from typing import Any, Dict, List, Optional, Tuple, Union from enum import Enum import gym import torch as th from torch import nn from torch.distributions import Normal, Independent, MultivariateNormal from math import pi from stable_baselines3.common.preprocessing import get_action_dim from stable_baselines3.common.distributions import sum_independent_dims from stable_baselines3.common.distributions import Distribution as SB3_Distribution from stable_baselines3.common.distributions import ( BernoulliDistribution, CategoricalDistribution, MultiCategoricalDistribution, # StateDependentNoiseDistribution, ) from stable_baselines3.common.distributions import DiagGaussianDistribution from ..misc.tensor_ops import fill_triangular from ..misc.tanhBijector import TanhBijector # TODO: Integrate and Test what I currently have before adding more complexity # TODO: Support Squashed Dists (tanh) # TODO: Contextual Cov # TODO: - Hybrid # TODO: Contextual SDE (Scalar + Diag + Full) # TODO: (SqrtInducedCov (Scalar + Diag + Full)) class Strength(Enum): NONE = 0 SCALAR = 1 DIAG = 2 FULL = 3 class ParametrizationType(Enum): NONE = 0 CHOL = 1 SPHERICAL_CHOL = 2 # Not (yet?) implemented: # GIVENS = 3 # NNLN_EIGEN = 4 class EnforcePositiveType(Enum): # This need to be implemented in this ugly fashion, # because cloudpickle does not like more complex enums NONE = 0 SOFTPLUS = 1 ABS = 2 RELU = 3 LOG = 4 def apply(self, x): # aaaaaa return [nn.Identity(), nn.Softplus(beta=1, threshold=20), th.abs, nn.ReLU(inplace=False), th.log][self.value](x) class ProbSquashingType(Enum): NONE = 0 TANH = 1 def apply(self, x): return [nn.Identity(), th.tanh][self.value](x) def apply_inv(self, x): return [nn.Identity(), TanhBijector.inverse][self.value](x) def cast_to_enum(inp, Class): if isinstance(inp, Enum): return inp else: return Class[inp] def get_legal_setups(allowedEPTs=None, allowedParStrength=None, allowedCovStrength=None, allowedPTs=None, allowedPSTs=None): allowedEPTs = allowedEPTs or EnforcePositiveType allowedParStrength = allowedParStrength or Strength allowedCovStrength = allowedCovStrength or Strength allowedPTs = allowedPTs or ParametrizationType allowedPSTs = allowedPSTs or ProbSquashingType for ps in allowedParStrength: for cs in allowedCovStrength: if ps.value > cs.value: continue if cs == Strength.NONE: yield (ps, cs, EnforcePositiveType.NONE, ParametrizationType.NONE) else: for ept in allowedEPTs: if cs == Strength.FULL: for pt in allowedPTs: if pt != ParametrizationType.NONE: yield (ps, cs, ept, pt) else: yield (ps, cs, ept, ParametrizationType.NONE) def make_proba_distribution( action_space: gym.spaces.Space, use_sde: bool = False, dist_kwargs: Optional[Dict[str, Any]] = None ) -> SB3_Distribution: """ Return an instance of Distribution for the correct type of action space :param action_space: the input action space :param use_sde: Force the use of StateDependentNoiseDistribution instead of DiagGaussianDistribution :param dist_kwargs: Keyword arguments to pass to the probability distribution :return: the appropriate Distribution object """ if dist_kwargs is None: dist_kwargs = {} dist_kwargs['use_sde'] = use_sde if isinstance(action_space, gym.spaces.Box): assert len( action_space.shape) == 1, "Error: the action space must be a vector" return UniversalGaussianDistribution(get_action_dim(action_space), **dist_kwargs) elif isinstance(action_space, gym.spaces.Discrete): return CategoricalDistribution(action_space.n, **dist_kwargs) elif isinstance(action_space, gym.spaces.MultiDiscrete): return MultiCategoricalDistribution(action_space.nvec, **dist_kwargs) elif isinstance(action_space, gym.spaces.MultiBinary): return BernoulliDistribution(action_space.n, **dist_kwargs) else: raise NotImplementedError( "Error: probability distribution, not implemented for action space" f"of type {type(action_space)}." " Must be of type Gym Spaces: Box, Discrete, MultiDiscrete or MultiBinary." ) class UniversalGaussianDistribution(SB3_Distribution): """ Gaussian distribution with configurable covariance matrix shape and optional contextual parametrization mechanism, for continuous actions. :param action_dim: Dimension of the action space. """ def __init__(self, action_dim: int, use_sde: bool = False, neural_strength: Strength = Strength.DIAG, cov_strength: Strength = Strength.DIAG, parameterization_type: ParametrizationType = ParametrizationType.NONE, enforce_positive_type: EnforcePositiveType = EnforcePositiveType.ABS, prob_squashing_type: ProbSquashingType = ProbSquashingType.NONE, epsilon=1e-6): super(UniversalGaussianDistribution, self).