"""Learning reward models using preference comparisons.
Trains a reward model and optionally a policy based on preferences
between trajectory fragments.
"""
import abc
import math
import pickle
import re
from collections import defaultdict
from typing import (
Any,
Callable,
Dict,
List,
Mapping,
NamedTuple,
NoReturn,
Optional,
Sequence,
Tuple,
Union,
cast,
overload,
)
import numpy as np
import torch as th
from scipy import special
from stable_baselines3.common import base_class, type_aliases, utils, vec_env
from torch import nn
from torch.utils import data as data_th
from tqdm.auto import tqdm
from imitation.algorithms import base
from imitation.data import rollout, types, wrappers
from imitation.data.types import (
AnyPath,
TrajectoryPair,
TrajectoryWithRew,
TrajectoryWithRewPair,
Transitions,
)
from imitation.policies import exploration_wrapper
from imitation.regularization import regularizers
from imitation.rewards import reward_function, reward_nets, reward_wrapper
from imitation.util import logger as imit_logger
from imitation.util import networks, util
[docs]class TrajectoryGenerator(abc.ABC):
"""Generator of trajectories with optional training logic."""
_logger: imit_logger.HierarchicalLogger
"""Object to log statistics and natural language messages to."""
[docs] def __init__(self, custom_logger: Optional[imit_logger.HierarchicalLogger] = None):
"""Builds TrajectoryGenerator.
Args:
custom_logger: Where to log to; if None (default), creates a new logger.
"""
self.logger = custom_logger or imit_logger.configure()
[docs] @abc.abstractmethod
def sample(self, steps: int) -> Sequence[TrajectoryWithRew]:
"""Sample a batch of trajectories.
Args:
steps: All trajectories taken together should
have at least this many steps.
Returns:
A list of sampled trajectories with rewards (which should
be the environment rewards, not ones from a reward model).
""" # noqa: DAR202
[docs] def train(self, steps: int, **kwargs: Any) -> None:
"""Train an agent if the trajectory generator uses one.
By default, this method does nothing and doesn't need
to be overridden in subclasses that don't require training.
Args:
steps: number of environment steps to train for.
**kwargs: additional keyword arguments to pass on to
the training procedure.
"""
@property
def logger(self) -> imit_logger.HierarchicalLogger:
return self._logger
@logger.setter
def logger(self, value: imit_logger.HierarchicalLogger) -> None:
self._logger = value
[docs]class TrajectoryDataset(TrajectoryGenerator):
"""A fixed dataset of trajectories."""
[docs] def __init__(
self,
trajectories: Sequence[TrajectoryWithRew],
rng: np.random.Generator,
custom_logger: Optional[imit_logger.HierarchicalLogger] = None,
):
"""Creates a dataset loaded from `path`.
Args:
trajectories: the dataset of rollouts.
rng: RNG used for shuffling dataset.
custom_logger: Where to log to; if None (default), creates a new logger.
"""
super().__init__(custom_logger=custom_logger)
self._trajectories = trajectories
self.rng = rng
[docs] def sample(self, steps: int) -> Sequence[TrajectoryWithRew]:
# make a copy before shuffling
trajectories = list(self._trajectories)
# NumPy's annotation here is overly-conservative, but this works at runtime
self.rng.shuffle(trajectories) # type: ignore[arg-type]
return _get_trajectories(trajectories, steps)
[docs]class AgentTrainer(TrajectoryGenerator):
"""Wrapper for training an SB3 algorithm on an arbitrary reward function."""
[docs] def __init__(
self,
algorithm: base_class.BaseAlgorithm,
reward_fn: Union[reward_function.RewardFn, reward_nets.RewardNet],
venv: vec_env.VecEnv,
rng: np.random.Generator,
exploration_frac: float = 0.0,
switch_prob: float = 0.5,
random_prob: float = 0.5,
custom_logger: Optional[imit_logger.HierarchicalLogger] = None,
) -> None:
"""Initialize the agent trainer.
Args:
algorithm: the stable-baselines algorithm to use for training.
reward_fn: either a RewardFn or a RewardNet instance that will supply
the rewards used for training the agent.
venv: vectorized environment to train in.
rng: random number generator used for exploration and for sampling.
exploration_frac: fraction of the trajectories that will be generated
partially randomly rather than only by the agent when sampling.
switch_prob: the probability of switching the current policy at each
step for the exploratory samples.
random_prob: the probability of picking the random policy when switching
during exploration.
custom_logger: Where to log to; if None (default), creates a new logger.
"""
self.algorithm = algorithm
# NOTE: this has to come after setting self.algorithm because super().__init__
# will set self.logger, which also sets the logger for the algorithm
super().__init__(custom_logger)
if isinstance(reward_fn, reward_nets.RewardNet):
utils.check_for_correct_spaces(
venv,
reward_fn.observation_space,
reward_fn.action_space,
)
reward_fn = reward_fn.predict_processed
self.reward_fn = reward_fn
self.exploration_frac = exploration_frac
self.rng = rng
# The BufferingWrapper records all trajectories, so we can return
# them after training. This should come first (before the wrapper that
# changes the reward function), so that we return the original environment
# rewards.
# When applying BufferingWrapper and RewardVecEnvWrapper, we should use `venv`
# instead of `algorithm.get_env()` because SB3 may apply some wrappers to
# `algorithm`'s env under the hood. In particular, in image-based environments,
# SB3 may move the image-channel dimension in the observation space, making
# `algorithm.get_env()` not match with `reward_fn`.
self.buffering_wrapper = wrappers.BufferingWrapper(venv)
self.venv = self.reward_venv_wrapper = reward_wrapper.RewardVecEnvWrapper(
self.buffering_wrapper,
reward_fn=self.reward_fn,
)
self.log_callback = self.reward_venv_wrapper.make_log_callback()
self.algorithm.set_env(self.venv)
# Unlike with BufferingWrapper, we should use `algorithm.get_env()` instead
# of `venv` when interacting with `algorithm`.
algo_venv = self.algorithm.get_env()
assert algo_venv is not None
# This wrapper will be used to ensure that rollouts are collected from a mixture
# of `self.algorithm` and a policy that acts randomly. The samples from
# `self.algorithm` are themselves stochastic if `self.algorithm` is stochastic.
# Otherwise, they are deterministic, and action selection is only stochastic
# when sampling from the random policy.
self.exploration_wrapper = exploration_wrapper.ExplorationWrapper(
policy=self.algorithm,
venv=algo_venv,
random_prob=random_prob,
switch_prob=switch_prob,
rng=self.rng,
)
[docs] def train(self, steps: int, **kwargs) -> None:
"""Train the agent using the reward function specified during instantiation.
Args:
steps: number of environment timesteps to train for
**kwargs: other keyword arguments to pass to BaseAlgorithm.train()
Raises:
RuntimeError: Transitions left in `self.buffering_wrapper`; call
`self.sample` first to clear them.
