PriorConditionedAnnealing/priorConditionedAnnealing/pca.py

403 lines
16 KiB
Python

from enum import Enum
import numpy as np
import torch as th
import scipy.spatial
from torch import nn
from stable_baselines3.common.distributions import Distribution as SB3_Distribution
from stable_baselines3.common.distributions import sum_independent_dims
from torch.distributions import Normal, MultivariateNormal
import torch.nn.functional as F
from priorConditionedAnnealing import noise, kernel
from priorConditionedAnnealing.tensor_ops import fill_triangular, fill_triangular_inverse
class Par_Strength(Enum):
SCALAR = 'SCALAR'
DIAG = 'DIAG'
FULL = 'FULL'
CONT_SCALAR = 'CONT_SCALAR'
CONT_DIAG = 'CONT_DIAG'
CONT_HYBRID = 'CONT_HYBRID'
CONT_FULL = 'CONT_FULL'
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=10, threshold=2), th.abs, nn.ReLU(inplace=False), th.log][self.value](x)
class Avaible_Kernel_Funcs(Enum):
RBF = 0
SE = 1
BROWN = 2
PINK = 3
def get_func(self):
# stil aaaaaaaa
return [kernel.rbf, kernel.se, kernel.brown, kernel.pink][self.value]
class Avaible_Noise_Funcs(Enum):
WHITE = 0
PINK = 1
COLOR = 2
PERLIN = 3
HARMONICPERLIN = 4
DIRTYPERLIN = 5
SDE = 6
SHORTPINK = 7
def get_func(self):
# stil aaaaaaaa
return [noise.White_Noise, noise.Pink_Noise, noise.Colored_Noise, noise.Perlin_Noise, noise.Harmonic_Perlin_Noise, noise.Dirty_Perlin_Noise, noise.SDE_Noise, noise.shortPink_Noise][self.value]
def cast_to_enum(inp, Class):
if isinstance(inp, Enum):
return inp
else:
return Class[inp]
def cast_to_kernel(inp):
if callable(inp):
return inp
else:
func, *pars = inp.split('_')
pars = [float(par) for par in pars]
return Avaible_Kernel_Funcs[func].get_func()(*pars)
def cast_to_Noise(Inp, known_shape):
if callable(Inp): # TODO: Allow instantiated?
return Inp(known_shape)
else:
func, *pars = Inp.split('_')
pars = [float(par) for par in pars]
return Avaible_Noise_Funcs[func].get_func()(known_shape, *pars)
class PCA_Distribution(SB3_Distribution):
def __init__(
self,
action_dim: int,
n_envs: int=1,
par_strength: Par_Strength = Par_Strength.CONT_DIAG,
kernel_func=kernel.rbf(),
init_std: float = 1,
cond_noise: float = 0,
window: int = 64,
epsilon: float = 1e-6,
skip_conditioning: bool = False,
temporal_gradient_emission: bool = False,
msqrt_induces_full: bool = False,
Base_Noise=noise.White_Noise,
):
super().__init__()
self.action_dim = action_dim
self.kernel_func = cast_to_kernel(kernel_func)
self.par_strength = cast_to_enum(par_strength, Par_Strength)
self.init_std = init_std
self.cond_noise = cond_noise
self.window = window
self.epsilon = epsilon
self.skip_conditioning = skip_conditioning
self.temporal_gradient_emission = temporal_gradient_emission
self.msqrt_induces_full = msqrt_induces_full
self.base_noise = cast_to_Noise(Base_Noise, (n_envs, action_dim))
assert not (not skip_conditioning and self.is_full()), 'Conditioning full Covariances not yet implemented'
# Premature optimization is the root of all evil
self._build_conditioner()
# *Optimizes it anyways*
def proba_distribution_net(self, latent_dim: int, return_log_std: bool = False):
mu_net = nn.Linear(latent_dim, self.action_dim)
std_net = StdNet(latent_dim, self.action_dim, self.init_std, self.par_strength, self.epsilon, return_log_std)
return mu_net, std_net
def proba_distribution(
self, mean_actions: th.Tensor, std_actions: th.Tensor) -> SB3_Distribution:
if self.is_full():
self.distribution = MultivariateNormal(mean_actions, scale_tril=std_actions, validate_args=not self.msqrt_induces_full)
#self.distribution.scale = th.diagonal(std_actions, dim1=-2, dim2=-1)
self.distribution._mark_mSqrt = self.msqrt_induces_full
else:
