mean-conc-beta 0.0.1__py3-none-any.whl

mean_conc_beta/__init__.py

@@ -0,0 +1 @@
+from mean_conc_beta.mean_conc_beta import Beta, exists

mean_conc_beta/mean_conc_beta.py

@@ -0,0 +1,157 @@
+from __future__ import annotations
+import math
+
+import torch
+from torch import Tensor, Size
+import torch.nn.functional as F
+from torch.nn import Module
+from torch.distributions import (
+    Distribution,
+    Beta as _Beta,
+    TransformedDistribution,
+    AffineTransform
+)
+
+# helpers
+
+def exists(v):
+    return v is not None
+
+def default(v, d):
+    return v if exists(v) else d
+
+# beta distribution policy - unimodal mean-concentration reparameterization on (-1, 1),
+# an affine shift of a unit-interval beta (y = 2x - 1)
+
+class Beta(Module):
+    def __init__(
+        self,
+        pos_fn = 'exp',
+        init_conc = 10.,
+        min_conc = 0.,
+        eps = 1e-5,
+        detach_unimodal = True,
+        clamp_exp = (-10., 10.)
+    ):
+        super().__init__()
+        assert pos_fn in ('exp', 'softplus'), f'pos_fn must be either exp or softplus, got {pos_fn}'
+        assert min_conc >= 0., f'min_conc must be non-negative, got {min_conc}'
+        assert init_conc > min_conc, f'init_conc ({init_conc}) must be greater than min_conc ({min_conc})'
+
+        self.pos_fn = pos_fn
+        self.init_conc = init_conc
+        self.min_conc = min_conc
+        self.eps = eps
+        self.detach_unimodal = detach_unimodal
+
+        if isinstance(clamp_exp, (int, float)):
+            clamp_exp = (-float(clamp_exp), float(clamp_exp))
+
+        self.clamp_exp = clamp_exp
+
+        # raw offset so concentration at raw_conc = 0 is exactly init_conc
+
+        self.raw_init_conc = math.log(init_conc - min_conc) if pos_fn == 'exp' else math.log(math.expm1(init_conc - min_conc))
+
+    @property
+    def has_rsample(self) -> bool:
+        return True
+
+    def to_dist(self, params_or_dist: Tensor | Distribution) -> Distribution:
+        return params_or_dist if isinstance(params_or_dist, Distribution) else self(params_or_dist)
+
+    def concentration(
+        self,
+        raw_conc: Tensor
+    ) -> Tensor:
+        if self.pos_fn == 'exp':
+            if exists(self.clamp_exp):
+                min_val, max_val = self.clamp_exp
+                raw_conc = raw_conc.clamp(min = min_val, max = max_val)
+
+            return (raw_conc + self.raw_init_conc).exp() + self.min_conc
+
+        return F.softplus(raw_conc + self.raw_init_conc) + self.min_conc
+
+    def mean(
+        self,
+        params: Tensor
+    ) -> Tensor:
+        raw_mean, _ = params.unbind(dim = -1)
+        return raw_mean.tanh().clamp(min = -1. + self.eps, max = 1. - self.eps)
+
+    def mode(
+        self,
+        params_or_dist: Tensor | Distribution
+    ) -> Tensor:
+        dist = self.to_dist(params_or_dist)
+        base_dist = getattr(dist, 'base_dist', dist)
+        return base_dist.mode * 2. - 1.
+
+    def entropy(
+        self,
+        params_or_dist: Tensor | Distribution,
+        sum_action_dim = True
+    ) -> Tensor:
+        # shifted beta entropy = base entropy + log(2), the affine jacobian
+
+        dist = self.to_dist(params_or_dist)
+        base_dist = getattr(dist, 'base_dist', dist)
+        entropy = base_dist.entropy() + math.log(2.)
+        return entropy.sum(dim = -1) if sum_action_dim else entropy
+
+    def log_prob(
+        self,
+        params_or_dist: Tensor | Distribution,
+        action: Tensor,
+        sum_action_dim = True,
+        eps = None
+    ) -> Tensor:
+        eps = default(eps, self.eps)
+        action = action.clamp(min = -1. + eps, max = 1. - eps)
+        dist = self.to_dist(params_or_dist)
+        log_prob = dist.log_prob(action)
+        return log_prob.sum(dim = -1) if sum_action_dim else log_prob
+
+    def sample(
+        self,
+        params: Tensor,
+        sample_shape = ()
+    ) -> Tensor:
+        sample_shape = Size(sample_shape)
+        return self(params).sample(sample_shape)
+
+    def rsample(
+        self,
+        params: Tensor,
+        sample_shape = ()
+    ) -> Tensor:
+        sample_shape = Size(sample_shape)
+        return self(params).rsample(sample_shape)
+
+    def forward(
+        self,
+        params: Tensor
+    ) -> TransformedDistribution:
+        raw_mean, raw_conc = params.unbind(dim = -1)
+
+        # map (-1, 1) mean onto unit interval
+
+        mean = self.mean(params)
+        unit_mean = (mean + 1.) / 2.
+
+        conc = self.concentration(raw_conc)
+
+        # keep the beta unimodal without changing its mean
+
+        min_unit_mean = torch.minimum(unit_mean, 1. - unit_mean).clamp(min = self.eps / 2.)
+
+        if self.detach_unimodal:
+            min_unit_mean = min_unit_mean.detach()
+
+        conc = conc + 1. / min_unit_mean
+
+        alpha = unit_mean * conc
+        beta = (1. - unit_mean) * conc
+
+        return TransformedDistribution(_Beta(alpha, beta), AffineTransform(loc = -1., scale = 2.))