"""
n_transitions = self.buffering_wrapper.n_transitions
if n_transitions:
raise RuntimeError(
f"There are {n_transitions} transitions left in the buffer. "
"Call AgentTrainer.sample() first to clear them.",
)
self.algorithm.learn(
total_timesteps=steps,
reset_num_timesteps=False,
callback=self.log_callback,
**kwargs,
)
[docs] def sample(self, steps: int) -> Sequence[types.TrajectoryWithRew]:
agent_trajs, _ = self.buffering_wrapper.pop_finished_trajectories()
# We typically have more trajectories than are needed.
# In that case, we use the final trajectories because
# they are the ones with the most relevant version of
# the agent.
# The easiest way to do this will be to first invert the
# list and then later just take the first trajectories:
agent_trajs = agent_trajs[::-1]
avail_steps = sum(len(traj) for traj in agent_trajs)
exploration_steps = int(self.exploration_frac * steps)
if self.exploration_frac > 0 and exploration_steps == 0:
self.logger.warn(
"No exploration steps included: exploration_frac = "
f"{self.exploration_frac} > 0 but steps={steps} is too small.",
)
agent_steps = steps - exploration_steps
if avail_steps < agent_steps:
self.logger.log(
f"Requested {agent_steps} transitions but only {avail_steps} in buffer."
f" Sampling {agent_steps - avail_steps} additional transitions.",
)
sample_until = rollout.make_sample_until(
min_timesteps=agent_steps - avail_steps,
min_episodes=None,
)
# Important note: we don't want to use the trajectories returned
# here because 1) they might miss initial timesteps taken by the RL agent
# and 2) their rewards are the ones provided by the reward model!
# Instead, we collect the trajectories using the BufferingWrapper.
algo_venv = self.algorithm.get_env()
assert algo_venv is not None
rollout.generate_trajectories(
self.algorithm,
algo_venv,
sample_until=sample_until,
# By setting deterministic_policy to False, we ensure that the rollouts
# are collected from a deterministic policy only if self.algorithm is
# deterministic. If self.algorithm is stochastic, then policy_callable
# will also be stochastic.
deterministic_policy=False,
rng=self.rng,
)
additional_trajs, _ = self.buffering_wrapper.pop_finished_trajectories()
agent_trajs = list(agent_trajs) + list(additional_trajs)
agent_trajs = _get_trajectories(agent_trajs, agent_steps)
trajectories = list(agent_trajs)
if exploration_steps > 0:
self.logger.log(f"Sampling {exploration_steps} exploratory transitions.")
sample_until = rollout.make_sample_until(
min_timesteps=exploration_steps,
min_episodes=None,
)
algo_venv = self.algorithm.get_env()
assert algo_venv is not None
rollout.generate_trajectories(
policy=self.exploration_wrapper,
venv=algo_venv,
sample_until=sample_until,
# buffering_wrapper collects rollouts from a non-deterministic policy,
# so we do that here as well for consistency.
deterministic_policy=False,
rng=self.rng,
)
exploration_trajs, _ = self.buffering_wrapper.pop_finished_trajectories()
exploration_trajs = _get_trajectories(exploration_trajs, exploration_steps)
# We call _get_trajectories separately on agent_trajs and exploration_trajs
# and then just concatenate. This could mean we return slightly too many
# transitions, but it gets the proportion of exploratory and agent
# transitions roughly right.
trajectories.extend(list(exploration_trajs))
return trajectories
@property
def logger(self) -> imit_logger.HierarchicalLogger:
return super().logger
@logger.setter
def logger(self, value: imit_logger.HierarchicalLogger) -> None:
self._logger = value
self.algorithm.set_logger(self.logger)
def _get_trajectories(
trajectories: Sequence[TrajectoryWithRew],
steps: int,
) -> Sequence[TrajectoryWithRew]:
"""Get enough trajectories to have at least `steps` transitions in total."""
if steps == 0:
return []
available_steps = sum(len(traj) for traj in trajectories)
if available_steps < steps:
raise RuntimeError(
f"Asked for {steps} transitions but only {available_steps} available",
)
# We need the cumulative sum of trajectory lengths
# to determine how many trajectories to return:
steps_cumsum = np.cumsum([len(traj) for traj in trajectories])
# Now we find the first index that gives us enough
# total steps:
idx = int((steps_cumsum >= steps).argmax())
# we need to include the element at position idx
trajectories = trajectories[: idx + 1]
# sanity check
assert sum(len(traj) for traj in trajectories) >= steps
return trajectories
[docs]class PreferenceModel(nn.Module):
"""Class to convert two fragments' rewards into preference probability."""
[docs] def __init__(
self,
model: reward_nets.RewardNet,
noise_prob: float = 0.0,
discount_factor: float = 1.0,
threshold: float = 50,
) -> None:
"""Create Preference Prediction Model.
Args:
model: base model to compute reward.
noise_prob: assumed probability with which the preference
is uniformly random (used for the model of preference generation
that is used for the loss).
discount_factor: the model of preference generation uses a softmax
of returns as the probability that a fragment is preferred.
This is the discount factor used to calculate those returns.
Default is 1, i.e. undiscounted sums of rewards (which is what
the DRLHP paper uses).
threshold: the preference model used to compute the loss contains
a softmax of returns. To avoid overflows, we clip differences
in returns that are above this threshold. This threshold
is therefore in logspace. The default value of 50 means
that probabilities below 2e-22 are rounded up to 2e-22.
Raises:
ValueError: if `RewardEnsemble` is wrapped around a class
other than `AddSTDRewardWrapper`.
"""
super().__init__()
self.model = model
self.noise_prob = noise_prob
self.discount_factor = discount_factor
self.threshold = threshold
base_model = get_base_model(model)
self.ensemble_model = None
# if the base model is an ensemble model, then keep the base model as
# model to get rewards from all networks
if isinstance(base_model, reward_nets.RewardEnsemble):
# reward_model may include an AddSTDRewardWrapper for RL training; but we
# must train directly on the base model for reward model training.
is_base = model is base_model
is_std_wrapper = (
isinstance(model, reward_nets.AddSTDRewardWrapper)
and model.base is base_model
)
if not (is_base or is_std_wrapper):
raise ValueError(
"RewardEnsemble can only be wrapped"
f" by AddSTDRewardWrapper but found {type(model).__name__}.",
)
self.ensemble_model = base_model
self.member_pref_models = []
for member in self.ensemble_model.members:
member_pref_model = PreferenceModel(
cast(reward_nets.RewardNet, member), # nn.ModuleList is not generic
self.noise_prob,
self.discount_factor,
self.threshold,
)
self.member_pref_models.append(member_pref_model)
[docs] def forward(
self,
fragment_pairs: Sequence[TrajectoryPair],
) -> Tuple[th.Tensor, Optional[th.Tensor]]:
"""Computes the preference probability of the first fragment for all pairs.
Note: This function passes the gradient through for non-ensemble models.
For an ensemble model, this function should not be used for loss
calculation. It can be used in case where passing the gradient is not
required such as during active selection or inference time.
Therefore, the EnsembleTrainer passes each member network through this
function instead of passing the EnsembleNetwork object with the use of
`ensemble_member_index`.
Args:
fragment_pairs: batch of pair of fragments.
Returns:
A tuple with the first element as the preference probabilities for the
first fragment for all fragment pairs given by the network(s).
If the ground truth rewards are available, it also returns gt preference
probabilities in the second element of the tuple (else None).
Reward probability shape - (num_fragment_pairs, ) for non-ensemble reward
network and (num_fragment_pairs, num_networks) for an ensemble of networks.
"""
probs = th.empty(len(fragment_pairs), dtype=th.float32)
gt_reward_available = _trajectory_pair_includes_reward(fragment_pairs[0])
if gt_reward_available:
gt_probs = th.empty(len(fragment_pairs), dtype=th.float32)
for i, fragment in enumerate(fragment_pairs):
frag1, frag2 = fragment
trans1 = rollout.flatten_trajectories([frag1])
trans2 = rollout.flatten_trajectories([frag2])
rews1 = self.rewards(trans1)
rews2 = self.rewards(trans2)
probs[i] = self.probability(rews1, rews2)
if gt_reward_available:
frag1 = cast(TrajectoryWithRew, frag1)
frag2 = cast(TrajectoryWithRew, frag2)
gt_rews_1 = th.from_numpy(frag1.rews)
gt_rews_2 = th.from_numpy(frag2.rews)
gt_probs[i] = self.probability(gt_rews_1, gt_rews_2)
return probs, (gt_probs if gt_reward_available else None)
[docs] def rewards(self, transitions: Transitions) -> th.Tensor:
"""Computes the reward for all transitions.
