Source code for tvo.models.gmm

# -*- coding: utf-8 -*-
# Copyright (C) 2021 Machine Learning Group of the University of Oldenburg.
# Licensed under the Academic Free License version 3.0


import torch as to
import math
from torch.distributions.one_hot_categorical import OneHotCategorical

from torch import Tensor
from typing import Union, Tuple

import tvo
from tvo.utils.parallel import pprint, all_reduce, broadcast
from tvo.variational.TVOVariationalStates import TVOVariationalStates
from tvo.variational._utils import mean_posterior
from tvo.utils.model_protocols import Optimized, Sampler, Reconstructor
from tvo.utils.sanity import fix_theta


[docs] class GMM(Optimized, Sampler, Reconstructor): def __init__( self, H: int, D: int, W_init: Tensor = None, sigma2_init: Tensor = None, pies_init: Tensor = None, precision: to.dtype = to.float64, ): """Gaussian Mixture model (GMM). :param H: Number of hidden units. :param D: Number of observables. :param W_init: Tensor with shape (D,H), initializes GM weights. :param pies_init: Tensor with shape (H,), initializes GM priors. :param precision: Floating point precision required. Must be one of torch.float32 or torch.float64. """ assert precision in (to.float32, to.float64), "precision must be one of torch.float{32,64}" self._precision = precision device = tvo.get_device() if W_init is not None: assert W_init.shape == (D, H) W_init = W_init.to(dtype=precision, device=device) else: W_init = to.rand((D, H), dtype=precision, device=device) broadcast(W_init) if pies_init is not None: assert pies_init.shape == (H,) pies_init = pies_init.to(dtype=precision, device=device) else: pies_init = to.full((H,), 1.0 / H, dtype=precision, device=device) if sigma2_init is not None: assert sigma2_init.shape == (1,) sigma2_init = sigma2_init.to(dtype=precision, device=device) else: sigma2_init = to.tensor([1.0], dtype=precision, device=device) self._theta = {"pies": pies_init, "W": W_init, "sigma2": sigma2_init} eps, inf = 1.0e-5, math.inf self.policy = { "W": [None, to.full_like(self._theta["W"], -inf), to.full_like(self._theta["W"], inf)], "pies": [ None, to.full_like(self._theta["pies"], eps), to.full_like(self._theta["pies"], 1.0 - eps), ], "sigma2": [ None, to.full_like(self._theta["sigma2"], eps), to.full_like(self._theta["sigma2"], inf), ], } self.my_Wp = to.zeros((D, H), dtype=precision, device=device) self.my_Wq = to.zeros((H), dtype=precision, device=device) self.my_pies = to.zeros(H, dtype=precision, device=device) self.my_sigma2 = to.zeros(1, dtype=precision, device=device) self.my_N = to.tensor([0], dtype=to.int, device=device) self._config = dict(H=H, D=D, precision=precision, device=device) self._shape = self.theta["W"].shape
[docs] def log_pseudo_joint(self, data: Tensor, states: Tensor) -> Tensor: # type: ignore """Evaluate log-pseudo-joints for GMM.""" Kfloat = states.to( dtype=self.theta["W"].dtype ) # N,C,C # TODO Find solution to avoid byte->float casting Wbar = to.matmul( Kfloat, self.theta["W"].t() ) # N,C,D # TODO Pre-allocate tensor and use `out` argument of to.matmul lpj = to.mul( to.sum(to.pow(Wbar - data[:, None, :], 2), dim=2), -1 / 2 / self.theta["sigma2"] ) + to.matmul(Kfloat, to.log(self.theta["pies"])) return lpj.to(device=states.device)
[docs] def log_joint(self, data: Tensor, states: Tensor, lpj: Tensor = None) -> Tensor: """Evaluate log-joints for GMM.""" if lpj is None: lpj = self.log_pseudo_joint(data, states) D = self.shape[0] return lpj - D / 2 * to.log(2 * math.pi * self.theta["sigma2"])