self.distribution = Normal(
mean_actions, std_actions)
return self
def log_prob(self, actions: th.Tensor) -> th.Tensor:
return self._log_prob(actions, self.distribution)
def conditioned_log_prob(self, actions: th.Tensor, trajectory: th.Tensor = None) -> th.Tensor:
pi_mean, pi_std = self.distribution.mean.cpu(), self.distribution.scale.cpu()
rho_mean, rho_std = self._conditioning_engine(trajectory, pi_mean, pi_std)
dist = Normal(rho_mean, rho_std)
return self._log_prob(dist)
def _log_prob(self, actions: th.Tensor, dist: Normal):
return sum_independent_dims(dist.log_prob(actions.to(dist.mean.device)))
def entropy(self) -> th.Tensor:
return sum_independent_dims(self.distribution.entropy())
def get_actions(self, deterministic: bool = False, trajectory: th.Tensor = None) -> th.Tensor:
"""
Return actions according to the probability distribution.
:param deterministic:
:return:
"""
if deterministic:
return self.mode()
return self.sample(traj=trajectory)
def sample(self, traj: th.Tensor, f_sigma: float = 1.0, epsilon=None) -> th.Tensor:
assert self.skip_conditioning or type(traj) != type(None), 'A past trajectory has to be supplied if conditinoning is performed'
pi_mean, pi_decomp = self.distribution.mean.cpu(), self.distribution.scale_tril.cpu() if self.is_full() else self.distribution.scale.cpu()
rho_mean, rho_std = self._conditioning_engine(traj, pi_mean, pi_decomp)
rho_std = rho_std * f_sigma
eta = self._get_rigged(pi_mean, pi_decomp,
rho_mean, rho_std,
epsilon)
# reparameterization with rigged samples
if self.is_full():
actions = pi_mean + th.matmul(pi_decomp, eta.unsqueeze(-1)).squeeze(-1)
else:
actions = pi_mean + pi_decomp * eta
self.gaussian_actions = actions
return actions
def is_contextual(self):
return self.par_strength in [Par_Strength.CONT_SCALAR, Par_Strength.CONT_DIAG, Par_Strength.CONT_HYBRID, Par_Strength.CONT_FULL]
def is_full(self):
return self.par_strength in [Par_Strength.FULL, Par_Strength.CONT_FULL]
def _get_rigged(self, pi_mean, pi_std, rho_mean, rho_std, epsilon=None):
# Ugly function to ensure, that the gradients flow as intended for each modus operandi
if not self.temporal_gradient_emission or self.skip_conditioning:
with th.no_grad():
return self._get_emitting_rigged(pi_mean, pi_std, rho_mean, rho_std, epsilon=epsilon).detach()
return self._get_emitting_rigged(pi_mean.detach(), pi_std.detach(), rho_mean, rho_std, epsilon=epsilon)
def _get_emitting_rigged(self, pi_mean, pi_std, rho_mean, rho_std, epsilon=None):
if epsilon == None:
epsilon = self.base_noise(pi_mean.shape)
if self.skip_conditioning:
return epsilon.detach()
Delta = rho_mean - pi_mean
Pi_mu = 1 / pi_std
Pi_sigma = rho_std / pi_std
eta = Pi_mu * Delta + Pi_sigma * epsilon
return eta
def _pad_and_cut_trajectory(self, traj, value=0):
if traj.shape[-2] < self.window:
if traj.shape[-2] == 0:
shape = list(traj.shape)
shape[-2] = 1
traj = th.ones(shape)*value
missing = self.window - traj.shape[-2]
return F.pad(input=traj, pad=(0, 0, missing, 0, 0, 0), value=value)
return traj[:, -self.window:, :]
def _conditioning_engine(self, trajectory, pi_mean, pi_std):
if self.skip_conditioning:
return pi_mean, pi_std
traj = self._pad_and_cut_trajectory(trajectory)
y = th.cat((traj.transpose(-1, -2), pi_mean.unsqueeze(-1).unsqueeze(0).repeat(traj.shape[0], 1, traj.shape[-2])), dim=1)
with th.no_grad():
conditioners = th.Tensor(self._adapt_conditioner(pi_std))
S = th.cholesky_solve(self.Sig12.expand(conditioners.shape[:-1]).unsqueeze(-1), conditioners).squeeze(-1)
rho_std = self.Sig22 - (S @ self.Sig12)
rho_mean = th.einsum('bai,bai->ba', S, y)
return rho_mean, rho_std
def _build_conditioner(self):