Args:
transitions: batch of obs-act-obs-done for a fragment of a trajectory.
Returns:
The reward given by the network(s) for all the transitions.
Shape - (num_transitions, ) for Single reward network and
(num_transitions, num_networks) for ensemble of networks.
"""
state = types.assert_not_dictobs(transitions.obs)
action = transitions.acts
next_state = types.assert_not_dictobs(transitions.next_obs)
done = transitions.dones
if self.ensemble_model is not None:
rews_np = self.ensemble_model.predict_processed_all(
state,
action,
next_state,
done,
)
assert rews_np.shape == (len(state), self.ensemble_model.num_members)
rews = util.safe_to_tensor(rews_np).to(self.ensemble_model.device)
else:
preprocessed = self.model.preprocess(state, action, next_state, done)
rews = self.model(*preprocessed)
assert rews.shape == (len(state),)
return rews
[docs] def probability(self, rews1: th.Tensor, rews2: th.Tensor) -> th.Tensor:
"""Computes the Boltzmann rational probability the first trajectory is best.
Args:
rews1: array/matrix of rewards for the first trajectory fragment.
matrix for ensemble models and array for non-ensemble models.
rews2: array/matrix of rewards for the second trajectory fragment.
matrix for ensemble models and array for non-ensemble models.
Returns:
The softmax of the difference between the (discounted) return of the
first and second trajectory.
Shape - (num_ensemble_members, ) for ensemble model and
() for non-ensemble model which is a torch scalar.
"""
# check rews has correct shape based on the model
expected_dims = 2 if self.ensemble_model is not None else 1
assert rews1.ndim == rews2.ndim == expected_dims
# First, we compute the difference of the returns of
# the two fragments. We have a special case for a discount
# factor of 1 to avoid unnecessary computation (especially
# since this is the default setting).
if self.discount_factor == 1:
returns_diff = (rews2 - rews1).sum(axis=0) # type: ignore[call-overload]
else:
device = rews1.device
assert device == rews2.device
discounts = self.discount_factor ** th.arange(len(rews1), device=device)
if self.ensemble_model is not None:
discounts = discounts.reshape(-1, 1)
returns_diff = (discounts * (rews2 - rews1)).sum(axis=0)
# Clip to avoid overflows (which in particular may occur
# in the backwards pass even if they do not in the forward pass).
returns_diff = th.clip(returns_diff, -self.threshold, self.threshold)
# We take the softmax of the returns. model_probability
# is the first dimension of that softmax, representing the
# probability that fragment 1 is preferred.
model_probability = 1 / (1 + returns_diff.exp())
probability = self.noise_prob * 0.5 + (1 - self.noise_prob) * model_probability
if self.ensemble_model is not None:
assert probability.shape == (self.model.num_members,)
else:
assert probability.shape == ()
return probability
[docs]class Fragmenter(abc.ABC):
"""Class for creating pairs of trajectory fragments from a set of trajectories."""
[docs] def __init__(self, custom_logger: Optional[imit_logger.HierarchicalLogger] = None):
"""Initialize the fragmenter.
Args:
custom_logger: Where to log to; if None (default), creates a new logger.
"""
self.logger = custom_logger or imit_logger.configure()
@abc.abstractmethod
def __call__(
self,
trajectories: Sequence[TrajectoryWithRew],
fragment_length: int,
num_pairs: int,
) -> Sequence[TrajectoryWithRewPair]:
"""Create fragment pairs out of a sequence of trajectories.
Args:
trajectories: collection of trajectories that will be split up into
fragments
fragment_length: the length of each sampled fragment
num_pairs: the number of fragment pairs to sample
Returns:
a sequence of fragment pairs
""" # noqa: DAR202
[docs]class RandomFragmenter(Fragmenter):
"""Sample fragments of trajectories uniformly at random with replacement.
Note that each fragment is part of a single episode and has a fixed
length. This leads to a bias: transitions at the beginning and at the
end of episodes are less likely to occur as part of fragments (this affects
the first and last fragment_length transitions).
An additional bias is that trajectories shorter than the desired fragment
length are never used.
"""
[docs] def __init__(
self,
rng: np.random.Generator,
warning_threshold: int = 10,
custom_logger: Optional[imit_logger.HierarchicalLogger] = None,
) -> None:
"""Initialize the fragmenter.
Args:
rng: the random number generator
warning_threshold: give a warning if the number of available
transitions is less than this many times the number of
required samples. Set to 0 to disable this warning.
custom_logger: Where to log to; if None (default), creates a new logger.
"""
super().__init__(custom_logger)
self.rng = rng
self.warning_threshold = warning_threshold
def __call__(
self,
trajectories: Sequence[TrajectoryWithRew],
fragment_length: int,
num_pairs: int,
) -> Sequence[TrajectoryWithRewPair]:
fragments: List[TrajectoryWithRew] = []
prev_num_trajectories = len(trajectories)
# filter out all trajectories that are too short
trajectories = [traj for traj in trajectories if len(traj) >= fragment_length]
if len(trajectories) == 0:
raise ValueError(
"No trajectories are long enough for the desired fragment length "
f"of {fragment_length}.",
)
num_discarded = prev_num_trajectories - len(trajectories)
if num_discarded:
self.logger.log(
f"Discarded {num_discarded} out of {prev_num_trajectories} "
"trajectories because they are shorter than the desired length "
f"of {fragment_length}.",
)
weights = [len(traj) for traj in trajectories]
# number of transitions that will be contained in the fragments
num_transitions = 2 * num_pairs * fragment_length
if sum(weights) < num_transitions:
self.logger.warn(
"Fewer transitions available than needed for desired number "
"of fragment pairs. Some transitions will appear multiple times.",
)
elif (
self.warning_threshold
and sum(weights) < self.warning_threshold * num_transitions
):
# If the number of available transitions is not much larger
# than the number of requires ones, we already give a warning.
# But only if self.warning_threshold is non-zero.
self.logger.warn(
f"Samples will contain {num_transitions} transitions in total "
f"and only {sum(weights)} are available. "
f"Because we sample with replacement, a significant number "
"of transitions are likely to appear multiple times.",
)
# we need two fragments for each comparison
for _ in range(2 * num_pairs):
# NumPy's annotation here is overly-conservative, but this works at runtime
traj = self.rng.choice(
trajectories, # type: ignore[arg-type]
p=np.array(weights) / sum(weights),
)
n = len(traj)
start = self.rng.integers(0, n - fragment_length, endpoint=True)
end = start + fragment_length
terminal = (end == n) and traj.terminal
fragment = TrajectoryWithRew(
obs=traj.obs[start : end + 1],
acts=traj.acts[start:end],
infos=traj.infos[start:end] if traj.infos is not None else None,
rews=traj.rews[start:end],
terminal=terminal,
)
fragments.append(fragment)
# fragments is currently a list of single fragments. We want to pair up
# fragments to get a list of (fragment1, fragment2) tuples. To do so,
# we create a single iterator of the list and zip it with itself:
iterator = iter(fragments)
return list(zip(iterator, iterator))
[docs]class ActiveSelectionFragmenter(Fragmenter):
"""Sample fragments of trajectories based on active selection.
