Shortcuts

Source code for torchrl.objectives.value.functional

# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
from __future__ import annotations

import math

import warnings
from functools import wraps
from typing import Optional, Tuple, Union

import torch


__all__ = [
    "generalized_advantage_estimate",
    "vec_generalized_advantage_estimate",
    "td0_advantage_estimate",
    "td0_return_estimate",
    "td1_return_estimate",
    "vec_td1_return_estimate",
    "td1_advantage_estimate",
    "vec_td1_advantage_estimate",
    "td_lambda_return_estimate",
    "vec_td_lambda_return_estimate",
    "td_lambda_advantage_estimate",
    "vec_td_lambda_advantage_estimate",
    "vtrace_advantage_estimate",
]

from torchrl.objectives.value.utils import (
    _custom_conv1d,
    _get_num_per_traj,
    _inv_pad_sequence,
    _make_gammas_tensor,
    _split_and_pad_sequence,
)

SHAPE_ERR = (
    "All input tensors (value, reward and done states) must share a unique shape."
)


def _transpose_time(fun):
    """Checks the time_dim argument of the function to allow for any dim.

    If not -2, makes a transpose of all the multi-dim input tensors to bring
    time at -2, and does the opposite transform for the outputs.
    """
    ERROR = (
        "The tensor shape and the time dimension are not compatible: "
        "got {} and time_dim={}."
    )

    @wraps(fun)
    def transposed_fun(*args, **kwargs):
        time_dim = kwargs.pop("time_dim", -2)

        def transpose_tensor(tensor):
            if not isinstance(tensor, torch.Tensor) or tensor.numel() <= 1:
                return tensor, False
            if time_dim >= 0:
                timedim = time_dim - tensor.ndim
            else:
                timedim = time_dim
            if timedim < -tensor.ndim or timedim >= 0:
                raise RuntimeError(ERROR.format(tensor.shape, timedim))
            if tensor.ndim >= 2:
                single_dim = False
                tensor = tensor.transpose(timedim, -2)
            elif tensor.ndim == 1 and timedim == -1:
                single_dim = True
                tensor = tensor.unsqueeze(-1)
            else:
                raise RuntimeError(ERROR.format(tensor.shape, timedim))
            return tensor, single_dim

        if time_dim != -2:
            single_dim = False
            if args:
                args, single_dim = zip(*(transpose_tensor(arg) for arg in args))
                single_dim = any(single_dim)
            for k, item in list(kwargs.items()):
                item, sd = transpose_tensor(item)
                single_dim = single_dim or sd
                kwargs[k] = item
            # We don't pass time_dim because it isn't supposed to be used thereafter
            out = fun(*args, **kwargs)
            if isinstance(out, torch.Tensor):
                out = transpose_tensor(out)[0]
                if single_dim:
                    out = out.squeeze(-2)
                return out
            if single_dim:
                return tuple(transpose_tensor(_out)[0].squeeze(-2) for _out in out)
            return tuple(transpose_tensor(_out)[0] for _out in out)
        # We don't pass time_dim because it isn't supposed to be used thereafter
        out = fun(*args, **kwargs)
        if isinstance(out, tuple):
            for _out in out:
                if _out.ndim < 2:
                    raise RuntimeError(ERROR.format(_out.shape, time_dim))
        else:
            if out.ndim < 2:
                raise RuntimeError(ERROR.format(out.shape, time_dim))
        return out

