Module mogptk.gpr.mean

Expand source code Browse git 
import torch
from . import Parameter, config

class Mean(torch.nn.Module):
    """
    Defines a trainable mean function, complementary to the the way we also have a trainable covariance function (the kernel).
    """
    def __init__(self):
        super().__init__()

    def __call__(self, X):
        """
        Return the mean for a given `X`. This is the same as calling `mean(X)` but `X` doesn't necessarily have to be a tensor.

        Args:
            X (torch.tensor): Input of shape (data_points,input_dims).

        Returns:
            torch.tensor: Mean values of shape (data_points,).
        """
        X = self._check_input(X)
        return self.mean(X)

    def name(self):
        return self.__class__.__name__

    def __setattr__(self, name, val):
        if name == 'train':
            for p in self.parameters():
                p.train = val
            return
        if hasattr(self, name) and isinstance(getattr(self, name), Parameter):
            raise AttributeError("parameter is read-only, use Parameter.assign()")
        if isinstance(val, Parameter) and val._name is None:
            val._name = '%s.%s' % (self.__class__.__name__, name)
        elif isinstance(val, torch.nn.ModuleList):
            for i, item in enumerate(val):
                for p in item.parameters():
                    p._name = '%s[%d].%s' % (self.__class__.__name__, i, p._name)
        super().__setattr__(name, val)

    def _check_input(self, X):
        if not torch.is_tensor(X):
            X = torch.tensor(X, device=config.device, dtype=config.dtype)
        elif X.device != config.device or X.dtype != config.dtype:
            X = X.to(device, dtype)
        if X.ndim != 2:
            raise ValueError("X should have two dimensions (data_points,input_dims)")
        if X.shape[0] == 0 or X.shape[1] == 0:
            raise ValueError("X must not be empty")
        return X

    def mean(self, X):
        """
        Return the mean for a given `X`.

        Args:
            X (torch.tensor): Input of shape (data_points,input_dims).

        Returns:
            torch.tensor: Mean values of shape (data_points,).
        """
        raise NotImplementedError()

class MultiOutputMean(Mean):
    """
    Multi-output mean to assign a different mean per channel.

    Args:
        means (mogptk.gpr.mean.Mean): List of means equal to the number of output dimensions.
    """
    def __init__(self, *means):
        super().__init__()

        if isinstance(means, tuple):
            if len(means) == 1 and isinstance(means[0], list):
                means = means[0]
            else:
                means = list(means)
        elif not isinstance(means, list):
            means = [means]
        if len(means) == 0:
            raise ValueError("must pass at least one mean")
        for i, mean in enumerate(means):
            if not issubclass(type(mean), Mean):
                raise ValueError("must pass means")
            elif isinstance(mean, MultiOutputMean):
                raise ValueError("can not nest MultiOutputMeans")

        self.output_dims = len(means)
        self.means = means

    def name(self):
        names = [mean.name() for mean in self.means]
        return '[%s]' % (','.join(names),)

    def _channel_indices(self, X):
        c = X[:,0].long()
        m = [c==j for j in range(self.output_dims)]
        r = [torch.nonzero(m[i], as_tuple=False).reshape(-1) for i in range(self.output_dims)]
        return r

    def mean(self, X):
        r = self._channel_indices(X)
        res = torch.empty(X.shape[0], 1, device=config.device, dtype=config.dtype)  # Nx1
        for i in range(self.output_dims):
            res[r[i]] = self.means[i].mean(X[r[i],1:])
        return res

class ConstantMean(Mean):
    """
    Constant mean function:

    $$ m(X) = b $$

    with \\(b\\) the bias.

    Args:

    Attributes:
        bias (mogptk.gpr.parameter.Parameter): Bias \\(b\\).
    """
    def __init__(self):
        super().__init__()
        self.bias = Parameter(0.0)

    def mean(self, X):
        return self.bias().repeat(X.shape[0], 1)

class LinearMean(Mean):
    """
    Linear mean function:

    $$ m(X) = aX + b $$

    with \\(a\\) the slope and \\(b\\) the bias.

    Args:
        input_dims (int): Number of input dimensions.

    Attributes:
        bias (mogptk.gpr.parameter.Parameter): Bias \\(b\\).
        slope (mogptk.gpr.parameter.Parameter): Slope \\(a\\).
    """
    def __init__(self, input_dims=1):
        super().__init__()
        self.bias = Parameter(0.0)
        self.slope = Parameter(torch.zeros(input_dims))

    def mean(self, X):
        return self.bias() + X.mm(self.slope().reshape(-1,1))

Classes

class ConstantMean

Constant mean function:

m(X) = b

with b the bias.