Actively picks the fragment pairs with the highest uncertainty (variance)
of rewards/probabilties/predictions from ensemble model.
"""
[docs] def __init__(
self,
preference_model: PreferenceModel,
base_fragmenter: Fragmenter,
fragment_sample_factor: float,
uncertainty_on: str = "logit",
custom_logger: Optional[imit_logger.HierarchicalLogger] = None,
) -> None:
"""Initialize the active selection fragmenter.
Args:
preference_model: an ensemble model that predicts the
preference of the first fragment over the other.
base_fragmenter: fragmenter instance to get
fragment pairs from trajectories
fragment_sample_factor: the factor of the number of
fragment pairs to sample from the base_fragmenter
uncertainty_on: the variable to calculate the variance on.
Can be logit|probability|label.
custom_logger: Where to log to; if None (default), creates a new logger.
Raises:
ValueError: Preference model not wrapped over an ensemble of networks.
"""
super().__init__(custom_logger=custom_logger)
if preference_model.ensemble_model is None:
raise ValueError(
"PreferenceModel not wrapped over an ensemble of networks.",
)
self.preference_model = preference_model
self.base_fragmenter = base_fragmenter
self.fragment_sample_factor = fragment_sample_factor
self._uncertainty_on = uncertainty_on
if not (uncertainty_on in ["logit", "probability", "label"]):
self.raise_uncertainty_on_not_supported()
@property
def uncertainty_on(self) -> str:
return self._uncertainty_on
[docs] def raise_uncertainty_on_not_supported(self) -> NoReturn:
raise ValueError(
f"""{self.uncertainty_on} not supported.
`uncertainty_on` should be from `logit`, `probability`, or `label`""",
)
def __call__(
self,
trajectories: Sequence[TrajectoryWithRew],
fragment_length: int,
num_pairs: int,
) -> Sequence[TrajectoryWithRewPair]:
# sample a large number (self.fragment_sample_factor*num_pairs)
# of fragments from all the trajectories
fragments_to_sample = int(self.fragment_sample_factor * num_pairs)
fragment_pairs = self.base_fragmenter(
trajectories=trajectories,
fragment_length=fragment_length,
num_pairs=fragments_to_sample,
)
var_estimates = np.zeros(len(fragment_pairs))
for i, fragment in enumerate(fragment_pairs):
frag1, frag2 = fragment
trans1 = rollout.flatten_trajectories([frag1])
trans2 = rollout.flatten_trajectories([frag2])
with th.no_grad():
rews1 = self.preference_model.rewards(trans1)
rews2 = self.preference_model.rewards(trans2)
var_estimate = self.variance_estimate(rews1, rews2)
var_estimates[i] = var_estimate
fragment_idxs = np.argsort(var_estimates)[::-1] # sort in descending order
# return fragment pairs that have the highest uncertainty
return [fragment_pairs[idx] for idx in fragment_idxs[:num_pairs]]
[docs] def variance_estimate(self, rews1: th.Tensor, rews2: th.Tensor) -> float:
"""Gets the variance estimate from the rewards of a fragment pair.
Args:
rews1: rewards obtained by all the ensemble models for the first fragment.
Shape - (fragment_length, num_ensemble_members)
rews2: rewards obtained by all the ensemble models for the second fragment.
Shape - (fragment_length, num_ensemble_members)
Returns:
the variance estimate based on the `uncertainty_on` flag.
"""
if self.uncertainty_on == "logit":
returns1, returns2 = rews1.sum(0), rews2.sum(0)
var_estimate = (returns1 - returns2).var().item()
else: # uncertainty_on is probability or label
probs = self.preference_model.probability(rews1, rews2)
probs_np = probs.cpu().numpy()
assert probs_np.shape == (self.preference_model.model.num_members,)
if self.uncertainty_on == "probability":
var_estimate = probs_np.var()
elif self.uncertainty_on == "label": # uncertainty_on is label
preds = (probs_np > 0.5).astype(np.float32)
# probability estimate of Bernoulli random variable
prob_estimate = preds.mean()
# variance estimate of Bernoulli random variable
var_estimate = prob_estimate * (1 - prob_estimate)
else:
self.raise_uncertainty_on_not_supported()
return var_estimate
[docs]class PreferenceGatherer(abc.ABC):
"""Base class for gathering preference comparisons between trajectory fragments."""
[docs] def __init__(
self,
rng: Optional[np.random.Generator] = None,
custom_logger: Optional[imit_logger.HierarchicalLogger] = None,
) -> None:
"""Initializes the preference gatherer.
Args:
rng: random number generator, if applicable.
custom_logger: Where to log to; if None (default), creates a new logger.
"""
# The random seed isn't used here, but it's useful to have this
# as an argument nevertheless because that means we can always
# pass in a seed in training scripts (without worrying about whether
# the PreferenceGatherer we use needs one).
del rng
self.logger = custom_logger or imit_logger.configure()
@abc.abstractmethod
def __call__(self, fragment_pairs: Sequence[TrajectoryWithRewPair]) -> np.ndarray:
"""Gathers the probabilities that fragment 1 is preferred in `fragment_pairs`.
Args:
fragment_pairs: sequence of pairs of trajectory fragments
Returns:
A numpy array with shape (b, ), where b is the length of the input
(i.e. batch size). Each item in the array is the probability that
fragment 1 is preferred over fragment 2 for the corresponding
pair of fragments.
Note that for human feedback, these probabilities are simply 0 or 1
(or 0.5 in case of indifference), but synthetic models may yield other
probabilities.
""" # noqa: DAR202
[docs]class SyntheticGatherer(PreferenceGatherer):
"""Computes synthetic preferences using ground-truth environment rewards."""
[docs] def __init__(
self,
temperature: float = 1,
discount_factor: float = 1,
sample: bool = True,
rng: Optional[np.random.Generator] = None,
threshold: float = 50,
custom_logger: Optional[imit_logger.HierarchicalLogger] = None,
) -> None:
"""Initialize the synthetic preference gatherer.
Args:
temperature: the preferences are sampled from a softmax, this is
the temperature used for sampling. temperature=0 leads to deterministic
results (for equal rewards, 0.5 will be returned).
discount_factor: discount factor that is used to compute
how good a fragment is. Default is to use undiscounted
sums of rewards (as in the DRLHP paper).
sample: if True (default), the preferences are 0 or 1, sampled from
a Bernoulli distribution (or 0.5 in the case of ties with zero
temperature). If False, then the underlying Bernoulli probabilities
are returned instead.
rng: random number generator, only used if
``temperature > 0`` and ``sample=True``
threshold: preferences are sampled from a softmax of returns.
To avoid overflows, we clip differences in returns that are
above this threshold (after multiplying with temperature).
This threshold is therefore in logspace. The default value
of 50 means that probabilities below 2e-22 are rounded up to 2e-22.
custom_logger: Where to log to; if None (default), creates a new logger.
Raises:
ValueError: if `sample` is true and no random state is provided.
