Revision

**241bf7ad2b806f6677a5e23534247f35f3a70f10**authored by rballester on**19 February 2023, 19:35:27 UTC**, committed by rballester on**19 February 2023, 19:35:27 UTC****1 parent**be80cb2

anova.py

```
import copy
import numpy as np
import torch
import tntorch as tn
def anova_decomposition(t, marginals=None):
"""
Compute an extended tensor that contains all terms of the ANOVA decomposition for a given tensor.
Reference: R. Ballester-Ripoll, E. G. Paredes, and R. Pajarola: `"Sobol Tensor Trains for Global Sensitivity Analysis" (2017) <https://www.sciencedirect.com/science/article/pii/S0951832018303132?dgcid=rss_sd_all>`_
:param t: ND input tensor
:param marginals: list of N vectors, each containing the PMF for each variable (use None for uniform distributions)
:return: a :class:`Tensor`
"""
if t.batch:
raise ValueError('Batched tensors are not supproted.')
marginals = copy.deepcopy(marginals)
if marginals is None:
marginals = [None] * t.dim()
for n in range(t.dim()):
if marginals[n] is None:
marginals[n] = torch.ones([t.shape[n]]) / float(t.shape[n])
cores = [c.clone() for c in t.cores]
Us = []
idxs = []
for n in range(t.dim()):
if t.Us[n] is None:
U = torch.eye(t.shape[n])
else:
U = t.Us[n]
expected = torch.sum(U * (marginals[n][:, None] / torch.sum(marginals[n])), dim=0, keepdim=True)
Us.append(torch.cat((expected, U-expected), dim=0))
idxs.append([0] + [1] * t.shape[n])
return tn.Tensor(cores, Us, idxs=idxs)
def undo_anova_decomposition(a):
"""
Undo the transformation done by :func:`anova_decomposition()`.
:param a: a :class:`Tensor` obtained with :func:`anova_decomposition()`
:return: a :class:`Tensor` t that has `a` as its ANOVA tensor
"""
cores = []
Us = []
for n in range(a.dim()):
if a.Us[n] is None:
cores.append(a.cores[n][..., 1:, :] + a.cores[n][..., 0:1, :])
Us.append(None)
else:
cores.append(a.cores[n].clone())
Us.append(a.Us[n][1:, :] + a.Us[n][0:1, :])
return tn.Tensor(cores, Us=Us)
def truncate_anova(t, mask, keepdim=False, marginals=None):
"""
Given a tensor and a mask, return the function that results after deleting all ANOVA terms that do not satisfy the
mask.
:Example:
>>> t = ... # an ND tensor
>>> x = tn.symbols(t.dim())[0]
>>> t2 = tn.truncate_anova(t, mask=tn.only(x), keepdim=False) # This tensor will depend on one variable only
:param t: an N-dimensional :class:`Tensor`
:param mask: an N-dimensional mask
:param keepdim: if True, all dummy dimensions will be preserved, otherwise they will disappear. Default is False
:param marginals: see :func:`anova_decomposition()`. Defaults to uniform marginals
:return: a :class:`Tensor`
"""
t = tn.undo_anova_decomposition(tn.mask(tn.anova_decomposition(t, marginals=marginals), mask=mask))
if not keepdim:
N = t.dim()
affecting = torch.sum(tn.accepted_inputs(mask).double(), dim=0)
slices = [0 for n in range(N)]
for i in np.where(affecting)[0]:
slices[int(i)] = slice(None)
t = t[tuple(slices)]
return t
def sobol(t, mask, marginals=None, normalize=True):
"""
Compute Sobol indices (as given by a certain mask) for a tensor and independently distributed input variables.
Reference: R. Ballester-Ripoll, E. G. Paredes, and R. Pajarola: `"Sobol Tensor Trains for Global Sensitivity Analysis" (2017) <https://www.sciencedirect.com/science/article/pii/S0951832018303132?dgcid=rss_sd_all>`_
:param t: an N-dimensional :class:`Tensor`
:param mask: an N-dimensional mask
:param marginals: a list of N vectors (will be normalized if not summing to 1). If None (default), uniform distributions are assumed for all variables
:param normalize: whether to normalize indices by the total variance of the model (True by default)
:return: a scalar >= 0
"""
if marginals is None:
marginals = [None] * t.dim()
a = tn.anova_decomposition(t, marginals)
a -= tn.Tensor([torch.cat((torch.ones(1, 1, 1),
torch.zeros(1, sh-1, 1)), dim=1)
for sh in a.shape])*a[(0,)*t.dim()] # Set empty tuple to 0
am = a.clone()
for n in range(t.dim()):
if marginals[n] is None:
m = torch.ones([t.shape[n]])
else:
m = marginals[n]
m /= torch.sum(m) # Make sure each marginal sums to 1
if am.Us[n] is None:
if am.cores[n].dim() == 3:
am.cores[n][:, 1:, :] *= m[None, :, None]
else:
am.cores[n][1:, :] *= m[:, None]
else:
am.Us[n][1:, :] *= m[:, None]
am_masked = tn.mask(am, mask)
if am_masked.cores[-1].shape[-1] > 1:
am_masked.cores.append(torch.eye(am_masked.cores[-1].shape[-1])[:, :, None])
am_masked.Us.append(None)
if normalize:
return tn.dot(a, am_masked) / tn.dot(a, am)
else:
return tn.dot(a, am_masked)
def mean_dimension(t, mask=None, marginals=None):
"""
Computes the mean dimension of a given tensor with given marginal distributions. This quantity measures how well the
represented function can be expressed as a sum of low-parametric functions. For example, mean dimension 1 (the
lowest possible value) means that it is a purely additive function: :math:`f(x_1, ..., x_N) = f_1(x_1) + ... + f_N(x_N)`.
Assumption: the input variables :math:`x_n` are independently distributed.
References:
- R. E. Caflisch, W. J. Morokoff, and A. B. Owen: `"Valuation of Mortgage Backed Securities Using Brownian Bridges to Reduce Effective Dimension" (1997) <http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.36.3160>`_
- R. Ballester-Ripoll, E. G. Paredes, and R. Pajarola: `"Tensor Algorithms for Advanced Sensitivity Metrics" (2017) <https://epubs.siam.org/doi/10.1137/17M1160252>`_
:param t: an N-dimensional :class:`Tensor`
:param marginals: a list of N vectors (will be normalized if not summing to 1). If None (default), uniform distributions are assumed for all variables
:return: a scalar >= 1
"""
if mask is None:
return tn.sobol(t, tn.weight(t.dim()), marginals=marginals)
else:
return tn.sobol(t, tn.mask(tn.weight(t.dim()), mask), marginals=marginals) / tn.sobol(t, mask, marginals=marginals)
def dimension_distribution(t, mask=None, order=None, marginals=None):
"""
Computes the dimension distribution of an ND tensor.
:param t: ND input :class:`Tensor`
:param mask: an optional mask :class:`Tensor` to restrict to
:param order: int, compute only this many order contributions. By default, all N are returned
:param marginals: PMFs for input variables. By default, uniform distributions
:return: a PyTorch vector containing N elements
"""
if order is None:
order = t.dim()
if mask is None:
return tn.sobol(t, tn.weight_one_hot(t.dim(), order+1), marginals=marginals).torch()[1:]
else:
mask2 = tn.mask(tn.weight_one_hot(t.dim(), order+1), mask)
return tn.sobol(t, mask2, marginals=marginals).torch()[1:] / tn.sobol(t, mask, marginals=marginals)
```

Computing file changes ...