    return transposed_fun


########################################################################
# GAE
# ---


[docs]@_transpose_time def generalized_advantage_estimate( gamma: float, lmbda: float, state_value: torch.Tensor, next_state_value: torch.Tensor, reward: torch.Tensor, done: torch.Tensor, terminated: torch.Tensor | None = None, *, time_dim: int = -2, ) -> Tuple[torch.Tensor, torch.Tensor]: """Generalized advantage estimate of a trajectory. Refer to "HIGH-DIMENSIONAL CONTINUOUS CONTROL USING GENERALIZED ADVANTAGE ESTIMATION" https://arxiv.org/pdf/1506.02438.pdf for more context. Args: gamma (scalar): exponential mean discount. lmbda (scalar): trajectory discount. state_value (Tensor): value function result with old_state input. next_state_value (Tensor): value function result with new_state input. reward (Tensor): reward of taking actions in the environment. done (Tensor): boolean flag for end of trajectory. terminated (Tensor): boolean flag for the end of episode. Defaults to ``done`` if not provided. time_dim (int): dimension where the time is unrolled. Defaults to -2. All tensors (values, reward and done) must have shape ``[*Batch x TimeSteps x *F]``, with ``*F`` feature dimensions. """ if terminated is None: terminated = done if not ( next_state_value.shape == state_value.shape == reward.shape == done.shape == terminated.shape ): raise RuntimeError(SHAPE_ERR) dtype = next_state_value.dtype device = state_value.device not_done = (~done).int() not_terminated = (~terminated).int() *batch_size, time_steps, lastdim = not_done.shape advantage = torch.empty( *batch_size, time_steps, lastdim, device=device, dtype=dtype ) prev_advantage = 0 g_not_terminated = gamma * not_terminated delta = reward + (g_not_terminated * next_state_value) - state_value discount = lmbda * gamma * not_done for t in reversed(range(time_steps)): prev_advantage = advantage[..., t, :] = delta[..., t, :] + ( prev_advantage * discount[..., t, :] ) value_target = advantage + state_value return advantage, value_target
def _geom_series_like(t, r, thr): """Creates a geometric series of the form [1, gammalmbda, gammalmbda**2] with the shape of `t`. Drops all elements which are smaller than `thr`. """ if isinstance(r, torch.Tensor): r = r.item() if r == 0.0: return torch.zeros_like(t) elif r >= 1.0: lim = t.numel() else: lim = int(math.log(thr) / math.log(r)) rs = torch.full_like(t[:lim], r) rs[0] = 1.0 rs = rs.cumprod(0) rs = rs.unsqueeze(-1) return rs def _fast_vec_gae( reward: torch.Tensor, state_value: torch.Tensor, next_state_value: torch.Tensor, done: torch.Tensor, terminated: torch.Tensor, gamma: float, lmbda: float, thr: float = 1e-7, ): """Fast vectorized Generalized Advantage Estimate when gamma and lmbda are scalars. In contrast to `vec_generalized_advantage_estimate` this function does not need to allocate a big tensor of the form [B, T, T]. Args: reward (torch.Tensor): a [*B, T, F] tensor containing rewards state_value (torch.Tensor): a [*B, T, F] tensor containing state values (value function) next_state_value (torch.Tensor): a [*B, T, F] tensor containing next state values (value function) done (torch.Tensor): a [B, T] boolean tensor containing the done states. terminated (torch.Tensor): a [B, T] boolean tensor containing the terminated states. gamma (scalar): the gamma decay (trajectory discount) lmbda (scalar): the lambda decay (exponential mean discount) thr (float): threshold for the filter. Below this limit, components will ignored. Defaults to 1e-7. All tensors (values, reward and done) must have shape ``[*Batch x TimeSteps x F]``, with ``F`` feature dimensions. """ # _gen_num_per_traj and _split_and_pad_sequence need # time dimension at last position done = done.transpose(-2, -1) terminated = terminated.transpose(-2, -1) reward = reward.transpose(-2, -1) state_value = state_value.transpose(-2, -1) next_state_value = next_state_value.transpose(-2, -1) gammalmbda = gamma * lmbda not_terminated = (~terminated).int() td0 = reward + not_terminated * gamma * next_state_value - state_value num_per_traj = _get_num_per_traj(done) td0_flat, mask = _split_and_pad_sequence(td0, num_per_traj, return_mask=True) gammalmbdas = _geom_series_like(td0_flat[0], gammalmbda, thr=thr) advantage = _custom_conv1d(td0_flat.unsqueeze(1), gammalmbdas) advantage = advantage.squeeze(1) advantage = advantage[mask].view_as(reward) value_target = advantage + state_value advantage = advantage.transpose(-1, -2) value_target = value_target.transpose(-1, -2) return advantage, value_target
