Source code for torchrl.modules.distributions.continuous
# 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 warnings
import weakref
from numbers import Number
from typing import Dict, Optional, Sequence, Tuple, Union
import numpy as np
import torch
from torch import distributions as D, nn
try:
from torch.compiler import assume_constant_result
except ImportError:
from torch._dynamo import assume_constant_result
from torch.distributions import constraints
from torch.distributions.transforms import _InverseTransform
from torchrl.modules.distributions.truncated_normal import (
TruncatedNormal as _TruncatedNormal,
)
from torchrl.modules.distributions.utils import (
_cast_device,
FasterTransformedDistribution,
safeatanh_noeps,
safetanh_noeps,
)
from torchrl.modules.utils import mappings
# speeds up distribution construction
D.Distribution.set_default_validate_args(False)
try:
from torch.compiler import is_dynamo_compiling
except ImportError:
from torch._dynamo import is_compiling as is_dynamo_compiling
[docs]class IndependentNormal(D.Independent):
"""Implements a Normal distribution with location scaling.
Location scaling prevents the location to be "too far" from 0, which ultimately
leads to numerically unstable samples and poor gradient computation (e.g. gradient explosion).
In practice, the location is computed according to
.. math::
loc = tanh(loc / upscale) * upscale.
This behavior can be disabled by switching off the tanh_loc parameter (see below).
Args:
loc (torch.Tensor): normal distribution location parameter
scale (torch.Tensor): normal distribution sigma parameter (squared root of variance)
upscale (torch.Tensor or number, optional): 'a' scaling factor in the formula:
.. math::
loc = tanh(loc / upscale) * upscale.
Default is 5.0
tanh_loc (bool, optional): if ``False``, the above formula is used for
the location scaling, otherwise the raw value
is kept. Default is ``False``;
"""
num_params: int = 2
def __init__(
self,
loc: torch.Tensor,
scale: torch.Tensor,
upscale: float = 5.0,
tanh_loc: bool = False,
event_dim: int = 1,
**kwargs,
):
self.tanh_loc = tanh_loc
self.upscale = upscale
self._event_dim = event_dim
self._kwargs = kwargs
super().__init__(D.Normal(loc, scale, **kwargs), event_dim)
def update(self, loc, scale):
if self.tanh_loc:
loc = self.upscale * (loc / self.upscale).tanh()
super().__init__(D.Normal(loc, scale, **self._kwargs), self._event_dim)
@property
def mode(self):
return self.base_dist.mean
@property
def deterministic_sample(self):
return self.mean
class SafeTanhTransform(D.TanhTransform):
"""TanhTransform subclass that ensured that the transformation is numerically invertible."""
def _call(self, x: torch.Tensor) -> torch.Tensor:
return safetanh_noeps(x)
def _inverse(self, y: torch.Tensor) -> torch.Tensor:
return safeatanh_noeps(y)
@property
def inv(self):
inv = None
if self._inv is not None:
inv = self._inv()
if inv is None:
inv = _InverseTransform(self)
if not is_dynamo_compiling():
self._inv = weakref.ref(inv)
return inv
[docs]class NormalParamWrapper(nn.Module):
"""A wrapper for normal distribution parameters.
Args:
operator (nn.Module): operator whose output will be transformed_in in location and scale parameters
scale_mapping (str, optional): positive mapping function to be used with the std.
default = "biased_softplus_1.0" (i.e. softplus map with bias such that fn(0.0) = 1.0)
choices: "softplus", "exp", "relu", "biased_softplus_1";
scale_lb (Number, optional): The minimum value that the variance can take. Default is 1e-4.