__init__() self.action_dim = action_dim self.par_strength = cast_to_enum(neural_strength, Strength) self.cov_strength = cast_to_enum(cov_strength, Strength) self.par_type = cast_to_enum( parameterization_type, ParametrizationType) self.enforce_positive_type = cast_to_enum( enforce_positive_type, EnforcePositiveType) self.prob_squashing_type = cast_to_enum( prob_squashing_type, ProbSquashingType) self.epsilon = epsilon self.distribution = None self.gaussian_actions = None if use_sde: raise Exception('SDE is not yet implemented') assert (self.par_type != ParametrizationType.NONE) == ( self.cov_strength == Strength.FULL), 'You should set an ParameterizationType iff the cov-strength is full' if self.par_type == ParametrizationType.SPHERICAL_CHOL and self.enforce_positive_type == EnforcePositiveType.NONE: raise Exception( 'You need to specify an enforce_positive_type for spherical_cholesky') def new_dist_like_me(self, mean: th.Tensor, chol: th.Tensor): p = self.distribution if isinstance(p, Independent): if p.stddev.shape != chol.shape: chol = th.diagonal(chol, dim1=1, dim2=2) np = Independent(Normal(mean, chol), 1) elif isinstance(p, MultivariateNormal): np = MultivariateNormal(mean, scale_tril=chol) new = UniversalGaussianDistribution(self.action_dim, neural_strength=self.par_strength, cov_strength=self.cov_strength, parameterization_type=self.par_type, enforce_positive_type=self.enforce_positive_type, prob_squashing_type=self.prob_squashing_type) new.distribution = np return new def proba_distribution_net(self, latent_dim: int, latent_sde_dim: int, std_init: float = 0.0) -> Tuple[nn.Module, nn.Module]: """ Create the layers and parameter that represent the distribution: one output will be the mean of the Gaussian, the other parameter will be the standard deviation :param latent_dim: Dimension of the last layer of the policy (before the action layer) :param std_init: Initial value for the standard deviation :return: We return two nn.Modules (mean, chol). chol can be a vector if the full chol would be a diagonal. """ assert std_init >= 0.0, "std can not be initialized to a negative value." # TODO: Implement SDE self.latent_sde_dim = latent_sde_dim mean_actions = nn.Linear(latent_dim, self.action_dim) chol = CholNet(latent_dim, self.action_dim, std_init, self.par_strength, self.cov_strength, self.par_type, self.enforce_positive_type, self.prob_squashing_type) return mean_actions, chol def proba_distribution_from_sqrt(self, mean_actions: th.Tensor, cov_sqrt: th.Tensor, latent_pi: nn.Module) -> "UniversalGaussianDistribution": """ Create the distribution given its parameters (mean, cov_sqrt) :param mean_actions: :param cov_sqrt: :return: """ cov = cov_sqrt.T @ cov_sqrt chol = th.linalg.cholesky(cov) self.cov_sqrt = cov_sqrt return self.proba_distribution(mean_actions, chol, latent_pi) def proba_distribution(self, mean_actions: th.Tensor, chol: th.Tensor, latent_pi: nn.Module) -> "UniversalGaussianDistribution": """ Create the distribution given its parameters (mean, chol) :param mean_actions: :param chol: :return: """ # TODO: latent_pi is for SDE, implement. if self.cov_strength in [Strength.NONE, Strength.SCALAR, Strength.DIAG]: self.distribution = Independent(Normal(mean_actions, chol), 1) elif self.cov_strength in [Strength.FULL]: self.distribution = MultivariateNormal( mean_actions, scale_tril=chol) if self.distribution == None: raise Exception('Unable to create torch distribution') return self def log_prob(self, actions: th.Tensor, gaussian_actions: Optional[th.Tensor] = None) -> th.Tensor: """ Get the log probabilities of actions according to the distribution. Note that you must first call the ``proba_distribution()`` method. :param actions: :return: """ if self.prob_squashing_type == ProbSquashingType.NONE: log_prob = self.distribution.log_prob(actions) return log_prob if gaussian_actions is None: # It will be clipped to avoid NaN when inversing tanh gaussian_actions = self.prob_squashing_type.apply_inv(actions) log_prob = self.distribution.log_prob(gaussian_actions) if self.prob_squashing_type == ProbSquashingType.TANH: log_prob -= th.sum(th.log(1 - actions ** 2 + self.epsilon), dim=1) return log_prob raise Exception() def entropy(self) -> th.Tensor: # TODO: This will return incorrect results when using prob-squashing return self.distribution.entropy() def sample(self) -> th.Tensor: # Reparametrization trick to pass gradients sample = self.distribution.rsample() self.gaussian_actions = sample return self.prob_squashing_type.apply(sample) def mode(self) -> th.Tensor: mode = self.distribution.mean self.gaussian_actions = mode return self.prob_squashing_type.apply(mode) def actions_from_params(self, mean_actions: th.Tensor, log_std: th.Tensor, deterministic: bool = False, latent_pi=None) -> th.Tensor: # Update the proba distribution self.proba_distribution(mean_actions, log_std, latent_pi=latent_pi) return self.get_actions(deterministic=deterministic) def log_prob_from_params(self, mean_actions: th.Tensor, log_std: th.Tensor) -> Tuple[th.Tensor, th.Tensor]: """ Compute the log probability of taking an action given the distribution parameters. :param mean_actions: :param log_std: :return: """ actions = self.actions_from_params(mean_actions, log_std) log_prob = self.log_prob(actions, self.gaussian_actions) return actions, log_prob class CholNet(nn.Module): def __init__(self, latent_dim: int, action_dim: int, std_init: float, par_strength: Strength, cov_strength: Strength, par_type: ParametrizationType, enforce_positive_type: EnforcePositiveType, prob_squashing_type: ProbSquashingType): super().__init__() self.latent_dim = latent_dim self.action_dim = action_dim self.par_strength = par_strength self.cov_strength = cov_strength self.par_type = par_type self.enforce_positive_type = enforce_positive_type self.prob_squashing_type = prob_squashing_type self._flat_chol_len = action_dim * (action_dim + 1) // 2 # Yes, this is ugly. # But I don't know how this mess could be elegantly abstracted away... if self.par_strength == Strength.NONE: if self.cov_strength == Strength.NONE: self.chol = th.ones(self.action_dim) * std_init elif self.cov_strength == Strength.SCALAR: self.param = nn.Parameter( th.Tensor([std_init]), requires_grad=True) elif self.cov_strength == Strength.DIAG: self.params = nn.Parameter( th.ones(self.action_dim) * std_init, requires_grad=True) elif self.cov_strength == Strength.FULL: # TODO: Init Off-axis differently? self.params = nn.Parameter( th.ones(self._full_params_len) * std_init, requires_grad=True) elif self.par_strength == self.cov_strength: if self.par_strength == Strength.SCALAR: self.std = nn.Linear(latent_dim, 1) elif self.par_strength == Strength.DIAG: self.diag_chol = nn.Linear(latent_dim, self.action_dim) elif self.par_strength == Strength.FULL: self.params = nn.Linear(latent_dim, self._full_params_len) elif self.par_strength.value > self.cov_strength.value: raise Exception( 'The parameterization can not be stronger than the actual covariance.') else: if self.par_strength == Strength.SCALAR and self.cov_strength == Strength.DIAG: self.factor = nn.Linear(latent_dim, 1) self.param = nn.Parameter( th.ones(self.action_dim), requires_grad=True) elif self.par_strength == Strength.DIAG and self.cov_strength == Strength.FULL: if self.enforce_positive_type == EnforcePositiveType.NONE: raise Exception( 'For Hybrid[Diag=>Full] enforce_positive_type has to be not NONE. Otherwise required SPD-contraint can not be ensured for cov.') self.stds = nn.Linear(latent_dim, self.action_dim) self.padder = th.nn.ZeroPad2d((0, 1, 1, 0)) # TODO: Init Non-zero? self.params = nn.Parameter( th.ones(self._full_params_len - self.action_dim) * 0, requires_grad=True) elif self.par_strength == Strength.SCALAR and self.cov_strength == Strength.FULL: self.factor = nn.Linear(latent_dim, 1) # TODO: Init Off-axis differently? self.params = nn.Parameter( th.ones(self._full_params_len) * std_init, requires_grad=True) else: raise Exception("This Exception can't happen") def forward(self, x: th.Tensor) -> th.Tensor: # Ugly mess pt.2: if self.par_strength == Strength.NONE: if self.cov_strength == Strength.NONE: return self.chol elif self.cov_strength == Strength.SCALAR: return self._ensure_positive_func( th.ones(self.action_dim) * self.param[0]) elif self.cov_strength == Strength.DIAG: return self._ensure_positive_func(self.params) elif self.cov_strength == Strength.FULL: return self._parameterize_full(self.params) elif self.par_strength == self.cov_strength: if self.par_strength == Strength.SCALAR: std = self.std(x) diag_chol = th.ones(self.action_dim) * std return self._ensure_positive_func(diag_chol) elif self.par_strength == Strength.DIAG: diag_chol = self.diag_chol(x) return self._ensure_positive_func(diag_chol) elif self.par_strength == Strength.FULL: params = self.params(x) return self._parameterize_full(params) else: if self.par_strength == Strength.SCALAR and self.cov_strength == Strength.DIAG: factor = self.factor(x)[0] diag_chol = self._ensure_positive_func( self.param * factor) return diag_chol elif self.par_strength == Strength.DIAG and self.cov_strength == Strength.FULL: # TODO: Maybe possible to improve speed and stability by making conversion from pearson correlation + stds to cov in cholesky-form. stds = self._ensure_positive_func(self.stds(x)) smol = self._parameterize_full(self.params) big = self.padder(smol) pearson_cor_chol = big + th.eye(stds.shape[-1]) pearson_cor = (pearson_cor_chol.T @ pearson_cor_chol) if len(stds.shape) > 1: # batched operation, we need to expand pearson_cor = pearson_cor.expand( (stds.shape[0],)+pearson_cor.shape) stds = stds.unsqueeze(2) cov = stds.mT * pearson_cor * stds chol = th.linalg.cholesky(cov) return chol elif self.par_strength == Strength.SCALAR and self.cov_strength == Strength.FULL: # TODO: Maybe possible to improve speed and stability by multiplying with factor in cholesky-form. factor = self._ensure_positive_func(self.factor(x)) par_chol = self._parameterize_full(self.params) cov = (par_chol.T @ par_chol) if len(factor) > 1: factor = factor.unsqueeze(2) cov = cov * factor chol = th.linalg.cholesky(cov) return chol raise Exception() @property def _full_params_len(self): if self.par_type == ParametrizationType.CHOL: return self._flat_chol_len elif self.par_type == ParametrizationType.SPHERICAL_CHOL: return self._flat_chol_len raise Exception() def _parameterize_full(self, params): if self.par_type == ParametrizationType.CHOL: return self._chol_from_flat(params) elif self.par_type == ParametrizationType.SPHERICAL_CHOL: return self._chol_from_flat_sphe_chol(params) raise Exception() def _chol_from_flat(self, flat_chol): chol = fill_triangular(flat_chol) return self._ensure_diagonal_positive(chol) def _chol_from_flat_sphe_chol(self, flat_sphe_chol): pos_flat_sphe_chol = self._ensure_positive_func(flat_sphe_chol) sphe_chol = fill_triangular(pos_flat_sphe_chol) chol = self._chol_from_sphe_chol(sphe_chol) return chol def _chol_from_sphe_chol(self, sphe_chol): # TODO: Test with batched data # TODO: Make efficient more # Note: # We must should ensure: # S[i,1] > 0 where i = 1..n # S[i,j] e (0, pi) where i = 2..n, j = 2..i # We already ensure S > 0 in _chol_from_flat_sphe_chol # We ensure < pi by applying tanh*pi to all applicable elements batch = (len(sphe_chol.shape) == 3) S = sphe_chol n = sphe_chol.shape[-1] L = th.zeros_like(sphe_chol) for i in range(n): t = 1 #s = '' for j in range(i+1): maybe_cos = 1 #s_maybe_cos = '' if i != j and j < n-1 and i < n: if batch: maybe_cos = th.cos(th.tanh(S[:, i, j+1])*pi) else: maybe_cos = th.cos(th.tanh(S[i, j+1])*pi) #s_maybe_cos = 'cos([l_'+str(i+1)+']_'+str(j+2)+')' if batch: L[:, i, j] = S[:, i, 0] * t * maybe_cos else: L[i, j] = S[i, 0] * t * maybe_cos # print('[L_'+str(i+1)+']_'+str(j+1) + # '=[l_'+str(i+1)+']_1'+s+s_maybe_cos) if j <= i and j < n-1 and i < n: if batch: t *= th.sin(th.tanh(S[:, i, j+1])*pi) else: t *= th.sin(th.tanh(S[i, j+1])*pi) #s += 'sin([l_'+str(i+1)+']_'+str(j+2)+')' return L def _ensure_positive_func(self, x): return self.enforce_positive_type.apply(x) def _ensure_diagonal_positive(self, chol): if len(chol.shape) == 1: # If our chol is a vector (representing a diagonal chol) return self._ensure_positive_func(chol) return chol.tril(-1) + self._ensure_positive_func(chol.diagonal(dim1=-2, dim2=-1)).diag_embed() + chol.triu(1) def string(self): # TODO return '' AnyDistribution = Union[SB3_Distribution, UniversalGaussianDistribution]