"""
super().__init__(custom_logger=custom_logger)
self.temperature = temperature
self.discount_factor = discount_factor
self.sample = sample
self.rng = rng
self.threshold = threshold
if self.sample and self.rng is None:
raise ValueError("If `sample` is True, then `rng` must be provided.")
def __call__(self, fragment_pairs: Sequence[TrajectoryWithRewPair]) -> np.ndarray:
"""Computes probability fragment 1 is preferred over fragment 2."""
returns1, returns2 = self._reward_sums(fragment_pairs)
if self.temperature == 0:
return (np.sign(returns1 - returns2) + 1) / 2
returns1 /= self.temperature
returns2 /= self.temperature
# clip the returns to avoid overflows in the softmax below
returns_diff = np.clip(returns2 - returns1, -self.threshold, self.threshold)
# Instead of computing exp(rews1) / (exp(rews1) + exp(rews2)) directly,
# we divide enumerator and denominator by exp(rews1) to prevent overflows:
model_probs = 1 / (1 + np.exp(returns_diff))
# Compute the mean binary entropy. This metric helps estimate
# how good we can expect the performance of the learned reward
# model to be at predicting preferences.
entropy = -(
special.xlogy(model_probs, model_probs)
+ special.xlogy(1 - model_probs, 1 - model_probs)
).mean()
self.logger.record("entropy", entropy)
if self.sample:
assert self.rng is not None
return self.rng.binomial(n=1, p=model_probs).astype(np.float32)
return model_probs
def _reward_sums(self, fragment_pairs) -> Tuple[np.ndarray, np.ndarray]:
rews1, rews2 = zip(
*[
(
rollout.discounted_sum(f1.rews, self.discount_factor),
rollout.discounted_sum(f2.rews, self.discount_factor),
)
for f1, f2 in fragment_pairs
],
)
return np.array(rews1, dtype=np.float32), np.array(rews2, dtype=np.float32)
[docs]class PreferenceDataset(data_th.Dataset):
"""A PyTorch Dataset for preference comparisons.
Each item is a tuple consisting of two trajectory fragments
and a probability that fragment 1 is preferred over fragment 2.
This dataset is meant to be generated piece by piece during the
training process, which is why data can be added via the .push()
method.
"""
[docs] def __init__(self, max_size: Optional[int] = None) -> None:
"""Builds an empty PreferenceDataset.
Args:
max_size: Maximum number of preference comparisons to store in the dataset.
If None (default), the dataset can grow indefinitely. Otherwise, the
dataset acts as a FIFO queue, and the oldest comparisons are evicted
when `push()` is called and the dataset is at max capacity.
"""
self.fragments1: List[TrajectoryWithRew] = []
self.fragments2: List[TrajectoryWithRew] = []
self.max_size = max_size
self.preferences: np.ndarray = np.array([])
[docs] def push(
self,
fragments: Sequence[TrajectoryWithRewPair],
preferences: np.ndarray,
) -> None:
"""Add more samples to the dataset.
Args:
fragments: list of pairs of trajectory fragments to add
preferences: corresponding preference probabilities (probability
that fragment 1 is preferred over fragment 2)
Raises:
ValueError: `preferences` shape does not match `fragments` or
has non-float32 dtype.
"""
fragments1, fragments2 = zip(*fragments)
if preferences.shape != (len(fragments),):
raise ValueError(
f"Unexpected preferences shape {preferences.shape}, "
f"expected {(len(fragments),)}",
)
if preferences.dtype != np.float32:
raise ValueError("preferences should have dtype float32")
self.fragments1.extend(fragments1)
self.fragments2.extend(fragments2)
self.preferences = np.concatenate((self.preferences, preferences))
# Evict old samples if the dataset is at max capacity
if self.max_size is not None:
extra = len(self.preferences) - self.max_size
if extra > 0:
self.fragments1 = self.fragments1[extra:]
self.fragments2 = self.fragments2[extra:]
self.preferences = self.preferences[extra:]
@overload
def __getitem__(self, key: int) -> Tuple[TrajectoryWithRewPair, float]:
pass
@overload
def __getitem__(
self,
key: slice,
) -> Tuple[types.Pair[Sequence[TrajectoryWithRew]], Sequence[float]]:
pass
def __getitem__(self, key):
return (self.fragments1[key], self.fragments2[key]), self.preferences[key]
def __len__(self) -> int:
assert len(self.fragments1) == len(self.fragments2) == len(self.preferences)
return len(self.fragments1)
[docs] def save(self, path: AnyPath) -> None:
with open(path, "wb") as file:
pickle.dump(self, file)
[docs] @staticmethod
def load(path: AnyPath) -> "PreferenceDataset":
with open(path, "rb") as file:
return pickle.load(file)
[docs]def preference_collate_fn(
batch: Sequence[Tuple[TrajectoryWithRewPair, float]],
) -> Tuple[Sequence[TrajectoryWithRewPair], np.ndarray]:
fragment_pairs, preferences = zip(*batch)
return list(fragment_pairs), np.array(preferences)
[docs]class LossAndMetrics(NamedTuple):
"""Loss and auxiliary metrics for reward network training."""
loss: th.Tensor
metrics: Mapping[str, th.Tensor]
[docs]class RewardLoss(nn.Module, abc.ABC):
"""A loss function over preferences."""
[docs] @abc.abstractmethod
def forward(
self,
fragment_pairs: Sequence[TrajectoryPair],
preferences: np.ndarray,
preference_model: PreferenceModel,
) -> LossAndMetrics:
"""Computes the loss.
Args:
fragment_pairs: Batch consisting of pairs of trajectory fragments.
preferences: The probability that the first fragment is preferred
over the second. Typically 0, 1 or 0.5 (tie).
preference_model: model to predict the preferred fragment from a pair.
Returns: # noqa: DAR202
loss: the loss
metrics: a dictionary of metrics that can be logged
"""
def _trajectory_pair_includes_reward(fragment_pair: TrajectoryPair) -> bool:
"""Return true if and only if both fragments in the pair include rewards."""
frag1, frag2 = fragment_pair
return isinstance(frag1, TrajectoryWithRew) and isinstance(frag2, TrajectoryWithRew)
[docs]class CrossEntropyRewardLoss(RewardLoss):
"""Compute the cross entropy reward loss."""
[docs] def __init__(self) -> None:
"""Create cross entropy reward loss."""
super().__init__()
[docs] def forward(
self,
fragment_pairs: Sequence[TrajectoryPair],
preferences: np.ndarray,
preference_model: PreferenceModel,
) -> LossAndMetrics:
"""Computes the loss.
Args:
fragment_pairs: Batch consisting of pairs of trajectory fragments.
preferences: The probability that the first fragment is preferred
over the second. Typically 0, 1 or 0.5 (tie).
preference_model: model to predict the preferred fragment from a pair.
Returns:
The cross-entropy loss between the probability predicted by the
reward model and the target probabilities in `preferences`. Metrics
are accuracy, and gt_reward_loss, if the ground truth reward is
available.
"""
probs, gt_probs = preference_model(fragment_pairs)
# TODO(ejnnr): Here and below, > 0.5 is problematic
# because getting exactly 0.5 is actually somewhat
# common in some environments (as long as sample=False or temperature=0).
# In a sense that "only" creates class imbalance
# but it's still misleading.
predictions = probs > 0.5
preferences_th = th.as_tensor(preferences, dtype=th.float32)
ground_truth = preferences_th > 0.5
metrics = {}
metrics["accuracy"] = (predictions == ground_truth).float().mean()
if gt_probs is not None:
metrics["gt_reward_loss"] = th.nn.functional.binary_cross_entropy(
gt_probs,
preferences_th,
)
metrics = {key: value.detach().cpu() for key, value in metrics.items()}
return LossAndMetrics(
loss=th.nn.functional.binary_cross_entropy(probs, preferences_th),
metrics=metrics,
)
[docs]class RewardTrainer(abc.ABC):
"""Abstract base class for training reward models using preference comparisons.