Examples:
>>> from torch import nn
>>> import torch
>>> module = nn.Linear(3, 4)
>>> module_normal = NormalParamWrapper(module)
>>> tensor = torch.randn(3)
>>> loc, scale = module_normal(tensor)
>>> print(loc.shape, scale.shape)
torch.Size([2]) torch.Size([2])
>>> assert (scale > 0).all()
>>> # with modules that return more than one tensor
>>> module = nn.LSTM(3, 4)
>>> module_normal = NormalParamWrapper(module)
>>> tensor = torch.randn(4, 2, 3)
>>> loc, scale, others = module_normal(tensor)
>>> print(loc.shape, scale.shape)
torch.Size([4, 2, 2]) torch.Size([4, 2, 2])
>>> assert (scale > 0).all()
"""
def __init__(
self,
operator: nn.Module,
scale_mapping: str = "biased_softplus_1.0",
scale_lb: Number = 1e-4,
) -> None:
warnings.warn(
"The NormalParamWrapper class will be deprecated in v0.7 in favor of :class:`~tensordict.nn.NormalParamExtractor`.",
category=DeprecationWarning,
)
super().__init__()
self.operator = operator
self.scale_mapping = scale_mapping
self.scale_lb = scale_lb
[docs] def forward(self, *tensors: torch.Tensor) -> Tuple[torch.Tensor]:
net_output = self.operator(*tensors)
others = ()
if not isinstance(net_output, torch.Tensor):
net_output, *others = net_output
loc, scale = net_output.chunk(2, -1)
scale = mappings(self.scale_mapping)(scale).clamp_min(self.scale_lb)
return (loc, scale, *others)
[docs]class TruncatedNormal(D.Independent):
"""Implements a Truncated Normal distribution with location scaling.
Location scaling prevents the location to be "too far" from 0, which ultimately
leads to numerically unstable samples and poor gradient computation (e.g. gradient explosion).
In practice, the location is computed according to
.. math::
loc = tanh(loc / upscale) * upscale.
This behavior can be disabled by switching off the tanh_loc parameter (see below).
Args:
loc (torch.Tensor): normal distribution location parameter
scale (torch.Tensor): normal distribution sigma parameter (squared root of variance)
upscale (torch.Tensor or number, optional): 'a' scaling factor in the formula:
.. math::
loc = tanh(loc / upscale) * upscale.
Default is 5.0
min (torch.Tensor or number, optional): minimum value of the distribution. Default = -1.0;
max (torch.Tensor or number, optional): maximum value of the distribution. Default = 1.0;
tanh_loc (bool, optional): if ``True``, the above formula is used for
the location scaling, otherwise the raw value is kept.
Default is ``False``;
"""
num_params: int = 2
base_dist: _TruncatedNormal
arg_constraints = {
"loc": constraints.real,
"scale": constraints.greater_than(1e-6),
}
def __init__(
self,
loc: torch.Tensor,
scale: torch.Tensor,
upscale: Union[torch.Tensor, float] = 5.0,
low: Union[torch.Tensor, float] = -1.0,
high: Union[torch.Tensor, float] = 1.0,
tanh_loc: bool = False,
):
err_msg = "TanhNormal high values must be strictly greater than low values"
if isinstance(high, torch.Tensor) or isinstance(low, torch.Tensor):
if not (high > low).all():
raise RuntimeError(err_msg)
elif isinstance(high, Number) and isinstance(low, Number):
if not high > low:
raise RuntimeError(err_msg)
else:
if not all(high > low):
raise RuntimeError(err_msg)
if isinstance(high, torch.Tensor):
self.non_trivial_max = (high != 1.0).any()
else:
self.non_trivial_max = high != 1.0
if isinstance(low, torch.Tensor):
self.non_trivial_min = (low != -1.0).any()
else:
self.non_trivial_min = low != -1.0
self.tanh_loc = tanh_loc
self.device = loc.device
self.upscale = torch.as_tensor(upscale, device=self.device)
high = torch.as_tensor(high, device=self.device)
low = torch.as_tensor(low, device=self.device)