# Precomputes the Cholesky decomp of the cov matrix to be used as a pseudoinverse.
# Also precomputes some auxilary stuff for _adapt_conditioner.
w = self.window
Z = np.linspace(0, w, w+1).reshape(-1, 1)
X = np.array([w]).reshape(-1, 1)
Sig11 = self.kernel_func(
Z, Z) + np.diag(np.hstack((np.repeat(self.cond_noise**2, w), 0)))
self.Sig12 = th.Tensor(self.kernel_func(Z, X)).squeeze(-1)
self.Sig22 = th.Tensor(self.kernel_func(
X, X)).squeeze(-1).squeeze(-1)
self.conditioner = np.linalg.cholesky(Sig11)
self.adapt_norm = np.linalg.norm(
self.conditioner[-1, :][:-1], axis=-1)**2
def _adapt_conditioner(self, pi_std):
# We can not actually precompute the cov inverse completely,
# since it also depends on the current policies sigma.
# But, because of the way the Cholesky Decomp works,
# we can use the precomputed L (conditioner)
# (which is computed by an efficient LAPACK implementation)
# and adapt it for our new k(x_w+1,x_w+1) value (in python)
# (Which is dependent on pi)
# S_{ij} = \frac{1}{D_j} \left( A_{ij} - \sum_{k=1}^{j-1} S_{ik} S_{jk} D_k \right), \qquad\text{for } i>j
# https://martin-thoma.com/images/2012/07/cholesky-zerlegung-numerik.png
# This way conditioning of the GP can be done in O(dim(A)) time.
if not self.is_contextual() and False:
# Always assuming contextual will merely waste cpu cycles
# TODO: fix, this does not work
# safe inplace
self.conditioner[-1, -
1] = np.sqrt(pi_std**2 + self.Sig22 - self.adapt_norm)
return np.expand_dims(np.expand_dims(self.conditioner, 0), 0)
else:
conditioner = np.zeros(
(pi_std.shape[0], pi_std.shape[1]) + self.conditioner.shape)
conditioner[:, :] = self.conditioner
conditioner[:, :, -1, -
1] = np.sqrt(pi_std**2 + self.Sig22 - self.adapt_norm)
return conditioner
def mode(self) -> th.Tensor:
return self.distribution.mean
def actions_from_params(
self, mean: th.Tensor, std: th.Tensor, deterministic: bool = False, trajectory: th.Tensor = None
) -> th.Tensor:
self.proba_distribution(mean, std)
return self.get_actions(deterministic=deterministic, trajectory=trajectory)
def log_prob_from_params(self, mean: th.Tensor, std: th.Tensor, trajectory: th.Tensor = None):
actions = self.actions_from_params(mean, std, trajectory=trajectory)
log_prob = self.log_prob(actions)
return actions, log_prob
def print_info(self, traj: th.Tensor):
pi_mean, pi_std = self.distribution.mean, self.distribution.scale,
rho_mean, rho_std = self._conditioning_engine(traj, pi_mean, pi_std)
eta = self._get_rigged(pi_mean, pi_std,
rho_mean, rho_std)
print('pi ~ N('+str(pi_mean)+','+str(pi_std)+')')
print('rho ~ N('+str(rho_mean)+','+str(rho_std)+')')
class StdNet(nn.Module):
def __init__(self, latent_dim: int, action_dim: int, std_init: float, par_strength: bool, epsilon: float, return_log_std: bool):
super().__init__()
self.action_dim = action_dim
