam(learning_rate=lr),
log_alpha=optim.Adam(learning_rate=lr),
)
self.module = Modules(
actor=GaussianPolicy.partial(action_dim=action_dim,
max_action=max_action),
critic=DoubleCritic.partial(),
alpha=Constant.partial(start_value=1),
)
self.optimizer = None
self.max_action = max_action
self.discount = discount
self.tau = tau
self.policy_freq = actor_freq
self.save_freq = save_freq
self.total_it = 0
self.model = None
def new_params(module: nn.Module, shape=None):
_, params = (module.init(next(self.rng)) if shape is None else
module.init_by_shape(next(self.rng), shape))
return params
def new_model(module: nn.Module, shape=None) -> nn.Model:
return nn.Model(module, new_params(module, shape))
def update_model(model: nn.Model, shape=None) -> nn.Model:
return model.replace(params=new_params(model.module, shape))
def reset_models() -> Models:
if self.model is None:
critic = new_model(self.module.critic, critic_input_dim)
return Models(
actor=new_model(self.module.actor, actor_input_dim),
critic=critic,
target_critic=critic.replace(params=critic.params),
alpha=new_model(self.module.alpha),
)
else:
critic = update_model(self.model.critic, critic_input_dim)
return Models(
actor=update_model(self.model.actor, actor_input_dim),
critic=critic,
target_critic=critic.replace(params=critic.params),
alpha=update_model(self.model.alpha),
)
self.reset_models = reset_models
def reset_optimizer(adam: Adam, model: nn.Model) -> Optimizer:
return jax.device_put(adam.create(model))
def reset_optimizers() -> Optimizers:
return Optimizers(
actor=reset_optimizer(self.adam.actor, self.model.actor),
critic=reset_optimizer(self.adam.critic, self.model.critic),
log_alpha=reset_optimizer(self.adam.log_alpha,
self.model.alpha),
)
self.reset_optimizers = reset_optimizers
self.i = 0