self.low = low
self.high = high
self.update(loc, scale)
@property
def min(self):
self._warn_minmax()
return self.low
@property
def max(self):
self._warn_minmax()
return self.high
def update(self, loc: torch.Tensor, scale: torch.Tensor) -> None:
if self.tanh_loc:
loc = (loc / self.upscale).tanh() * self.upscale
self.loc = loc
self.scale = scale
base_dist = _TruncatedNormal(
loc,
scale,
a=self.low.expand_as(loc),
b=self.high.expand_as(scale),
device=self.device,
)
super().__init__(base_dist, 1, validate_args=False)
@property
def mode(self):
m = self.base_dist.loc
a = self.base_dist._non_std_a + self.base_dist._dtype_min_gt_0
b = self.base_dist._non_std_b - self.base_dist._dtype_min_gt_0
m = torch.min(torch.stack([m, b], -1), dim=-1)[0]
return torch.max(torch.stack([m, a], -1), dim=-1)[0]
@property
def deterministic_sample(self):
return self.mean
[docs] def log_prob(self, value, **kwargs):
above_or_below = (self.low > value) | (self.high < value)
a = self.base_dist._non_std_a + self.base_dist._dtype_min_gt_0
a = a.expand_as(value)
b = self.base_dist._non_std_b - self.base_dist._dtype_min_gt_0
b = b.expand_as(value)
value = torch.min(torch.stack([value, b], -1), dim=-1)[0]
value = torch.max(torch.stack([value, a], -1), dim=-1)[0]
lp = super().log_prob(value, **kwargs)
if above_or_below.any():
if self.event_shape:
above_or_below = above_or_below.flatten(-len(self.event_shape), -1).any(
-1
)
lp = torch.masked_fill(
lp,
above_or_below.expand_as(lp),
torch.tensor(-float("inf"), device=lp.device, dtype=lp.dtype),
)
return lp
class _PatchedComposeTransform(D.ComposeTransform):
@property
def inv(self):
inv = None
if self._inv is not None:
inv = self._inv()
if inv is None:
inv = _PatchedComposeTransform([p.inv for p in reversed(self.parts)])
if not is_dynamo_compiling():
self._inv = weakref.ref(inv)
inv._inv = weakref.ref(self)
return inv
class _PatchedAffineTransform(D.AffineTransform):
@property
def inv(self):
inv = None
if self._inv is not None:
inv = self._inv()
if inv is None:
inv = _InverseTransform(self)
if not is_dynamo_compiling():
self._inv = weakref.ref(inv)
return inv
[docs]class TanhNormal(FasterTransformedDistribution):
"""Implements a TanhNormal distribution with location scaling.
Location scaling prevents the location to be "too far" from 0 when a
``TanhTransform`` is applied, but ultimately
leads to numerically unstable samples and poor gradient computation
(e.g. gradient explosion).
In practice, with location scaling the location is computed according to
.. math::
loc = tanh(loc / upscale) * upscale.
Args:
loc (torch.Tensor): normal distribution location parameter
scale (torch.Tensor): normal distribution sigma parameter (squared root of variance)
upscale (torch.Tensor or number): 'a' scaling factor in the formula:
.. math::
loc = tanh(loc / upscale) * upscale.
low (torch.Tensor or number, optional): minimum value of the distribution. Default is -1.0;
high (torch.Tensor or number, optional): maximum value of the distribution. Default is 1.0;
event_dims (int, optional): number of dimensions describing the action.
Default is 1. Setting ``event_dims`` to ``0`` will result in a log-probability that has the same shape
as the input, ``1`` will reduce (sum over) the last dimension, ``2`` the last two etc.
tanh_loc (bool, optional): if ``True``, the above formula is used for the location scaling, otherwise the raw
value is kept. Default is ``False``;
safe_tanh (bool, optional): if ``True``, the Tanh transform is done "safely", to avoid numerical overflows.
This will currently break with :func:`torch.compile`.
