utilities3.py

import torch
import numpy as np
import scipy.io
import h5py
import torch.nn as nn

import operator
from functools import reduce
#################################################
#
# Utilities
#
#################################################
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# reading data
class MatReader(object):
    def __init__(self, file_path, to_torch=True, to_cuda=False, to_float=True):
        super(MatReader, self).__init__()

        self.to_torch = to_torch
        self.to_cuda = to_cuda
        self.to_float = to_float

        self.file_path = file_path

        self.data = None
        self.old_mat = True
        self.h5 = False
        self._load_file()

    def _load_file(self):

        if self.file_path[-3:] == '.h5':
            self.data = h5py.File(self.file_path, 'r')
            self.h5 = True

        else:
            try:
                self.data = scipy.io.loadmat(self.file_path)
            except:
                self.data = h5py.File(self.file_path, 'r')
                self.old_mat = False

    def load_file(self, file_path):
        self.file_path = file_path
        self._load_file()

    def read_field(self, field):
        x = self.data[field]

        if self.h5:
            x = x[()]

        if not self.old_mat:
            x = x[()]
            x = np.transpose(x, axes=range(len(x.shape) - 1, -1, -1))

        if self.to_float:
            x = x.astype(np.float32)

        if self.to_torch:
            x = torch.from_numpy(x)

            if self.to_cuda:
                x = x.cuda()

        return x

    def set_cuda(self, to_cuda):
        self.to_cuda = to_cuda

    def set_torch(self, to_torch):
        self.to_torch = to_torch

    def set_float(self, to_float):
        self.to_float = to_float

# normalization, pointwise gaussian
class UnitGaussianNormalizer(object):
    def __init__(self, x, eps=0.00001):
        super(UnitGaussianNormalizer, self).__init__()

        # x could be in shape of ntrain*n or ntrain*T*n or ntrain*n*T
        self.mean = torch.mean(x, 0)
        self.std = torch.std(x, 0)
        self.eps = eps

    def encode(self, x):
        x = (x - self.mean) / (self.std + self.eps)
        return x

    def decode(self, x, sample_idx=None):
        if sample_idx is None:
            std = self.std + self.eps # n
            mean = self.mean
        else:
            if len(self.mean.shape) == len(sample_idx[0].shape):
                std = self.std[sample_idx] + self.eps  # batch*n
                mean = self.mean[sample_idx]
            if len(self.mean.shape) > len(sample_idx[0].shape):
                std = self.std[:,sample_idx]+ self.eps # T*batch*n
                mean = self.mean[:,sample_idx]

        # x is in shape of batch*n or T*batch*n
        x = (x * std) + mean
        return x

    def cuda(self):
        self.mean = self.mean.cuda()
        self.std = self.std.cuda()

    def cpu(self):
        self.mean = self.mean.cpu()
        self.std = self.std.cpu()

# normalization, Gaussian
class GaussianNormalizer(object):
    def __init__(self, x, eps=0.00001):
        super(GaussianNormalizer, self).__init__()

        self.mean = torch.mean(x)
        self.std = torch.std(x)
        self.eps = eps

    def encode(self, x):
        x = (x - self.mean) / (self.std + self.eps)
        return x

    def decode(self, x, sample_idx=None):
        x = (x * (self.std + self.eps)) + self.mean
        return x

    def cuda(self):
        self.mean = self.mean.cuda()
        self.std = self.std.cuda()

    def cpu(self):
        self.mean = self.mean.cpu()
        self.std = self.std.cpu()


# normalization, scaling by range
class RangeNormalizer(object):
    def __init__(self, x, low=0.0, high=1.0):
        super(RangeNormalizer, self).__init__()
        mymin = torch.min(x, 0)[0].view(-1)
        mymax = torch.max(x, 0)[0].view(-1)

        self.a = (high - low)/(mymax - mymin)
        self.b = -self.a*mymax + high

    def encode(self, x):
        s = x.size()
        x = x.view(s[0], -1)
        x = self.a*x + self.b
        x = x.view(s)
        return x

    def decode(self, x):
        s = x.size()
        x = x.view(s[0], -1)
        x = (x - self.b)/self.a
        x = x.view(s)
        return x

#loss function with rel/abs Lp loss
class LpLoss(object):
    def __init__(self, d=2, p=2, size_average=True, reduction=True):
        super(LpLoss, self).__init__()

        #Dimension and Lp-norm type are postive
        assert d > 0 and p > 0

        self.d = d
        self.p = p
        self.reduction = reduction
        self.size_average = size_average

    def abs(self, x, y):
        num_examples = x.size()[0]

        #Assume uniform mesh
        h = 1.0 / (x.size()[1] - 1.0)

        all_norms = (h**(self.d/self.p))*torch.norm(x.reshape(num_examples,-1) - y.reshape(num_examples,-1), self.p, 1)

        if self.reduction:
            if self.size_average:
                return torch.mean(all_norms)
            else:
                return torch.sum(all_norms)

        return all_norms

    def rel(self, x, y):
        num_examples = x.size()[0]

        diff_norms = torch.norm(x.reshape(num_examples,-1) - y.reshape(num_examples,-1), self.p, 1)
        y_norms = torch.norm(y.reshape(num_examples,-1), self.p, 1)

        if self.reduction:
            if self.size_average:
                return torch.mean(diff_norms/y_norms)
            else:
                return torch.sum(diff_norms/y_norms)

        return diff_norms/y_norms

    def __call__(self, x, y):
        return self.rel(x, y)