[docs]@_transpose_time def vec_generalized_advantage_estimate( gamma: Union[float, torch.Tensor], lmbda: Union[float, torch.Tensor], state_value: torch.Tensor, next_state_value: torch.Tensor, reward: torch.Tensor, done: torch.Tensor, terminated: torch.Tensor | None = None, *, time_dim: int = -2, ) -> Tuple[torch.Tensor, torch.Tensor]: """Vectorized Generalized advantage estimate of a trajectory. Refer to "HIGH-DIMENSIONAL CONTINUOUS CONTROL USING GENERALIZED ADVANTAGE ESTIMATION" https://arxiv.org/pdf/1506.02438.pdf for more context. Args: gamma (scalar): exponential mean discount. lmbda (scalar): trajectory discount. state_value (Tensor): value function result with old_state input. next_state_value (Tensor): value function result with new_state input. reward (Tensor): reward of taking actions in the environment. done (Tensor): boolean flag for end of trajectory. terminated (Tensor): boolean flag for the end of episode. Defaults to ``done`` if not provided. time_dim (int): dimension where the time is unrolled. Defaults to -2. All tensors (values, reward and done) must have shape ``[*Batch x TimeSteps x *F]``, with ``*F`` feature dimensions. """ if terminated is None: terminated = done if not ( next_state_value.shape == state_value.shape == reward.shape == done.shape == terminated.shape ): raise RuntimeError(SHAPE_ERR) dtype = state_value.dtype *batch_size, time_steps, lastdim = terminated.shape value = gamma * lmbda if isinstance(value, torch.Tensor) and value.numel() > 1: # create tensor while ensuring that gradients are passed not_done = (~done).to(dtype) gammalmbdas = not_done * value else: # when gamma and lmbda are scalars, use fast_vec_gae implementation return _fast_vec_gae( reward=reward, state_value=state_value, next_state_value=next_state_value, done=done, terminated=terminated, gamma=gamma, lmbda=lmbda, ) gammalmbdas = _make_gammas_tensor(gammalmbdas, time_steps, True) gammalmbdas = gammalmbdas.cumprod(-2) first_below_thr = gammalmbdas < 1e-7 # if we have multiple gammas, we only want to truncate if _all_ of # the geometric sequences fall below the threshold first_below_thr = first_below_thr.flatten(0, 1).all(0).all(-1) if first_below_thr.any(): first_below_thr = torch.where(first_below_thr)[0][0].item() gammalmbdas = gammalmbdas[..., :first_below_thr, :] not_terminated = (~terminated).to(dtype) td0 = reward + not_terminated * gamma * next_state_value - state_value if len(batch_size) > 1: td0 = td0.flatten(0, len(batch_size) - 1) elif not len(batch_size): td0 = td0.unsqueeze(0) td0_r = td0.transpose(-2, -1) shapes = td0_r.shape[:2] if lastdim != 1: # then we flatten again the first dims and reset a singleton in between td0_r = td0_r.flatten(0, 1).unsqueeze(1) advantage = _custom_conv1d(td0_r, gammalmbdas) if lastdim != 1: advantage = advantage.squeeze(1).unflatten(0, shapes) if len(batch_size) > 1: advantage = advantage.unflatten(0, batch_size) elif not len(batch_size): advantage = advantage.squeeze(0) advantage = advantage.transpose(-2, -1) value_target = advantage + state_value return advantage, value_target
######################################################################## # TD(0) # -----
[docs]def td0_advantage_estimate( gamma: float, state_value: torch.Tensor, next_state_value: torch.Tensor, reward: torch.Tensor, done: torch.Tensor, terminated: torch.Tensor | None = None, ) -> Tuple[torch.Tensor, torch.Tensor]: """TD(0) advantage estimate of a trajectory. Also known as bootstrapped Temporal Difference or one-step return. Args: gamma (scalar): exponential mean discount. state_value (Tensor): value function result with old_state input. next_state_value (Tensor): value function result with new_state input. reward (Tensor): reward of taking actions in the environment. done (Tensor): boolean flag for end of trajectory. terminated (Tensor): boolean flag for the end of episode. Defaults to ``done`` if not provided. All tensors (values, reward and done) must have shape ``[*Batch x TimeSteps x *F]``, with ``*F`` feature dimensions. """ if terminated is None: terminated = done if not ( next_state_value.shape == state_value.shape == reward.shape == done.shape == terminated.shape ): raise RuntimeError(SHAPE_ERR) returns = td0_return_estimate(gamma, next_state_value, reward, terminated) advantage = returns - state_value return advantage
[docs]def td0_return_estimate( gamma: float, next_state_value: torch.Tensor, reward: torch.Tensor, terminated: torch.Tensor | None = None, *, done: torch.Tensor | None = None, ) -> Tuple[torch.Tensor, torch.Tensor]: # noqa: D417 """TD(0) discounted return estimate of a trajectory. Also known as bootstrapped Temporal Difference or one-step return. Args: gamma (scalar): exponential mean discount. next_state_value (Tensor): value function result with new_state input. must be a [Batch x TimeSteps x 1] or [Batch x TimeSteps] tensor reward (Tensor): reward of taking actions in the environment. must be a [Batch x TimeSteps x 1] or [Batch x TimeSteps] tensor terminated (Tensor): boolean flag for the end of episode. Defaults to ``done`` if not provided. Keyword Args: done (Tensor): Deprecated. Use ``terminated`` instead. All tensors (values, reward and done) must have shape ``[*Batch x TimeSteps x *F]``, with ``*F`` feature dimensions. """ if done is not None and terminated is None: terminated = done warnings.warn( "done for td0_return_estimate is deprecated. Pass ``terminated`` instead." ) if not (next_state_value.shape == reward.shape == terminated.shape): raise RuntimeError(SHAPE_ERR) not_terminated = (~terminated).int() advantage = reward + gamma * not_terminated * next_state_value return advantage
######################################################################## # TD(1) # ----------