"""
arg_constraints = {
"loc": constraints.real,
"scale": constraints.greater_than(1e-6),
}
num_params = 2
def __init__(
self,
loc: torch.Tensor,
scale: torch.Tensor,
upscale: Union[torch.Tensor, Number] = 5.0,
low: Union[torch.Tensor, Number] = -1.0,
high: Union[torch.Tensor, Number] = 1.0,
event_dims: int | None = None,
tanh_loc: bool = False,
safe_tanh: bool = True,
**kwargs,
):
if not isinstance(loc, torch.Tensor):
loc = torch.as_tensor(loc, dtype=torch.get_default_dtype())
if not isinstance(scale, torch.Tensor):
scale = torch.as_tensor(scale, dtype=torch.get_default_dtype())
if event_dims is None:
event_dims = min(1, loc.ndim)
err_msg = "TanhNormal high values must be strictly greater than low values"
if isinstance(high, torch.Tensor) or isinstance(low, torch.Tensor):
if not (high > low).all():
raise RuntimeError(err_msg)
elif isinstance(high, Number) and isinstance(low, Number):
if not high > low:
raise RuntimeError(err_msg)
else:
if not all(high > low):
raise RuntimeError(err_msg)
high = torch.as_tensor(high, device=loc.device)
low = torch.as_tensor(low, device=loc.device)
self.non_trivial_max = (high != 1.0).any()
self.non_trivial_min = (low != -1.0).any()
self.tanh_loc = tanh_loc
self._event_dims = event_dims
self.device = loc.device
self.upscale = (
upscale
if not isinstance(upscale, torch.Tensor)
else upscale.to(self.device)
)
if isinstance(high, torch.Tensor):
high = high.to(loc.device)
if isinstance(low, torch.Tensor):
low = low.to(loc.device)
self.low = low
self.high = high
if safe_tanh:
if is_dynamo_compiling():
_err_compile_safetanh()
t = SafeTanhTransform()
else:
t = D.TanhTransform()
# t = D.TanhTransform()
if is_dynamo_compiling() or (self.non_trivial_max or self.non_trivial_min):
t = _PatchedComposeTransform(
[
t,
_PatchedAffineTransform(
loc=(high + low) / 2, scale=(high - low) / 2
),
]
)
self._t = t
self.update(loc, scale)
@property
def min(self):
self._warn_minmax()
return self.low
@property
def max(self):
self._warn_minmax()
return self.high
def update(self, loc: torch.Tensor, scale: torch.Tensor) -> None:
if self.tanh_loc:
loc = (loc / self.upscale).tanh() * self.upscale
# loc must be rescaled if tanh_loc
if is_dynamo_compiling() or (self.non_trivial_max or self.non_trivial_min):
loc = loc + (self.high - self.low) / 2 + self.low
self.loc = loc
self.scale = scale
if (
hasattr(self, "base_dist")
and (self.root_dist.loc.shape == self.loc.shape)
and (self.root_dist.scale.shape == self.scale.shape)
):
self.root_dist.loc = self.loc
self.root_dist.scale = self.scale
else:
if self._event_dims > 0:
base = D.Independent(D.Normal(self.loc, self.scale), self._event_dims)
super().__init__(base, self._t)
else:
base = D.Normal(self.loc, self.scale)
super().__init__(base, self._t)
@property
def support(self):
return D.constraints.real()
@property
def root_dist(self):
bd = self
while hasattr(bd, "base_dist"):
bd = bd.base_dist
return bd
@property
def mode(self):
raise RuntimeError(
f"The distribution {type(self).__name__} has not analytical mode. "
f"Use ExplorationMode.DETERMINISTIC to get a deterministic sample from it."
)
@property
def deterministic_sample(self):
m = self.root_dist.mean
for t in self.transforms:
m = t(m)
return m
[docs] @torch.enable_grad()
def get_mode(self):
"""Computes an estimation of the mode using the Adam optimizer."""