self.latent_dim = latent_dim
self.std_init = std_init
self.par_strength = par_strength
self.enforce_positive_type = EnforcePositiveType.SOFTPLUS
self.epsilon = epsilon
self.return_log_std = return_log_std
if self.par_strength == Par_Strength.SCALAR:
self.param = nn.Parameter(
th.Tensor([std_init]), requires_grad=True)
elif self.par_strength == Par_Strength.DIAG:
self.param = nn.Parameter(
th.Tensor(th.ones(action_dim)*std_init), requires_grad=True)
elif self.par_strength == Par_Strength.FULL:
ident = th.eye(action_dim)*std_init
ident_chol = fill_triangular_inverse(ident)
self.param = nn.Parameter(
th.Tensor(ident_chol), requires_grad=True)
elif self.par_strength == Par_Strength.CONT_SCALAR:
self.net = nn.Linear(latent_dim, 1)
elif self.par_strength == Par_Strength.CONT_HYBRID:
self.net = nn.Linear(latent_dim, 1)
self.param = nn.Parameter(
th.Tensor(th.ones(action_dim)*std_init), requires_grad=True)
elif self.par_strength == Par_Strength.CONT_DIAG:
self.net = nn.Linear(latent_dim, self.action_dim)
self.bias = th.ones(action_dim)*self.std_init
elif self.par_strength == Par_Strength.CONT_FULL:
self.net = nn.Linear(latent_dim, action_dim * (action_dim + 1) // 2)
self.bias = fill_triangular_inverse(th.eye(action_dim)*self.std_init)
def forward(self, x: th.Tensor) -> th.Tensor:
if self.par_strength == Par_Strength.SCALAR:
return self._ensure_positive_func(
th.ones(self.action_dim) * self.param[0])
elif self.par_strength == Par_Strength.DIAG:
return self._ensure_positive_func(self.param)
elif self.par_strength == Par_Strength.FULL:
return self._chol_from_flat(self.param)
elif self.par_strength == Par_Strength.CONT_SCALAR:
cont = self.net(x)
diag_chol = th.ones(self.action_dim, device=cont.device) * cont * self.std_init
return self._ensure_positive_func(diag_chol)
elif self.par_strength == Par_Strength.CONT_HYBRID:
cont = self.net(x)
return self._ensure_positive_func(self.param * cont)
elif self.par_strength == Par_Strength.CONT_DIAG:
cont = self.net(x)
bias = self.bias.to(cont.device)
diag_chol = cont + bias
return self._ensure_positive_func(diag_chol)
elif self.par_strength == Par_Strength.CONT_FULL:
cont = self.net(x)
bias = self.bias.to(device=cont.device)
return self._chol_from_flat(cont + bias)
raise Exception()
def _ensure_positive_func(self, x):
return self.enforce_positive_type.apply(x) + self.epsilon
def _chol_from_flat(self, flat_chol):
chol = fill_triangular(flat_chol)
return self._ensure_diagonal_positive(chol)
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):
return '<StdNet />'
def test():
mu = th.Tensor([[0.0, 0.0]])
sigma = th.Tensor([[0.9, 0.1]])
traj = th.Tensor([[[-1.0, -1.0], [-0.4, -0.4], [0.3, 0.3]]])
d = PCA_Distribution(2, window=3)
d.proba_distribution(mu, sigma)
d.print_info(traj)
print(d.sample(traj))
return d
if __name__ == '__main__':
test()