[docs]@_transpose_time def td1_return_estimate( gamma: float, next_state_value: torch.Tensor, reward: torch.Tensor, done: torch.Tensor, terminated: torch.Tensor | None = None, rolling_gamma: bool = None, *, time_dim: int = -2, ) -> torch.Tensor: r"""TD(1) return estimate. Args: gamma (scalar): exponential mean discount. next_state_value (Tensor): value function result with new_state input. reward (Tensor): reward of taking actions in the environment. done (Tensor): boolean flag for end of trajectory. terminated (Tensor): boolean flag for the end of episode. Defaults to ``done`` if not provided. rolling_gamma (bool, optional): if ``True``, it is assumed that each gamma if a gamma tensor is tied to a single event: gamma = [g1, g2, g3, g4] value = [v1, v2, v3, v4] return = [ v1 + g1 v2 + g1 g2 v3 + g1 g2 g3 v4, v2 + g2 v3 + g2 g3 v4, v3 + g3 v4, v4, ] if False, it is assumed that each gamma is tied to the upcoming trajectory: gamma = [g1, g2, g3, g4] value = [v1, v2, v3, v4] return = [ v1 + g1 v2 + g1**2 v3 + g**3 v4, v2 + g2 v3 + g2**2 v4, v3 + g3 v4, v4, ] Default is True. time_dim (int): dimension where the time is unrolled. Defaults to -2. All tensors (values, reward and done) must have shape ``[*Batch x TimeSteps x *F]``, with ``*F`` feature dimensions. """ if terminated is None: terminated = done if not (next_state_value.shape == reward.shape == done.shape == terminated.shape): raise RuntimeError(SHAPE_ERR) not_done = (~done).int() not_terminated = (~terminated).int() returns = torch.empty_like(next_state_value) T = returns.shape[-2] single_gamma = False if not (isinstance(gamma, torch.Tensor) and gamma.shape == not_done.shape): single_gamma = True gamma = torch.full_like(next_state_value, gamma) if rolling_gamma is None: rolling_gamma = True elif not rolling_gamma and single_gamma: raise RuntimeError( "rolling_gamma=False is expected only with time-sensitive gamma values" ) done_but_not_terminated = (done & ~terminated).int() if rolling_gamma: gamma = gamma * not_terminated g = next_state_value[..., -1, :] for i in reversed(range(T)): # if not done (and hence not terminated), get the bootstrapped value # if done but not terminated, get nex_val # if terminated, take nothing (gamma = 0) dnt = done_but_not_terminated[..., i, :] g = returns[..., i, :] = reward[..., i, :] + gamma[..., i, :] * ( (1 - dnt) * g + dnt * next_state_value[..., i, :] ) else: for k in range(T): g = 0 _gamma = gamma[..., k, :] nd = not_terminated _gamma = _gamma.unsqueeze(-2) * nd for i in reversed(range(k, T)): dnt = done_but_not_terminated[..., i, :] g = reward[..., i, :] + _gamma[..., i, :] * ( (1 - dnt) * g + dnt * next_state_value[..., i, :] ) returns[..., k, :] = g return returns
[docs]def td1_advantage_estimate( gamma: float, state_value: torch.Tensor, next_state_value: torch.Tensor, reward: torch.Tensor, done: torch.Tensor, terminated: torch.Tensor | None = None, rolling_gamma: bool = None, time_dim: int = -2, ) -> torch.Tensor: """TD(1) advantage estimate. Args: gamma (scalar): exponential mean discount. state_value (Tensor): value function result with old_state input. next_state_value (Tensor): value function result with new_state input. reward (Tensor): reward of taking actions in the environment. done (Tensor): boolean flag for end of trajectory. terminated (Tensor): boolean flag for the end of episode. Defaults to ``done`` if not provided. rolling_gamma (bool, optional): if ``True``, it is assumed that each gamma if a gamma tensor is tied to a single event: gamma = [g1, g2, g3, g4] value = [v1, v2, v3, v4] return = [ v1 + g1 v2 + g1 g2 v3 + g1 g2 g3 v4, v2 + g2 v3 + g2 g3 v4, v3 + g3 v4, v4, ] if False, it is assumed that each gamma is tied to the upcoming trajectory: gamma = [g1, g2, g3, g4] value = [v1, v2, v3, v4] return = [ v1 + g1 v2 + g1**2 v3 + g**3 v4, v2 + g2 v3 + g2**2 v4, v3 + g3 v4, v4, ] Default is True. time_dim (int): dimension where the time is unrolled. Defaults to -2. All tensors (values, reward and done) must have shape ``[*Batch x TimeSteps x *F]``, with ``*F`` feature dimensions. """ if terminated is None: terminated = done if not ( next_state_value.shape == state_value.shape == reward.shape == done.shape == terminated.shape ): raise RuntimeError(SHAPE_ERR) if not state_value.shape == next_state_value.shape: raise RuntimeError("shape of state_value and next_state_value must match") returns = td1_return_estimate( gamma, next_state_value, reward, done, terminated=terminated, rolling_gamma=rolling_gamma, time_dim=time_dim, ) advantage = returns - state_value return advantage
[docs]@_transpose_time def vec_td1_return_estimate( gamma, next_state_value, reward, done: torch.Tensor, terminated: torch.Tensor | None = None, rolling_gamma: Optional[bool] = None, time_dim: int = -2, ): """Vectorized TD(1) return estimate. Args: gamma (scalar, Tensor): exponential mean discount. If tensor-valued, next_state_value (Tensor): value function result with new_state input. reward (Tensor): reward of taking actions in the environment. done (Tensor): boolean flag for end of trajectory. terminated (Tensor): boolean flag for the end of episode. Defaults to ``done`` if not provided. rolling_gamma (bool, optional): if ``True``, it is assumed that each gamma if a gamma tensor is tied to a single event: gamma = [g1, g2, g3, g4] value = [v1, v2, v3, v4] return = [ v1 + g1 v2 + g1 g2 v3 + g1 g2 g3 v4, v2 + g2 v3 + g2 g3 v4, v3 + g3 v4, v4, ] if False, it is assumed that each gamma is tied to the upcoming trajectory: gamma = [g1, g2, g3, g4] value = [v1, v2, v3, v4] return = [ v1 + g1 v2 + g1**2 v3 + g**3 v4, v2 + g2 v3 + g2**2 v4, v3 + g3 v4, v4, ] Default is True. time_dim (int): dimension where the time is unrolled. Defaults to -2. All tensors (values, reward and done) must have shape ``[*Batch x TimeSteps x *F]``, with ``*F`` feature dimensions. """ return vec_td_lambda_return_estimate( gamma=gamma, next_state_value=next_state_value, reward=reward, done=done, terminated=terminated, rolling_gamma=rolling_gamma, lmbda=1, time_dim=time_dim, )