This class contains only the actual reward model training code,
it is not responsible for gathering trajectories and preferences
or for agent training (see :class: `PreferenceComparisons` for that).
"""
[docs] def __init__(
self,
preference_model: PreferenceModel,
custom_logger: Optional[imit_logger.HierarchicalLogger] = None,
) -> None:
"""Initialize the reward trainer.
Args:
preference_model: the preference model to train the reward network.
custom_logger: Where to log to; if None (default), creates a new logger.
"""
self._preference_model = preference_model
self._logger = custom_logger or imit_logger.configure()
@property
def logger(self) -> imit_logger.HierarchicalLogger:
return self._logger
@logger.setter
def logger(self, custom_logger: imit_logger.HierarchicalLogger) -> None:
self._logger = custom_logger
[docs] def train(self, dataset: PreferenceDataset, epoch_multiplier: float = 1.0) -> None:
"""Train the reward model on a batch of fragment pairs and preferences.
Args:
dataset: the dataset of preference comparisons to train on.
epoch_multiplier: how much longer to train for than usual
(measured relatively).
"""
with networks.training(self._preference_model.model):
self._train(dataset, epoch_multiplier)
@abc.abstractmethod
def _train(self, dataset: PreferenceDataset, epoch_multiplier: float) -> None:
"""Train the reward model; see ``train`` for details."""
[docs]class BasicRewardTrainer(RewardTrainer):
"""Train a basic reward model."""
regularizer: Optional[regularizers.Regularizer]
[docs] def __init__(
self,
preference_model: PreferenceModel,
loss: RewardLoss,
rng: np.random.Generator,
batch_size: int = 32,
minibatch_size: Optional[int] = None,
epochs: int = 1,
lr: float = 1e-3,
custom_logger: Optional[imit_logger.HierarchicalLogger] = None,
regularizer_factory: Optional[regularizers.RegularizerFactory] = None,
) -> None:
"""Initialize the reward model trainer.
Args:
preference_model: the preference model to train the reward network.
loss: the loss to use
rng: the random number generator to use for splitting the dataset into
training and validation.
batch_size: number of fragment pairs per batch
minibatch_size: size of minibatch to calculate gradients over.
The gradients are accumulated until `batch_size` examples
are processed before making an optimization step. This
is useful in GPU training to reduce memory usage, since
fewer examples are loaded into memory at once,
facilitating training with larger batch sizes, but is
generally slower. Must be a factor of `batch_size`.
Optional, defaults to `batch_size`.
epochs: number of epochs in each training iteration (can be adjusted
on the fly by specifying an `epoch_multiplier` in `self.train()`
if longer training is desired in specific cases).
lr: the learning rate
custom_logger: Where to log to; if None (default), creates a new logger.
regularizer_factory: if you would like to apply regularization during
training, specify a regularizer factory here. The factory will be
used to construct a regularizer. See
``imitation.regularization.RegularizerFactory`` for more details.
Raises:
ValueError: if the batch size is not a multiple of the minibatch size.
"""
super().__init__(preference_model, custom_logger)
self.loss = loss
self.batch_size = batch_size
self.minibatch_size = minibatch_size or batch_size
if self.batch_size % self.minibatch_size != 0:
raise ValueError("Batch size must be a multiple of minibatch size.")
self.epochs = epochs
self.optim = th.optim.AdamW(self._preference_model.parameters(), lr=lr)
self.rng = rng
self.regularizer = (
regularizer_factory(optimizer=self.optim, logger=self.logger)
if regularizer_factory is not None
else None
)
def _make_data_loader(self, dataset: data_th.Dataset) -> data_th.DataLoader:
"""Make a dataloader."""
return data_th.DataLoader(
dataset,
batch_size=self.minibatch_size,
shuffle=True,
collate_fn=preference_collate_fn,
)
@property
def requires_regularizer_update(self) -> bool:
"""Whether the regularizer requires updating.
Returns:
If true, this means that a validation dataset will be used.
"""
return self.regularizer is not None and self.regularizer.val_split is not None
def _train(
self,
dataset: PreferenceDataset,
epoch_multiplier: float = 1.0,
) -> None:
"""Trains for `epoch_multiplier * self.epochs` epochs over `dataset`."""
if self.regularizer is not None and self.regularizer.val_split is not None:
val_length = int(len(dataset) * self.regularizer.val_split)
train_length = len(dataset) - val_length
if val_length < 1 or train_length < 1:
raise ValueError(
"Not enough data samples to split into training and validation, "
"or the validation split is too large/small. "
"Make sure you've generated enough initial preference data. "
"You can adjust this through initial_comparison_frac in "
"PreferenceComparisons.",
)
train_dataset, val_dataset = data_th.random_split(
dataset,
lengths=[train_length, val_length],
# we convert the numpy generator to the pytorch generator.
generator=th.Generator().manual_seed(util.make_seeds(self.rng)),
)
dataloader = self._make_data_loader(train_dataset)
val_dataloader = self._make_data_loader(val_dataset)
else:
dataloader = self._make_data_loader(dataset)
val_dataloader = None
epochs = round(self.epochs * epoch_multiplier)
assert epochs > 0, "Must train for at least one epoch."
with self.logger.accumulate_means("reward"):
for epoch_num in tqdm(range(epochs), desc="Training reward model"):
with self.logger.add_key_prefix(f"epoch-{epoch_num}"):
train_loss = 0.0
accumulated_size = 0
self.optim.zero_grad()
for fragment_pairs, preferences in dataloader:
with self.logger.add_key_prefix("train"):
loss = self._training_inner_loop(
fragment_pairs,
preferences,
)
# Renormalise the loss to be averaged over
# the whole batch size instead of the
# minibatch size. If there is an incomplete
# batch, its gradients will be smaller,
# which may be helpful for stability.
loss *= len(fragment_pairs) / self.batch_size
train_loss += loss.item()
if self.regularizer:
self.regularizer.regularize_and_backward(loss)
else:
loss.backward()
accumulated_size += len(fragment_pairs)
if accumulated_size >= self.batch_size:
self.optim.step()
self.optim.zero_grad()
accumulated_size = 0
if accumulated_size != 0:
self.optim.step() # if there remains an incomplete batch
if not self.requires_regularizer_update:
continue
assert val_dataloader is not None
assert self.regularizer is not None
val_loss = 0.0
for fragment_pairs, preferences in val_dataloader:
with self.logger.add_key_prefix("val"):
val_loss += self._training_inner_loop(
fragment_pairs,
preferences,
).item()
self.regularizer.update_params(train_loss, val_loss)
# after training all the epochs,
# record also the final value in a separate key for easy access.
keys = list(self.logger.name_to_value.keys())
outer_prefix = self.logger.get_accumulate_prefixes()
for key in keys:
base_path = f"{outer_prefix}reward/" # existing prefix + accum_means ctx
epoch_path = f"mean/{base_path}epoch-{epoch_num}/" # mean for last epoch
final_path = f"{base_path}final/" # path to record last epoch
pattern = rf"{epoch_path}(.+)"
if regex_match := re.match(pattern, key):
(key_name,) = regex_match.groups()
val = self.logger.name_to_value[key]
new_key = f"{final_path}{key_name}"
self.logger.record(new_key, val)
def _training_inner_loop(
self,
fragment_pairs: Sequence[TrajectoryPair],
preferences: np.ndarray,
) -> th.Tensor:
output = self.loss.forward(fragment_pairs, preferences, self._preference_model)
loss = output.loss
self.logger.record("loss", loss.item())
for name, value in output.metrics.items():
self.logger.record(name, value.item())
return loss
[docs]class EnsembleTrainer(BasicRewardTrainer):
"""Train a reward ensemble."""
