def _predict_obs(self, predict_fn, rng):
obs_repr = np.zeros(
(self._steps.shape[0], self._obs_repr_length), dtype=np.int32,
)
for (i, subrng) in enumerate(jax_random.split(rng, self._obs_repr_length)):
log_probs = predict_fn(self._last_symbols, rng=subrng)
self._last_symbols = utils.gumbel_sample(log_probs)
obs_repr[:, i] = self._last_symbols[:, 0]
return self._obs_serializer.deserialize(obs_repr)
def PolicySchedule(
history,
observation_metrics=(
('train', 'metrics/accuracy'),
('train', 'metrics/loss'),
('eval', 'metrics/accuracy'),
('eval', 'metrics/loss'),
),
include_controls_in_observation=False,
control_configs=(
# (name, start, (low, high), flip)
('learning_rate', 1e-3, (1e-9, 10.0), False), ),
observation_range=(0.0, 10.0),
action_multipliers=(1.0 / 1.5, 1.0 / 1.25, 1.0, 1.25, 1.5),
policy_and_value_model=trax_models.FrameStackMLP,
policy_and_value_two_towers=False,
policy_and_value_vocab_size=None,
policy_dir=gin.REQUIRED,
temperature=1.0,
):
"""Learning rate schedule controlled by a learned policy.
Args:
history: the history of training and evaluation (History object).
observation_metrics: list of pairs (mode, metric), as in the History object.
include_controls_in_observation: bool, whether to include the controls in
observations.
control_configs: control configs, see trax.rl.envs.OnlineTuneEnv.
observation_range: tuple (low, high), range to clip the metrics to.
action_multipliers: sequence of LR multipliers that policy actions
correspond to.
policy_and_value_model: Trax model to use as the policy.
policy_and_value_two_towers: bool, whether the action distribution and value
prediction is computed by separate model towers.
policy_and_value_vocab_size: vocabulary size of a policy and value network
operating on serialized representation. If None, use raw continuous
representation.
policy_dir: directory with the policy checkpoint.
temperature: temperature for sampling from the policy.
Returns:
a function nontrainable_params(step): float -> {'name': float}, the
step-dependent schedule for nontrainable parameters.
"""
# Turn the history into observations for the policy. If we don't have any,
# return the initial learning rate.
start_time = time.time()
observations = online_tune.history_to_observations(
history, observation_metrics, observation_range,
control_configs if include_controls_in_observation else None)
logging.vlog(1, 'Building observations took %0.2f sec.',
time.time() - start_time)
if observations.shape[0] == 0:
controls = {
name: start_value
for (name, start_value, _, _) in control_configs
}
return lambda _: controls
# Build the policy network and load its parameters.
start_time = time.time()
net = ppo.policy_and_value_net(
n_controls=len(control_configs),
n_actions=len(action_multipliers),
vocab_size=policy_and_value_vocab_size,
bottom_layers_fn=policy_and_value_model,
two_towers=policy_and_value_two_towers,
)
logging.vlog(1, 'Building the policy network took %0.2f sec.',
time.time() - start_time)
start_time = time.time()
# (opt_state, state, epoch, opt_step, history)
(opt_state, state, _, _, _) = ppo.maybe_restore_opt_state(policy_dir)
assert opt_state is not None, 'Policy checkpoint not found.'
(params, _, _) = opt_state
logging.vlog(1, 'Restoring the policy parameters took %0.2f sec.',
time.time() - start_time)
# Run the policy and sample an action.
seed = random.randint(0, 2**31 - 1)
rng = jax_random.get_prng(seed=seed)
start_time = time.time()
(low, high) = observation_range
observation_space = gym.spaces.Box(shape=observations.shape[1:],
low=low,
high=high)
action_space = gym.spaces.MultiDiscrete(nvec=(len(action_multipliers), ) *
len(control_configs))
n_timesteps = observations.shape[0]
rewards_to_actions = ppo.init_rewards_to_actions(
policy_and_value_vocab_size, observation_space, action_space,
n_timesteps)
# (log_probs, value_preds, state, rng)
(log_probs, _, _, _) = ppo.run_policy(
policy_and_value_net_apply=net,
observations=np.array([observations]),
lengths=np.array([n_timesteps]),
weights=params,
state=state,
rng=rng,
vocab_size=policy_and_value_vocab_size,
observation_space=observation_space,
action_space=action_space,
rewards_to_actions=rewards_to_actions,
)
logging.vlog(1, 'Running the policy took %0.2f sec.',
time.time() - start_time)