[docs]def vec_td1_advantage_estimate( gamma, state_value, next_state_value, reward, done: torch.Tensor, terminated: torch.Tensor | None = None, rolling_gamma: bool = None, time_dim: int = -2, ): """Vectorized TD(1) advantage estimate. Args: gamma (scalar, Tensor): exponential mean discount. If tensor-valued, state_value (Tensor): value function result with old_state input. next_state_value (Tensor): value function result with new_state input. reward (Tensor): reward of taking actions in the environment. done (Tensor): boolean flag for end of trajectory. terminated (Tensor): boolean flag for the end of episode. Defaults to ``done`` if not provided. rolling_gamma (bool, optional): if ``True``, it is assumed that each gamma if a gamma tensor is tied to a single event: gamma = [g1, g2, g3, g4] value = [v1, v2, v3, v4] return = [ v1 + g1 v2 + g1 g2 v3 + g1 g2 g3 v4, v2 + g2 v3 + g2 g3 v4, v3 + g3 v4, v4, ] if False, it is assumed that each gamma is tied to the upcoming trajectory: gamma = [g1, g2, g3, g4] value = [v1, v2, v3, v4] return = [ v1 + g1 v2 + g1**2 v3 + g**3 v4, v2 + g2 v3 + g2**2 v4, v3 + g3 v4, v4, ] Default is True. time_dim (int): dimension where the time is unrolled. Defaults to -2. All tensors (values, reward and done) must have shape ``[*Batch x TimeSteps x *F]``, with ``*F`` feature dimensions. """ if terminated is None: terminated = done if not ( next_state_value.shape == state_value.shape == reward.shape == done.shape == terminated.shape ): raise RuntimeError(SHAPE_ERR) return ( vec_td1_return_estimate( gamma, next_state_value, reward, done=done, terminated=terminated, rolling_gamma=rolling_gamma, time_dim=time_dim, ) - state_value )
######################################################################## # TD(lambda) # ----------
[docs]@_transpose_time def td_lambda_return_estimate( gamma: float, lmbda: float, next_state_value: torch.Tensor, reward: torch.Tensor, done: torch.Tensor, terminated: torch.Tensor | None = None, rolling_gamma: bool = None, *, time_dim: int = -2, ) -> torch.Tensor: r"""TD(:math:`\lambda`) return estimate. Args: gamma (scalar): exponential mean discount. lmbda (scalar): trajectory discount. next_state_value (Tensor): value function result with new_state input. reward (Tensor): reward of taking actions in the environment. done (Tensor): boolean flag for end of trajectory. terminated (Tensor): boolean flag for the end of episode. Defaults to ``done`` if not provided. rolling_gamma (bool, optional): if ``True``, it is assumed that each gamma if a gamma tensor is tied to a single event: gamma = [g1, g2, g3, g4] value = [v1, v2, v3, v4] return = [ v1 + g1 v2 + g1 g2 v3 + g1 g2 g3 v4, v2 + g2 v3 + g2 g3 v4, v3 + g3 v4, v4, ] if False, it is assumed that each gamma is tied to the upcoming trajectory: gamma = [g1, g2, g3, g4] value = [v1, v2, v3, v4] return = [ v1 + g1 v2 + g1**2 v3 + g**3 v4, v2 + g2 v3 + g2**2 v4, v3 + g3 v4, v4, ] Default is True. time_dim (int): dimension where the time is unrolled. Defaults to -2. All tensors (values, reward and done) must have shape ``[*Batch x TimeSteps x *F]``, with ``*F`` feature dimensions. """ if terminated is None: terminated = done if not (next_state_value.shape == reward.shape == done.shape == terminated.shape): raise RuntimeError(SHAPE_ERR) not_terminated = (~terminated).int() returns = torch.empty_like(next_state_value) next_state_value = next_state_value * not_terminated *batch, T, lastdim = returns.shape # if gamma is not a tensor of the same shape as other inputs, we use rolling_gamma = True single_gamma = False if not (isinstance(gamma, torch.Tensor) and gamma.shape == done.shape): single_gamma = True gamma = torch.full_like(next_state_value, gamma) single_lambda = False if not (isinstance(lmbda, torch.Tensor) and lmbda.shape == done.shape): single_lambda = True lmbda = torch.full_like(next_state_value, lmbda) if rolling_gamma is None: rolling_gamma = True elif not rolling_gamma and single_gamma and single_lambda: raise RuntimeError( "rolling_gamma=False is expected only with time-sensitive gamma or lambda values" ) if rolling_gamma: g = next_state_value[..., -1, :] for i in reversed(range(T)): dn = done[..., i, :].int() nv = next_state_value[..., i, :] lmd = lmbda[..., i, :] # if done, the bootstrapped gain is the next value, otherwise it's the # value we computed during the previous iter g = g * (1 - dn) + nv * dn g = returns[..., i, :] = reward[..., i, :] + gamma[..., i, :] * ( (1 - lmd) * nv + lmd * g ) else: for k in range(T): g = next_state_value[..., -1, :] _gamma = gamma[..., k, :] _lambda = lmbda[..., k, :] for i in reversed(range(k, T)): dn = done[..., i, :].int() nv = next_state_value[..., i, :] g = g * (1 - dn) + nv * dn g = reward[..., i, :] + _gamma * ((1 - _lambda) * nv + _lambda * g) returns[..., k, :] = g return returns