# A simple feedforward neural network
class DenseNet(torch.nn.Module):
    def __init__(self, layers, nonlinearity, out_nonlinearity=None, normalize=False):
        super(DenseNet, self).__init__()

        self.n_layers = len(layers) - 1

        assert self.n_layers >= 1

        self.layers = nn.ModuleList()

        for j in range(self.n_layers):
            self.layers.append(nn.Linear(layers[j], layers[j+1]))

            if j != self.n_layers - 1:
                if normalize:
                    self.layers.append(nn.BatchNorm1d(layers[j+1]))

                self.layers.append(nonlinearity())

        if out_nonlinearity is not None:
            self.layers.append(out_nonlinearity())

    def forward(self, x):
        for _, l in enumerate(self.layers):
            x = l(x)

        return x

def pdist(sample_1, sample_2, norm=2, eps=1e-5):
    r"""Compute the matrix of all squared pairwise distances.
    Arguments
    ---------
    sample_1 : torch.Tensor or Variable
        The first sample, should be of shape ``(n_1, d)``.
    sample_2 : torch.Tensor or Variable
        The second sample, should be of shape ``(n_2, d)``.
    norm : float
        The l_p norm to be used.
    Returns
    -------
    torch.Tensor or Variable
        Matrix of shape (n_1, n_2). The [i, j]-th entry is equal to
        ``|| sample_1[i, :] - sample_2[j, :] ||_p``."""
    n_1, n_2 = sample_1.size(0), sample_2.size(0)
    norm = float(norm)
    if norm == 2.:
        norms_1 = torch.sum(sample_1**2, dim=1, keepdim=True)
        norms_2 = torch.sum(sample_2**2, dim=1, keepdim=True)
        norms = (norms_1.expand(n_1, n_2) +
                 norms_2.transpose(0, 1).expand(n_1, n_2))
        distances_squared = norms - 2 * sample_1.mm(sample_2.t())
        return torch.sqrt(eps + torch.abs(distances_squared))
    else:
        dim = sample_1.size(1)
        expanded_1 = sample_1.unsqueeze(1).expand(n_1, n_2, dim)
        expanded_2 = sample_2.unsqueeze(0).expand(n_1, n_2, dim)
        differences = torch.abs(expanded_1 - expanded_2) ** norm
        inner = torch.sum(differences, dim=2, keepdim=False)
        return (eps + inner) ** (1. / norm)

class MMDStatistic:
    r"""The *unbiased* MMD test of :cite:`gretton2012kernel`.
    The kernel used is equal to:
    .. math ::
        k(x, x') = \sum_{j=1}^k e^{-\alpha_j\|x - x'\|^2},
    for the :math:`\alpha_j` proved in :py:meth:`~.MMDStatistic.__call__`.
    Arguments
    ---------
    n_1: int
        The number of points in the first sample.
    n_2: int
        The number of points in the second sample."""

    def __init__(self, n_1, n_2):
        self.n_1 = n_1
        self.n_2 = n_2

        # The three constants used in the test.
        self.a00 = 1. / (n_1 * (n_1 - 1))
        self.a11 = 1. / (n_2 * (n_2 - 1))
        self.a01 = - 1. / (n_1 * n_2)

    def __call__(self, sample_1, sample_2, alphas, ret_matrix=False):
        r"""Evaluate the statistic.
        The kernel used is
        .. math::
            k(x, x') = \sum_{j=1}^k e^{-\alpha_j \|x - x'\|^2},
        for the provided ``alphas``.
        Arguments
        ---------
        sample_1: :class:`torch:torch.autograd.Variable`
            The first sample, of size ``(n_1, d)``.
        sample_2: variable of shape (n_2, d)
            The second sample, of size ``(n_2, d)``.
        alphas : list of :class:`float`
            The kernel parameters.
        ret_matrix: bool
            If set, the call with also return a second variable.
            This variable can be then used to compute a p-value using
            :py:meth:`~.MMDStatistic.pval`.
        Returns
        -------
        :class:`float`
            The test statistic.
        :class:`torch:torch.autograd.Variable`
            Returned only if ``ret_matrix`` was set to true."""
        sample_12 = torch.cat((sample_1, sample_2), 0)
        distances = pdist(sample_12, sample_12, norm=2)

        kernels = None
        for alpha in alphas:
            kernels_a = torch.exp(- alpha * distances ** 2)
            if kernels is None:
                kernels = kernels_a
            else:
                kernels = kernels + kernels_a

        k_1 = kernels[:self.n_1, :self.n_1]
        k_2 = kernels[self.n_1:, self.n_1:]
        k_12 = kernels[:self.n_1, self.n_1:]

        mmd = (2 * self.a01 * k_12.sum() +
               self.a00 * (k_1.sum() - torch.trace(k_1)) +
               self.a11 * (k_2.sum() - torch.trace(k_2)))
        if ret_matrix:
            return mmd, kernels
        else:
            return mmd

# print the number of parameters
def count_params(model):
    c = 0
    for p in list(model.parameters()):
        c += reduce(operator.mul, list(p.size()))
    return c