[docs] def __init__(
self,
preference_model: PreferenceModel,
loss: RewardLoss,
rng: np.random.Generator,
batch_size: int = 32,
minibatch_size: Optional[int] = None,
epochs: int = 1,
lr: float = 1e-3,
custom_logger: Optional[imit_logger.HierarchicalLogger] = None,
regularizer_factory: Optional[regularizers.RegularizerFactory] = None,
) -> None:
"""Initialize the reward model trainer.
Args:
preference_model: the preference model to train the reward network.
loss: the loss to use
rng: random state for the internal RNG used in bagging
batch_size: number of fragment pairs per batch
minibatch_size: size of minibatch to calculate gradients over.
The gradients are accumulated until `batch_size` examples
are processed before making an optimization step. This
is useful in GPU training to reduce memory usage, since
fewer examples are loaded into memory at once,
facilitating training with larger batch sizes, but is
generally slower. Must be a factor of `batch_size`.
Optional, defaults to `batch_size`.
epochs: number of epochs in each training iteration (can be adjusted
on the fly by specifying an `epoch_multiplier` in `self.train()`
if longer training is desired in specific cases).
lr: the learning rate
custom_logger: Where to log to; if None (default), creates a new logger.
regularizer_factory: A factory for creating a regularizer. If None,
no regularization is used.
Raises:
TypeError: if model is not a RewardEnsemble.
"""
if preference_model.ensemble_model is None:
raise TypeError(
"PreferenceModel of a RewardEnsemble expected by EnsembleTrainer.",
)
super().__init__(
preference_model,
loss=loss,
batch_size=batch_size,
minibatch_size=minibatch_size,
epochs=epochs,
lr=lr,
custom_logger=custom_logger,
rng=rng,
regularizer_factory=regularizer_factory,
)
self.member_trainers = []
for member_pref_model in self._preference_model.member_pref_models:
reward_trainer = BasicRewardTrainer(
member_pref_model,
loss=loss,
batch_size=batch_size,
minibatch_size=minibatch_size,
epochs=epochs,
lr=lr,
custom_logger=self.logger,
regularizer_factory=regularizer_factory,
rng=self.rng,
)
self.member_trainers.append(reward_trainer)
@property
def logger(self) -> imit_logger.HierarchicalLogger:
return super().logger
@logger.setter
def logger(self, custom_logger: imit_logger.HierarchicalLogger) -> None:
self._logger = custom_logger
for member_trainer in self.member_trainers:
member_trainer.logger = custom_logger
def _train(self, dataset: PreferenceDataset, epoch_multiplier: float = 1.0) -> None:
"""Trains for `epoch_multiplier * self.epochs` epochs over `dataset`."""
sampler = data_th.RandomSampler(
dataset,
replacement=True,
num_samples=len(dataset),
# we convert the numpy generator to the pytorch generator.
generator=th.Generator().manual_seed(util.make_seeds(self.rng)),
)
for member_idx in range(len(self.member_trainers)):
# sampler gives new indexes on every call
bagging_dataset = data_th.Subset(dataset, list(sampler))
with self.logger.add_accumulate_prefix(f"member-{member_idx}"):
self.member_trainers[member_idx].train(
bagging_dataset,
epoch_multiplier=epoch_multiplier,
)
# average the metrics across the member models
metrics = defaultdict(list)
keys = list(self.logger.name_to_value.keys())
for key in keys:
if re.match(r"member-(\d+)/reward/(.+)", key) and "final" in key:
val = self.logger.name_to_value[key]
key_list = key.split("/")
key_list.pop(0)
metrics["/".join(key_list)].append(val)
for k, v in metrics.items():
self.logger.record(k, np.mean(v))
self.logger.record(k + "_std", np.std(v))
[docs]def get_base_model(reward_model: reward_nets.RewardNet) -> reward_nets.RewardNet:
base_model = reward_model
while hasattr(base_model, "base"):
base_model = cast(reward_nets.RewardNet, base_model.base)
return base_model
def _make_reward_trainer(
preference_model: PreferenceModel,
loss: RewardLoss,
rng: np.random.Generator,
reward_trainer_kwargs: Optional[Mapping[str, Any]] = None,
) -> RewardTrainer:
"""Construct the correct type of reward trainer for this reward function."""
if reward_trainer_kwargs is None:
reward_trainer_kwargs = {}
if preference_model.ensemble_model is not None:
return EnsembleTrainer(
preference_model,
loss,
rng=rng,
**reward_trainer_kwargs,
)
else:
return BasicRewardTrainer(
preference_model,
loss=loss,
rng=rng,
**reward_trainer_kwargs,
)
QUERY_SCHEDULES: Dict[str, type_aliases.Schedule] = {
"constant": lambda t: 1.0,
"hyperbolic": lambda t: 1.0 / (1.0 + t),
"inverse_quadratic": lambda t: 1.0 / (1.0 + t**2),
}
[docs]class PreferenceComparisons(base.BaseImitationAlgorithm):
"""Main interface for reward learning using preference comparisons."""
[docs] def __init__(
self,
trajectory_generator: TrajectoryGenerator,
reward_model: reward_nets.RewardNet,
num_iterations: int,
fragmenter: Optional[Fragmenter] = None,
preference_gatherer: Optional[PreferenceGatherer] = None,
reward_trainer: Optional[RewardTrainer] = None,
comparison_queue_size: Optional[int] = None,
fragment_length: int = 100,
transition_oversampling: float = 1,
initial_comparison_frac: float = 0.1,
initial_epoch_multiplier: float = 200.0,
custom_logger: Optional[imit_logger.HierarchicalLogger] = None,
allow_variable_horizon: bool = False,
rng: Optional[np.random.Generator] = None,
query_schedule: Union[str, type_aliases.Schedule] = "hyperbolic",
) -> None:
"""Initialize the preference comparison trainer.
The loggers of all subcomponents are overridden with the logger used
by this class.
Args:
trajectory_generator: generates trajectories while optionally training
an RL agent on the learned reward function (can also be a sampler
from a static dataset of trajectories though).
reward_model: a RewardNet instance to be used for learning the reward
num_iterations: number of times to train the agent against the reward model
and then train the reward model against newly gathered preferences.
fragmenter: takes in a set of trajectories and returns pairs of fragments
for which preferences will be gathered. These fragments could be random,
or they could be selected more deliberately (active learning).
Default is a random fragmenter.
preference_gatherer: how to get preferences between trajectory fragments.
Default (and currently the only option) is to use synthetic preferences
based on ground-truth rewards. Human preferences could be implemented
here in the future.
reward_trainer: trains the reward model based on pairs of fragments and
associated preferences. Default is to use the preference model
and loss function from DRLHP.
comparison_queue_size: the maximum number of comparisons to keep in the
queue for training the reward model. If None, the queue will grow
without bound as new comparisons are added.
fragment_length: number of timesteps per fragment that is used to elicit
preferences
transition_oversampling: factor by which to oversample transitions before
creating fragments. Since fragments are sampled with replacement,
this is usually chosen > 1 to avoid having the same transition
in too many fragments.
initial_comparison_frac: fraction of the total_comparisons argument
to train() that will be sampled before the rest of training begins
(using a randomly initialized agent). This can be used to pretrain the
reward model before the agent is trained on the learned reward, to
help avoid irreversibly learning a bad policy from an untrained reward.