[docs]def td_lambda_advantage_estimate( gamma: float, lmbda: float, state_value: torch.Tensor, next_state_value: torch.Tensor, reward: torch.Tensor, done: torch.Tensor, terminated: torch.Tensor | None = None, rolling_gamma: bool = None, # not a kwarg because used directly time_dim: int = -2, ) -> torch.Tensor: r"""TD(:math:`\lambda`) advantage estimate. Args: gamma (scalar): exponential mean discount. lmbda (scalar): trajectory discount. state_value (Tensor): value function result with old_state input. next_state_value (Tensor): value function result with new_state input. reward (Tensor): reward of taking actions in the environment. done (Tensor): boolean flag for end of trajectory. terminated (Tensor): boolean flag for the end of episode. Defaults to ``done`` if not provided. rolling_gamma (bool, optional): if ``True``, it is assumed that each gamma if a gamma tensor is tied to a single event: gamma = [g1, g2, g3, g4] value = [v1, v2, v3, v4] return = [ v1 + g1 v2 + g1 g2 v3 + g1 g2 g3 v4, v2 + g2 v3 + g2 g3 v4, v3 + g3 v4, v4, ] if False, it is assumed that each gamma is tied to the upcoming trajectory: gamma = [g1, g2, g3, g4] value = [v1, v2, v3, v4] return = [ v1 + g1 v2 + g1**2 v3 + g**3 v4, v2 + g2 v3 + g2**2 v4, v3 + g3 v4, v4, ] Default is True. time_dim (int): dimension where the time is unrolled. Defaults to -2. All tensors (values, reward and done) must have shape ``[*Batch x TimeSteps x *F]``, with ``*F`` feature dimensions. """ if terminated is None: terminated = done if not ( next_state_value.shape == state_value.shape == reward.shape == done.shape == terminated.shape ): raise RuntimeError(SHAPE_ERR) if not state_value.shape == next_state_value.shape: raise RuntimeError("shape of state_value and next_state_value must match") returns = td_lambda_return_estimate( gamma, lmbda, next_state_value, reward, done, terminated=terminated, rolling_gamma=rolling_gamma, time_dim=time_dim, ) advantage = returns - state_value return advantage
def _fast_td_lambda_return_estimate( gamma: Union[torch.Tensor, float], lmbda: float, next_state_value: torch.Tensor, reward: torch.Tensor, done: torch.Tensor, terminated: torch.Tensor, thr: float = 1e-7, ): """Fast vectorized TD lambda return estimate. In contrast to the generalized `vec_td_lambda_return_estimate` this function does not need to allocate a big tensor of the form [B, T, T], but it only works with gamma/lmbda being scalars. Args: gamma (scalar): the gamma decay, can be a tensor with a single element (trajectory discount) lmbda (scalar): the lambda decay (exponential mean discount) next_state_value (torch.Tensor): a [*B, T, F] tensor containing next state values (value function) reward (torch.Tensor): a [*B, T, F] tensor containing rewards done (Tensor): boolean flag for end of trajectory. terminated (Tensor): boolean flag for end of episode. thr (float): threshold for the filter. Below this limit, components will ignored. Defaults to 1e-7. All tensors (values, reward and done) must have shape ``[*Batch x TimeSteps x F]``, with ``F`` feature dimensions. """ device = reward.device done = done.transpose(-2, -1) terminated = terminated.transpose(-2, -1) reward = reward.transpose(-2, -1) next_state_value = next_state_value.transpose(-2, -1) # the only valid next states are those where the trajectory does not terminate next_state_value = (~terminated).int() * next_state_value gamma_tensor = torch.tensor([gamma], device=device) gammalmbda = gamma_tensor * lmbda num_per_traj = _get_num_per_traj(done) done = done.clone() done[..., -1] = 1 not_done = (~done).int() t = reward + next_state_value * gamma_tensor * (1 - not_done * lmbda) t_flat, mask = _split_and_pad_sequence(t, num_per_traj, return_mask=True) gammalmbdas = _geom_series_like(t_flat[0], gammalmbda, thr=thr) ret_flat = _custom_conv1d(t_flat.unsqueeze(1), gammalmbdas) ret = ret_flat.squeeze(1)[mask] return ret.view_as(reward).transpose(-1, -2)
[docs]@_transpose_time def vec_td_lambda_return_estimate( gamma, lmbda, next_state_value, reward, done, terminated: torch.Tensor | None = None, rolling_gamma: Optional[bool] = None, *, time_dim: int = -2, ): r"""Vectorized TD(:math:`\lambda`) return estimate. Args: gamma (scalar, Tensor): exponential mean discount. If tensor-valued, must be a [Batch x TimeSteps x 1] tensor. lmbda (scalar): trajectory discount. next_state_value (Tensor): value function result with new_state input. must be a [Batch x TimeSteps x 1] tensor reward (Tensor): reward of taking actions in the environment. must be a [Batch x TimeSteps x 1] or [Batch x TimeSteps] tensor done (Tensor): boolean flag for end of trajectory. terminated (Tensor): boolean flag for the end of episode. Defaults to ``done`` if not provided. rolling_gamma (bool, optional): if ``True``, it is assumed that each gamma if a gamma tensor is tied to a single event: gamma = [g1, g2, g3, g4] value = [v1, v2, v3, v4] return = [ v1 + g1 v2 + g1 g2 v3 + g1 g2 g3 v4, v2 + g2 v3 + g2 g3 v4, v3 + g3 v4, v4, ] if False, it is assumed that each gamma is tied to the upcoming trajectory: gamma = [g1, g2, g3, g4] value = [v1, v2, v3, v4] return = [ v1 + g1 v2 + g1**2 v3 + g**3 v4, v2 + g2 v3 + g2**2 v4, v3 + g3 v4, v4, ] Default is True. time_dim (int): dimension where