# Get starting point
m = self.sample((1000,)).mean(0)
m = torch.nn.Parameter(m.clamp(self.low, self.high).detach())
optim = torch.optim.Adam((m,), lr=1e-2)
self_copy = type(self)(
loc=self.loc.detach(),
scale=self.scale.detach(),
low=self.low.detach(),
high=self.high.detach(),
event_dims=self._event_dims,
upscale=self.upscale,
tanh_loc=False,
)
for _ in range(200):
lp = -self_copy.log_prob(m)
lp.mean().backward()
mc = m.clone().detach()
m.grad.clamp_max_(1)
optim.step()
optim.zero_grad()
m.data.clamp_(self_copy.low, self_copy.high)
nans = m.isnan()
if nans.any():
m.data = torch.where(nans, mc, m.data)
if (m - mc).norm() < 1e-3:
break
return m.detach()
@property
def mean(self):
raise NotImplementedError(
f"{type(self).__name__} does not have a closed form formula for the average. "
"Am estimate of this value can be computed using dist.sample((N,)).mean(dim=0), "
"where N is a large number of samples."
)
def uniform_sample_tanhnormal(dist: TanhNormal, size=None) -> torch.Tensor:
"""Defines what uniform sampling looks like for a TanhNormal distribution.
Args:
dist (TanhNormal): distribution defining the space where the sampling should occur.
size (torch.Size): batch-size of the output tensor
Returns:
a tensor sampled uniformly in the boundaries defined by the input distribution.
"""
if size is None:
size = torch.Size([])
return torch.rand_like(dist.sample(size)) * (dist.max - dist.min) + dist.min
[docs]class Delta(D.Distribution):
"""Delta distribution.
Args:
param (torch.Tensor): parameter of the delta distribution;
atol (number, optional): absolute tolerance to consider that a tensor matches the distribution parameter;
Default is 1e-6
rtol (number, optional): relative tolerance to consider that a tensor matches the distribution parameter;
Default is 1e-6
batch_shape (torch.Size, optional): batch shape;
event_shape (torch.Size, optional): shape of the outcome.
"""
arg_constraints: Dict = {}
def __init__(
self,
param: torch.Tensor,
atol: float = 1e-6,
rtol: float = 1e-6,
batch_shape: Union[torch.Size, Sequence[int]] = None,
event_shape: Union[torch.Size, Sequence[int]] = None,
):
if batch_shape is None:
batch_shape = torch.Size([])
if event_shape is None:
event_shape = torch.Size([])
self.update(param)
self.atol = atol
self.rtol = rtol
if not len(batch_shape) and not len(event_shape):
batch_shape = param.shape[:-1]
event_shape = param.shape[-1:]
super().__init__(batch_shape=batch_shape, event_shape=event_shape)
def update(self, param):
self.param = param
def _is_equal(self, value: torch.Tensor) -> torch.Tensor:
param = self.param.expand_as(value)
is_equal = abs(value - param) < self.atol + self.rtol * abs(param)
for i in range(-1, -len(self.event_shape) - 1, -1):
is_equal = is_equal.all(i)
return is_equal
[docs] def log_prob(self, value: torch.Tensor) -> torch.Tensor:
is_equal = self._is_equal(value)
out = torch.zeros_like(is_equal, dtype=value.dtype)
out.masked_fill_(is_equal, np.inf)
out.masked_fill_(~is_equal, -np.inf)
return out
[docs] @torch.no_grad()
def sample(self, size=None) -> torch.Tensor:
if size is None:
size = torch.Size([])
return self.param.expand(*size, *self.param.shape)
[docs] def rsample(self, size=None) -> torch.Tensor:
if size is None:
size = torch.Size([])
return self.param.expand(*size, *self.param.shape)
@property
def mode(self) -> torch.Tensor:
return self.param
@property
def deterministic_sample(self):
return self.mean
@property
def mean(self) -> torch.Tensor:
return self.param
[docs]class TanhDelta(FasterTransformedDistribution):
"""Implements a Tanh transformed_in Delta distribution.