mean_conc_beta-0.0.1.dist-info/METADATA

@@ -0,0 +1,112 @@
+Metadata-Version: 2.5
+Name: mean-conc-beta
+Version: 0.0.1
+Summary: Beta distribution with mean / concentration parameterization, for bounded continuous action spaces in RL
+Project-URL: Homepage, https://pypi.org/project/mean-conc-beta/
+Project-URL: Repository, https://github.com/lucidrains/mean-conc-beta
+Author-email: Phil Wang <lucidrains@gmail.com>
+License: MIT License
+        
+        Copyright (c) 2026 Phil Wang
+        
+        Permission is hereby granted, free of charge, to any person obtaining a copy
+        of this software and associated documentation files (the "Software"), to deal
+        in the Software without restriction, including without limitation the rights
+        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+        copies of the Software, and to permit persons to whom the Software is
+        furnished to do so, subject to the following conditions:
+        
+        The above copyright notice and this permission notice shall be included in all
+        copies or substantial portions of the Software.
+        
+        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+        SOFTWARE.
+License-File: LICENSE
+Keywords: artificial intelligence,continuous control,deep learning,probability distribution,reinforcement learning
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Requires-Dist: torch>=2.0
+Provides-Extra: test
+Requires-Dist: gymnasium; extra == 'test'
+Requires-Dist: mujoco; extra == 'test'
+Requires-Dist: numpy; extra == 'test'
+Requires-Dist: pytest; extra == 'test'
+Description-Content-Type: text/markdown
+
+## mean-conc-beta
+
+Beta distribution parameterized by mean and concentration for bounded continuous action spaces in reinforcement learning.
+
+## Install
+
+```bash
+$ pip install mean-conc-beta
+```
+
+## Usage
+
+```python
+import torch
+import torch.nn.functional as F
+from mean_conc_beta import Beta
+
+beta = Beta()
+
+# network output: (batch, num_actions, 2) for raw mean and concentration
+
+params = torch.randn(16, 4, 2, requires_grad = True)
+
+# distribution on (-1, 1)
+
+dist = beta(params)
+
+# sample actions
+
+actions = dist.sample()
+actions_reparam = dist.rsample()
+
+# log prob and entropy
+
+log_prob = beta.log_prob(dist, actions)
+entropy = beta.entropy(dist)
+
+# behavior cloning with mse loss on mean
+
+expert_actions = torch.rand(16, 4)
+
+pred_mean = beta.mean(params)
+
+bc_loss = F.mse_loss(pred_mean, expert_actions)
+bc_loss.backward()
+```
+
+## Citations
+
+```bibtex
+@article{Ferrari2004BetaRF,
+    title   = {Beta Regression for Modelling Rates and Proportions},
+    author  = {Silvia L. P. Ferrari and Francisco Cribari-Neto},
+    journal = {Journal of Applied Statistics},
+    year    = {2004},
+    volume  = {31},
+    pages   = {799 - 815}
+}
+```
+
+```bibtex
+@inproceedings{Chou2017TheBP,
+    title   = {The Beta Policy for Continuous Reinforcement Learning},
+    author  = {Po-Wei Chou and Daniel Maturana and Sebastian Scherer},
+    booktitle = {International Conference on Machine Learning},
+    year    = {2017}
+}
+```

mean_conc_beta-0.0.1.dist-info/RECORD

@@ -0,0 +1,6 @@
+mean_conc_beta/__init__.py,sha256=UAdOwHDMhSedUGbytcdXmR891d56gbZ0Z6k5vqF9GN0,55
+mean_conc_beta/mean_conc_beta.py,sha256=Hy6jUkFdLmHq3P6diFxXQ63Iz4eyi_lxJ8jc2VtH4xw,4653
+mean_conc_beta-0.0.1.dist-info/METADATA,sha256=6DV7ahvyXqYAzH2-a6X4mM7_3r9RnFzrANp0etn4IJI,3672
+mean_conc_beta-0.0.1.dist-info/WHEEL,sha256=zOwg4jB6zX2kU910N-cMawjivD6tO8NEWvE12je1bVk,87
+mean_conc_beta-0.0.1.dist-info/licenses/LICENSE,sha256=e6AOF7Z8EFdK3IdcL0x0fLw4cY7Q0d0kNR0o0TmBewM,1066
+mean_conc_beta-0.0.1.dist-info/RECORD,,

mean_conc_beta-0.0.1.dist-info/WHEEL

@@ -0,0 +1,4 @@
+Wheel-Version: 1.0
+Generator: hatchling 1.32.0
+Root-Is-Purelib: true
+Tag: py3-none-any

mean_conc_beta-0.0.1.dist-info/licenses/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Phil Wang
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.