the time is unrolled. Defaults to -2. All tensors (values, reward and done) must have shape ``[*Batch x TimeSteps x *F]``, with ``*F`` feature dimensions. """ if terminated is None: terminated = done if not (next_state_value.shape == reward.shape == done.shape == terminated.shape): raise RuntimeError(SHAPE_ERR) gamma_thr = 1e-7 shape = next_state_value.shape *batch, T, lastdim = shape def _is_scalar(tensor): return not isinstance(tensor, torch.Tensor) or tensor.numel() == 1 # There are two use-cases: if gamma/lmbda are scalars we can use the # fast implementation, if not we must construct a gamma tensor. if _is_scalar(gamma) and _is_scalar(lmbda): return _fast_td_lambda_return_estimate( gamma=gamma, lmbda=lmbda, next_state_value=next_state_value, reward=reward, done=done, terminated=terminated, thr=gamma_thr, ) next_state_value = next_state_value.transpose(-2, -1).unsqueeze(-2) if len(batch): next_state_value = next_state_value.flatten(0, len(batch)) reward = reward.transpose(-2, -1).unsqueeze(-2) if len(batch): reward = reward.flatten(0, len(batch)) """Vectorized version of td_lambda_advantage_estimate""" device = reward.device not_done = (~done).int() not_terminated = (~terminated).int().transpose(-2, -1).unsqueeze(-2) if len(batch): not_terminated = not_terminated.flatten(0, len(batch)) next_state_value = next_state_value * not_terminated if rolling_gamma is None: rolling_gamma = True if not rolling_gamma: terminated_follows_terminated = terminated[..., 1:, :][ terminated[..., :-1, :] ].all() if not terminated_follows_terminated: raise NotImplementedError( "When using rolling_gamma=False and vectorized TD(lambda) with time-dependent gamma, " "make sure that conseducitve trajectories are separated as different batch " "items. Propagating a gamma value across trajectories is not permitted with " "this method. Check that you need to use rolling_gamma=False, and if so " "consider using the non-vectorized version of the return computation or splitting " "your trajectories." ) if rolling_gamma: # Make the coefficient table gammas = _make_gammas_tensor(gamma * not_done, T, rolling_gamma) gammas_cp = torch.cumprod(gammas, -2) lambdas = torch.ones(T + 1, 1, device=device) lambdas[1:] = lmbda lambdas_cp = torch.cumprod(lambdas, -2) lambdas = lambdas[1:] dec = gammas_cp * lambdas_cp gammas = _make_gammas_tensor(gamma, T, rolling_gamma) gammas = gammas[..., 1:, :] if gammas.ndimension() == 4 and gammas.shape[1] > 1: gammas = gammas[:, :1] if lambdas.ndimension() == 4 and lambdas.shape[1] > 1: lambdas = lambdas[:, :1] not_done = not_done.transpose(-2, -1).unsqueeze(-2) if len(batch): not_done = not_done.flatten(0, len(batch)) # lambdas = lambdas * not_done v3 = (gammas * lambdas).squeeze(-1) * next_state_value * not_done v3[..., :-1] = 0 out = _custom_conv1d( reward + gammas.squeeze(-1) * next_state_value * (1 - lambdas.squeeze(-1) * not_done) + v3, dec, ) return out.view(*batch, lastdim, T).transpose(-2, -1) else: raise NotImplementedError( "The vectorized version of TD(lambda) with rolling_gamma=False is currently not available. " "To use this feature, use the non-vectorized version of TD(lambda). You can expect " "good speed improvements by decorating the function with torch.compile!" )
[docs]def vec_td_lambda_advantage_estimate( gamma, lmbda, state_value, next_state_value, reward, done, terminated: torch.Tensor | None = None, rolling_gamma: bool = None, # not a kwarg because used directly time_dim: int = -2, ): r"""Vectorized TD(:math:`\lambda`) advantage estimate. Args: gamma (scalar, Tensor): exponential mean discount. If tensor-valued, lmbda (scalar): trajectory discount. state_value (Tensor): value function result with old_state input. next_state_value (Tensor): value function result with new_state input. reward (Tensor): reward of taking actions in the environment. done (Tensor): boolean flag for end of trajectory. terminated (Tensor): boolean flag for the end of episode. Defaults to ``done`` if not provided. rolling_gamma (bool, optional): if ``True``, it is assumed that each gamma if a gamma tensor is tied to a single event: gamma = [g1, g2, g3, g4] value = [v1, v2, v3, v4] return = [ v1 + g1 v2 + g1 g2 v3 + g1 g2 g3 v4, v2 + g2 v3 + g2 g3 v4, v3 + g3 v4, v4, ] if False, it is assumed that each gamma is tied to the upcoming trajectory: gamma = [g1, g2, g3, g4] value = [v1, v2, v3, v4] return = [ v1 + g1 v2 + g1**2 v3 + g**3 v4, v2 + g2 v3 + g2**2 v4, v3 + g3 v4, v4, ] Default is True. time_dim (int): dimension where the time is unrolled. Defaults to -2. All tensors (values, reward and done) must have shape ``[*Batch x TimeSteps x *F]``, with ``*F`` feature dimensions. """ if terminated is None: terminated = done if not ( next_state_value.shape == state_value.shape == reward.shape == done.shape == terminated.shape ): raise RuntimeError(SHAPE_ERR) return ( vec_td_lambda_return_estimate( gamma, lmbda, next_state_value, reward, done=done, terminated=terminated, rolling_gamma=rolling_gamma, time_dim=time_dim, ) - state_value )