Note that there will often be some additional pretraining comparisons
since `comparisons_per_iteration` won't exactly divide the total number
of comparisons. How many such comparisons there are depends
discontinuously on `total_comparisons` and `comparisons_per_iteration`.
initial_epoch_multiplier: before agent training begins, train the reward
model for this many more epochs than usual (on fragments sampled from a
random agent).
custom_logger: Where to log to; if None (default), creates a new logger.
allow_variable_horizon: If False (default), algorithm will raise an
exception if it detects trajectories of different length during
training. If True, overrides this safety check. WARNING: variable
horizon episodes leak information about the reward via termination
condition, and can seriously confound evaluation. Read
https://imitation.readthedocs.io/en/latest/guide/variable_horizon.html
before overriding this.
rng: random number generator to use for initializing subcomponents such as
fragmenter.
Only used when default components are used; if you instantiate your
own fragmenter, preference gatherer, etc., you are responsible for
seeding them!
query_schedule: one of ("constant", "hyperbolic", "inverse_quadratic"), or
a function that takes in a float between 0 and 1 inclusive,
representing a fraction of the total number of timesteps elapsed up to
some time T, and returns a potentially unnormalized probability
indicating the fraction of `total_comparisons` that should be queried
at that iteration. This function will be called `num_iterations` times
in `__init__()` with values from `np.linspace(0, 1, num_iterations)`
as input. The outputs will be normalized to sum to 1 and then used to
apportion the comparisons among the `num_iterations` iterations.
Raises:
ValueError: if `query_schedule` is not a valid string or callable.
"""
super().__init__(
custom_logger=custom_logger,
allow_variable_horizon=allow_variable_horizon,
)
# for keeping track of the global iteration, in case train() is called
# multiple times
self._iteration = 0
self.model = reward_model
self.rng = rng
# are any of the optional args that require a rng None?
has_any_rng_args_none = None in (
preference_gatherer,
fragmenter,
reward_trainer,
)
if self.rng is None and has_any_rng_args_none:
raise ValueError(
"If you don't provide a random state, you must provide your own "
"seeded fragmenter, preference gatherer, and reward_trainer. "
"You can initialize a random state with `np.random.default_rng(seed)`.",
)
elif self.rng is not None and not has_any_rng_args_none:
raise ValueError(
"If you provide your own fragmenter, preference gatherer, "
"and reward trainer, you don't need to provide a random state.",
)
if reward_trainer is None:
assert self.rng is not None
preference_model = PreferenceModel(reward_model)
loss = CrossEntropyRewardLoss()
self.reward_trainer = _make_reward_trainer(
preference_model,
loss,
rng=self.rng,
)
else:
self.reward_trainer = reward_trainer
# If the reward trainer was created in the previous line, we've already passed
# the correct logger. But if the user created a RewardTrainer themselves and
# didn't manually set a logger, it would be annoying if a separate one was used.
self.reward_trainer.logger = self.logger
self.trajectory_generator = trajectory_generator
self.trajectory_generator.logger = self.logger
if fragmenter:
self.fragmenter = fragmenter
else:
assert self.rng is not None
self.fragmenter = RandomFragmenter(
custom_logger=self.logger,
rng=self.rng,
)
self.fragmenter.logger = self.logger
if preference_gatherer:
self.preference_gatherer = preference_gatherer
else:
assert self.rng is not None
self.preference_gatherer = SyntheticGatherer(
custom_logger=self.logger,
rng=self.rng,
)
self.preference_gatherer.logger = self.logger
self.fragment_length = fragment_length
self.initial_comparison_frac = initial_comparison_frac
self.initial_epoch_multiplier = initial_epoch_multiplier
self.num_iterations = num_iterations
self.transition_oversampling = transition_oversampling
if callable(query_schedule):
self.query_schedule = query_schedule
elif query_schedule in QUERY_SCHEDULES:
self.query_schedule = QUERY_SCHEDULES[query_schedule]
else:
raise ValueError(f"Unknown query schedule: {query_schedule}")
self.dataset = PreferenceDataset(max_size=comparison_queue_size)
[docs] def train(
self,
total_timesteps: int,
total_comparisons: int,
callback: Optional[Callable[[int], None]] = None,
) -> Mapping[str, Any]:
"""Train the reward model and the policy if applicable.
Args:
total_timesteps: number of environment interaction steps
total_comparisons: number of preferences to gather in total
callback: callback functions called at the end of each iteration
Returns:
A dictionary with final metrics such as loss and accuracy
of the reward model.
"""
initial_comparisons = int(total_comparisons * self.initial_comparison_frac)
total_comparisons -= initial_comparisons
# Compute the number of comparisons to request at each iteration in advance.
vec_schedule = np.vectorize(self.query_schedule)
unnormalized_probs = vec_schedule(np.linspace(0, 1, self.num_iterations))
probs = unnormalized_probs / np.sum(unnormalized_probs)
shares = util.oric(probs * total_comparisons)
schedule = [initial_comparisons] + shares.tolist()
print(f"Query schedule: {schedule}")
timesteps_per_iteration, extra_timesteps = divmod(
total_timesteps,
self.num_iterations,
)
reward_loss = None
reward_accuracy = None
for i, num_pairs in enumerate(schedule):
##########################
# Gather new preferences #
##########################
num_steps = math.ceil(
self.transition_oversampling * 2 * num_pairs * self.fragment_length,
)
self.logger.log(
f"Collecting {2 * num_pairs} fragments ({num_steps} transitions)",
)
trajectories = self.trajectory_generator.sample(num_steps)
# This assumes there are no fragments missing initial timesteps
# (but allows for fragments missing terminal timesteps).
horizons = (len(traj) for traj in trajectories if traj.terminal)
self._check_fixed_horizon(horizons)
self.logger.log("Creating fragment pairs")
fragments = self.fragmenter(trajectories, self.fragment_length, num_pairs)
with self.logger.accumulate_means("preferences"):
self.logger.log("Gathering preferences")
preferences = self.preference_gatherer(fragments)
self.dataset.push(fragments, preferences)
self.logger.log(f"Dataset now contains {len(self.dataset)} comparisons")
##########################
# Train the reward model #
##########################
# On the first iteration, we train the reward model for longer,
# as specified by initial_epoch_multiplier.
epoch_multiplier = 1.0
if i == 0:
epoch_multiplier = self.initial_epoch_multiplier
self.reward_trainer.train(self.dataset, epoch_multiplier=epoch_multiplier)
base_key = self.logger.get_accumulate_prefixes() + "reward/final/train"
assert f"{base_key}/loss" in self.logger.name_to_value
assert f"{base_key}/accuracy" in self.logger.name_to_value
reward_loss = self.logger.name_to_value[f"{base_key}/loss"]
reward_accuracy = self.logger.name_to_value[f"{base_key}/accuracy"]
###################
# Train the agent #
###################
num_steps = timesteps_per_iteration
# if the number of timesteps per iterations doesn't exactly divide
# the desired total number of timesteps, we train the agent a bit longer
# at the end of training (where the reward model is presumably best)
if i == self.num_iterations - 1:
num_steps += extra_timesteps
with self.logger.accumulate_means("agent"):
self.logger.log(f"Training agent for {num_steps} timesteps")
self.trajectory_generator.train(steps=num_steps)
self.logger.dump(self._iteration)
########################
# Additional Callbacks #
########################
if callback:
callback(self._iteration)
self._iteration += 1
return {"reward_loss": reward_loss, "reward_accuracy": reward_accuracy}