[docs] def update_param_batch(self, idx: Tensor, batch: Tensor, states: Tensor) -> None: lpj = states.lpj[idx] K = states.K[idx] batch_size, S, _ = K.shape Kfloat = K.to(dtype=lpj.dtype) # TODO Find solution to avoid byte->float casting Wbar = to.matmul( Kfloat, self.theta["W"].t() ) # N,S,D # TODO Find solution to re-use evaluations from E-step batch_s_pjc = mean_posterior(Kfloat, lpj) # is (batch_size,H) mean_posterior(Kfloat, lpj) batch_Wp = batch.unsqueeze(2) * batch_s_pjc.unsqueeze(1) # is (batch_size,D,H) batch_sigma2 = mean_posterior(to.sum((batch[:, None, :] - Wbar) ** 2, dim=2), lpj) self.my_pies.add_(to.sum(batch_s_pjc, dim=0)) self.my_Wp.add_(to.sum(batch_Wp, dim=0)) self.my_Wq.add_(to.sum(batch_s_pjc, dim=0)) self.my_sigma2.add_(to.sum(batch_sigma2)) self.my_N.add_(batch_size) return None
[docs] def update_param_epoch(self) -> None: theta = self.theta policy = self.policy all_reduce(self.my_Wp) all_reduce(self.my_Wq) all_reduce(self.my_pies) all_reduce(self.my_sigma2) all_reduce(self.my_N) N = self.my_N.item() D = self.shape[0] # Calculate updated W Wold_noisy = theta["W"] + 0.1 * to.randn_like(theta["W"]) broadcast(Wold_noisy) theta_new = {} try: theta_new["W"] = self.my_Wp / self.my_Wq[None, :] except RuntimeError: pprint("Inversion error. Will not update W but add some noise instead.") theta_new["W"] = Wold_noisy # Calculate updated pi theta_new["pies"] = self.my_pies / N # Calculate updated sigma^2 theta_new["sigma2"] = self.my_sigma2 / N / D policy["W"][0] = Wold_noisy policy["pies"][0] = theta["pies"] policy["sigma2"][0] = theta["sigma2"] fix_theta(theta_new, policy) for key in theta: theta[key] = theta_new[key] self.my_Wp[:] = 0.0 self.my_Wq[:] = 0.0 self.my_pies[:] = 0.0 self.my_sigma2[:] = 0.0 self.my_N[:] = 0.0
@property def shape(self) -> Tuple[int, ...]: return self.theta["W"].shape
[docs] def generate_data( self, N: int = None, hidden_state: to.Tensor = None ) -> Union[to.Tensor, Tuple[to.Tensor, to.Tensor]]: precision, device = self.precision, tvo.get_device() D, H = self.shape if hidden_state is None: assert N is not None pies = self.theta["pies"] hidden_state = OneHotCategorical(probs=pies).sample([N]) == 1 must_return_hidden_state = True else: shape = hidden_state.shape if N is None: N = shape[0] assert shape == (N, H), f"hidden_state has shape {shape}, expected ({N},{H})" must_return_hidden_state = False Wbar = to.zeros((N, D), dtype=precision, device=device) for n in range(N): for h in range(H): if hidden_state[n, h]: Wbar[n] += self.theta["W"][:, h] # Add noise according to the model parameters Y = Wbar + to.sqrt(self.theta["sigma2"]) * to.randn((N, D), dtype=precision, device=device) return (Y, hidden_state) if must_return_hidden_state else Y
[docs] def data_estimator(self, idx: Tensor, batch: Tensor, states: TVOVariationalStates) -> Tensor: # Not yet implemented """Estimator used for data reconstruction. Data reconstruction can only be supported by a model if it implements this method. The estimator to be implemented is defined as follows:""" r""" :math:`\\langle \langle y_d \rangle_{p(y_d|\vec{s},\Theta)} \rangle_{q(\vec{s}|\mathcal{K},\Theta)}` # noqa """ # Not K = states.K[idx] # TODO Find solution to avoid byte->float casting of `K` # TODO Pre-allocate tensor and use `out` argument of to.matmul return mean_posterior( to.matmul(K.to(dtype=self.precision), self.theta["W"].t()), states.lpj[idx] )