######################################################################## # V-Trace # ----- @_transpose_time def vtrace_advantage_estimate( gamma: float, log_pi: torch.Tensor, log_mu: torch.Tensor, state_value: torch.Tensor, next_state_value: torch.Tensor, reward: torch.Tensor, done: torch.Tensor, terminated: torch.Tensor | None = None, rho_thresh: Union[float, torch.Tensor] = 1.0, c_thresh: Union[float, torch.Tensor] = 1.0, # not a kwarg because used directly time_dim: int = -2, ) -> Tuple[torch.Tensor, torch.Tensor]: """Computes V-Trace off-policy actor critic targets. Refer to "IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures" https://arxiv.org/abs/1802.01561 for more context. Args: gamma (scalar): exponential mean discount. log_pi (Tensor): collection actor log probability of taking actions in the environment. log_mu (Tensor): current actor log probability of taking actions in the environment. state_value (Tensor): value function result with state input. next_state_value (Tensor): value function result with next_state input. reward (Tensor): reward of taking actions in the environment. done (Tensor): boolean flag for end of episode. terminated (torch.Tensor): a [B, T] boolean tensor containing the terminated states. rho_thresh (Union[float, Tensor]): rho clipping parameter for importance weights. c_thresh (Union[float, Tensor]): c clipping parameter for importance weights. time_dim (int): dimension where the time is unrolled. Defaults to -2. All tensors (values, reward and done) must have shape ``[*Batch x TimeSteps x *F]``, with ``*F`` feature dimensions. """ if not (next_state_value.shape == state_value.shape == reward.shape == done.shape): raise RuntimeError(SHAPE_ERR) device = state_value.device if not isinstance(rho_thresh, torch.Tensor): rho_thresh = torch.tensor(rho_thresh, device=device) if not isinstance(c_thresh, torch.Tensor): c_thresh = torch.tensor(c_thresh, device=device) c_thresh = c_thresh.to(device) rho_thresh = rho_thresh.to(device) not_done = (~done).int() not_terminated = not_done if terminated is None else (~terminated).int() *batch_size, time_steps, lastdim = not_done.shape done_discounts = gamma * not_done terminated_discounts = gamma * not_terminated rho = (log_pi - log_mu).exp() clipped_rho = rho.clamp_max(rho_thresh) deltas = clipped_rho * ( reward + terminated_discounts * next_state_value - state_value ) clipped_c = rho.clamp_max(c_thresh) vs_minus_v_xs = [torch.zeros_like(next_state_value[..., -1, :])] for i in reversed(range(time_steps)): discount_t, c_t, delta_t = ( done_discounts[..., i, :], clipped_c[..., i, :], deltas[..., i, :], ) vs_minus_v_xs.append(delta_t + discount_t * c_t * vs_minus_v_xs[-1]) vs_minus_v_xs = torch.stack(vs_minus_v_xs[1:], dim=time_dim) vs_minus_v_xs = torch.flip(vs_minus_v_xs, dims=[time_dim]) vs = vs_minus_v_xs + state_value vs_t_plus_1 = torch.cat( [vs[..., 1:, :], next_state_value[..., -1:, :]], dim=time_dim ) advantages = clipped_rho * ( reward + terminated_discounts * vs_t_plus_1 - state_value ) return advantages, vs ######################################################################## # Reward to go # ------------
[docs]@_transpose_time def reward2go( reward, done, gamma, *, time_dim: int = -2, ): """Compute the discounted cumulative sum of rewards given multiple trajectories and the episode ends. Args: reward (torch.Tensor): A tensor containing the rewards received at each time step over multiple trajectories. done (Tensor): boolean flag for end of episode. Differs from truncated, where the episode did not end but was interrupted. gamma (float, optional): The discount factor to use for computing the discounted cumulative sum of rewards. Defaults to 1.0. time_dim (int): dimension where the time is unrolled. Defaults to -2. Returns: torch.Tensor: A tensor of shape [B, T] containing the discounted cumulative sum of rewards (reward-to-go) at each time step. Examples: >>> reward = torch.ones(1, 10) >>> done = torch.zeros(1, 10, dtype=torch.bool) >>> done[:, [3, 7]] = True >>> reward2go(reward, done, 0.99, time_dim=-1) tensor([[3.9404], [2.9701], [1.9900], [1.0000], [3.9404], [2.9701], [1.9900], [1.0000], [1.9900], [1.0000]]) """ shape = reward.shape if shape != done.shape: raise ValueError( f"reward and done must share the same shape, got {reward.shape} and {done.shape}" ) # flatten if needed if reward.ndim > 2: # we know time dim is at -2, let's put it at -3 rflip = reward.transpose(-2, -3) rflip_shape = rflip.shape[-2:] r2go = reward2go( rflip.flatten(-2, -1), done.transpose(-2, -3).flatten(-2, -1), gamma=gamma ).unflatten(-1, rflip_shape) return r2go.transpose(-2, -3) # place time at dim -1 reward = reward.transpose(-2, -1) done = done.transpose(-2, -1) num_per_traj = _get_num_per_traj(done) td0_flat = _split_and_pad_sequence(reward, num_per_traj) gammas = _geom_series_like(td0_flat[0], gamma, thr=1e-7) cumsum = _custom_conv1d(td0_flat.unsqueeze(1), gammas) cumsum = cumsum.squeeze(1) cumsum = _inv_pad_sequence(cumsum, num_per_traj) cumsum = cumsum.reshape_as(reward) cumsum = cumsum.transpose(-2, -1) if cumsum.shape != shape: try: cumsum = cumsum.reshape(shape) except RuntimeError: raise RuntimeError( f"Wrong shape for output reward2go: {cumsum.shape} when {shape} was expected." ) return cumsum

Docs

Access comprehensive developer documentation for PyTorch

View Docs

Tutorials

Get in-depth tutorials for beginners and advanced developers

View Tutorials

Resources

Find development resources and get your questions answered

View Resources