# Sample from the action distribution for the last timestep.
assert log_probs.shape == (1, len(control_configs),
len(action_multipliers))
action = utils.gumbel_sample(log_probs[0] / temperature)
# Get new controls.
controls = {
# name: value
control_config[0]: online_tune.update_control( # pylint: disable=g-complex-comprehension
control_config, control_action, history, action_multipliers)
for (control_action, control_config) in zip(action, control_configs)
}
return lambda _: controls
def PolicySchedule(
history,
observation_metrics=(
("train", "metrics/accuracy"),
("train", "metrics/loss"),
("eval", "metrics/accuracy"),
("eval", "metrics/loss"),
),
include_controls_in_observation=False,
control_configs=(
# (name, start, (low, high), flip)
("learning_rate", 1e-3, (1e-9, 10.0), False), ),
observation_range=(0.0, 10.0),
action_multipliers=(1.0 / 1.5, 1.0 / 1.25, 1.0, 1.25, 1.5),
policy_and_value_model=trax_models.FrameStackMLP,
policy_and_value_two_towers=False,
policy_and_value_vocab_size=None,
policy_dir=gin.REQUIRED,
temperature=1.0,
):
"""Learning rate schedule controlled by a learned policy.
Args:
history: the history of training and evaluation (History object).
observation_metrics: list of pairs (mode, metric), as in the History object.
include_controls_in_observation: bool, whether to include the controls in
observations.
control_configs: control configs, see trax.rl.envs.OnlineTuneEnv.
observation_range: tuple (low, high), range to clip the metrics to.
action_multipliers: sequence of LR multipliers that policy actions
correspond to.
policy_and_value_model: Trax model to use as the policy.
policy_and_value_two_towers: bool, whether the action distribution and value
prediction is computed by separate model towers.
policy_and_value_vocab_size: vocabulary size of a policy and value network
operating on serialized representation. If None, use raw continuous
representation.
policy_dir: directory with the policy checkpoint.
temperature: temperature for sampling from the policy.
Returns:
a function nontrainable_params(step): float -> {"name": float}, the
step-dependent schedule for nontrainable parameters.
"""
# Turn the history into observations for the policy. If we don't have any,
# return the initial learning rate.
start_time = time.time()
observations = online_tune.history_to_observations(
history, observation_metrics, observation_range,
control_configs if include_controls_in_observation else None)
logging.vlog(1, "Building observations took %0.2f sec.",
time.time() - start_time)
if observations.shape[0] == 0:
controls = {
name: start_value
for (name, start_value, _, _) in control_configs
}
return lambda _: controls
assert policy_and_value_vocab_size is None, (
"Serialized policies are not supported yet.")
# Build the policy network and load its parameters.
start_time = time.time()
net = ppo.policy_and_value_net(
n_controls=len(control_configs),
n_actions=len(action_multipliers),
vocab_size=policy_and_value_vocab_size,
bottom_layers_fn=policy_and_value_model,
two_towers=policy_and_value_two_towers,
)
logging.vlog(1, "Building the policy network took %0.2f sec.",
time.time() - start_time)
start_time = time.time()
# (opt_state, state, epoch, opt_step)
(opt_state, state, _, _) = ppo.maybe_restore_opt_state(policy_dir)
assert opt_state is not None, "Policy checkpoint not found."
(params, _) = opt_state
logging.vlog(1, "Restoring the policy parameters took %0.2f sec.",
time.time() - start_time)
# Run the policy and sample an action.
seed = random.randint(0, 2**31 - 1)
rng = jax_random.get_prng(seed=seed)
start_time = time.time()
# ((log_probs, value_preds), state). We have no way to pass state to the next
# step, but that should be fine.
(log_probs, _) = (net(np.array([observations]),
params=params,
state=state,
rng=rng))
logging.vlog(1, "Running the policy took %0.2f sec.",
time.time() - start_time)
# Sample from the action distribution for the last timestep.
assert log_probs.shape == (1, len(control_configs) * observations.shape[0],
len(action_multipliers))
action = utils.gumbel_sample(log_probs[0, -len(control_configs):, :] /
temperature)
# Get new controls.
controls = {
# name: value
control_config[0]: online_tune.update_control( # pylint: disable=g-complex-comprehension
control_config, control_action, history, action_multipliers)
for (control_action, control_config) in zip(action, control_configs)
}
return lambda _: controls