Args:
param (torch.Tensor): parameter of the delta distribution;
low (torch.Tensor or number, optional): minimum value of the distribution. Default is -1.0;
high (torch.Tensor or number, optional): maximum value of the distribution. Default is 1.0;
event_dims (int, optional): number of dimensions describing the action.
Default is 1;
atol (number, optional): absolute tolerance to consider that a tensor matches the distribution parameter;
Default is 1e-6
rtol (number, optional): relative tolerance to consider that a tensor matches the distribution parameter;
Default is 1e-6
batch_shape (torch.Size, optional): batch shape;
event_shape (torch.Size, optional): shape of the outcome;
"""
arg_constraints = {
"loc": constraints.real,
}
def __init__(
self,
param: torch.Tensor,
low: Union[torch.Tensor, float] = -1.0,
high: Union[torch.Tensor, float] = 1.0,
event_dims: int = 1,
atol: float = 1e-6,
rtol: float = 1e-6,
):
minmax_msg = "high value has been found to be equal or less than low value"
if isinstance(high, torch.Tensor) or isinstance(low, torch.Tensor):
if not (high > low).all():
raise ValueError(minmax_msg)
elif isinstance(high, Number) and isinstance(low, Number):
if high <= low:
raise ValueError(minmax_msg)
else:
if not all(high > low):
raise ValueError(minmax_msg)
t = SafeTanhTransform()
non_trivial_min = (isinstance(low, torch.Tensor) and (low != -1.0).any()) or (
not isinstance(low, torch.Tensor) and low != -1.0
)
non_trivial_max = (isinstance(high, torch.Tensor) and (high != 1.0).any()) or (
not isinstance(high, torch.Tensor) and high != 1.0
)
self.non_trivial = non_trivial_min or non_trivial_max
self.low = _cast_device(low, param.device)
self.high = _cast_device(high, param.device)
loc = self.update(param)
if self.non_trivial:
t = _PatchedComposeTransform(
[
t,
_PatchedAffineTransform(
loc=(self.high + self.low) / 2, scale=(self.high - self.low) / 2
),
]
)
event_shape = param.shape[-event_dims:]
batch_shape = param.shape[:-event_dims]
base = Delta(
loc,
atol=atol,
rtol=rtol,
batch_shape=batch_shape,
event_shape=event_shape,
)
super().__init__(base, t)
@property
def min(self):
self._warn_minmax()
return self.low
@property
def max(self):
self._warn_minmax()
return self.high
def update(self, net_output: torch.Tensor) -> Optional[torch.Tensor]:
loc = net_output
if self.non_trivial:
device = loc.device
shift = _cast_device(self.high - self.low, device)
loc = loc + shift / 2 + _cast_device(self.low, device)
if hasattr(self, "base_dist"):
self.base_dist.update(loc)
else:
return loc
@property
def mode(self) -> torch.Tensor:
mode = self.base_dist.param
for t in self.transforms:
mode = t(mode)
return mode
@property
def deterministic_sample(self):
return self.mode
@property
def mean(self) -> torch.Tensor:
raise AttributeError("TanhDelta mean has not analytical form.")
def _uniform_sample_delta(dist: Delta, size=None) -> torch.Tensor:
if size is None:
size = torch.Size([])
return torch.randn_like(dist.sample(size))
uniform_sample_delta = _uniform_sample_delta
def _err_compile_safetanh():
raise RuntimeError(
"safe_tanh=True in TanhNormal is not compatible with torch.compile. To deactivate it, pass"
"safe_tanh=False. "
"If you are using a ProbabilisticTensorDictModule, this can be done via "
"`distribution_kwargs={'safe_tanh': False}`. "
"See https://github.com/pytorch/pytorch/issues/133529 for more details."
)
_warn_compile_safetanh = assume_constant_result(_err_compile_safetanh)