Args:

Attributes

bias : Parameter
Bias b.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

Expand source code Browse git 
class ConstantMean(Mean):
    """
    Constant mean function:

    $$ m(X) = b $$

    with \\(b\\) the bias.

    Args:

    Attributes:
        bias (mogptk.gpr.parameter.Parameter): Bias \\(b\\).
    """
    def __init__(self):
        super().__init__()
        self.bias = Parameter(0.0)

    def mean(self, X):
        return self.bias().repeat(X.shape[0], 1)

Ancestors

  • Mean
  • torch.nn.modules.module.Module

Inherited members

class LinearMean (input_dims=1)

Linear mean function:

m(X) = aX + b

with a the slope and b the bias.

Args

input_dims : int
Number of input dimensions.

Attributes

bias : Parameter
Bias b.
slope : Parameter
Slope a.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

Expand source code Browse git 
class LinearMean(Mean):
    """
    Linear mean function:

    $$ m(X) = aX + b $$

    with \\(a\\) the slope and \\(b\\) the bias.

    Args:
        input_dims (int): Number of input dimensions.

    Attributes:
        bias (mogptk.gpr.parameter.Parameter): Bias \\(b\\).
        slope (mogptk.gpr.parameter.Parameter): Slope \\(a\\).
    """
    def __init__(self, input_dims=1):
        super().__init__()
        self.bias = Parameter(0.0)
        self.slope = Parameter(torch.zeros(input_dims))

    def mean(self, X):
        return self.bias() + X.mm(self.slope().reshape(-1,1))

Ancestors

  • Mean
  • torch.nn.modules.module.Module

Inherited members

class Mean

Defines a trainable mean function, complementary to the the way we also have a trainable covariance function (the kernel).

Initializes internal Module state, shared by both nn.Module and ScriptModule.

Expand source code Browse git 
class Mean(torch.nn.Module):
    """
    Defines a trainable mean function, complementary to the the way we also have a trainable covariance function (the kernel).
    """
    def __init__(self):
        super().__init__()

    def __call__(self, X):
        """
        Return the mean for a given `X`. This is the same as calling `mean(X)` but `X` doesn't necessarily have to be a tensor.

        Args:
            X (torch.tensor): Input of shape (data_points,input_dims).

        Returns:
            torch.tensor: Mean values of shape (data_points,).
        """
        X = self._check_input(X)
        return self.mean(X)

    def name(self):
        return self.__class__.__name__

    def __setattr__(self, name, val):
        if name == 'train':
            for p in self.parameters():
                p.train = val
            return
        if hasattr(self, name) and isinstance(getattr(self, name), Parameter):
            raise AttributeError("parameter is read-only, use Parameter.assign()")
        if isinstance(val, Parameter) and val._name is None:
            val._name = '%s.%s' % (self.__class__.__name__, name)
        elif isinstance(val, torch.nn.ModuleList):
            for i, item in enumerate(val):
                for p in item.parameters():
                    p._name = '%s[%d].%s' % (self.__class__.__name__, i, p._name)
        super().__setattr__(name, val)

    def _check_input(self, X):
        if not torch.is_tensor(X):
            X = torch.tensor(X, device=config.device, dtype=config.dtype)
        elif X.device != config.device or X.dtype != config.dtype:
            X = X.to(device, dtype)
        if X.ndim != 2:
            raise ValueError("X should have two dimensions (data_points,input_dims)")
        if X.shape[0] == 0 or X.shape[1] == 0:
            raise ValueError("X must not be empty")
        return X

    def mean(self, X):
        """
        Return the mean for a given `X`.

        Args:
            X (torch.tensor): Input of shape (data_points,input_dims).

        Returns:
            torch.tensor: Mean values of shape (data_points,).
        """
        raise NotImplementedError()

Ancestors

  • torch.nn.modules.module.Module

Subclasses

Methods

def mean(self, X)

Return the mean for a given X.

Args

X : torch.tensor
Input of shape (data_points,input_dims).

Returns

torch.tensor
Mean values of shape (data_points,).
Expand source code Browse git 
def mean(self, X):
    """
    Return the mean for a given `X`.

    Args:
        X (torch.tensor): Input of shape (data_points,input_dims).

    Returns:
        torch.tensor: Mean values of shape (data_points,).
    """
    raise NotImplementedError()
def name(self)
Expand source code Browse git 
def name(self):
    return self.__class__.__name__
class MultiOutputMean (*means)

Multi-output mean to assign a different mean per channel.

Args

means : Mean
List of means equal to the number of output dimensions.

Initializes internal Module state, shared by both nn.Module and ScriptModule.

Expand source code Browse git 
class MultiOutputMean(Mean):
    """
    Multi-output mean to assign a different mean per channel.

    Args:
        means (mogptk.gpr.mean.Mean): List of means equal to the number of output dimensions.
    """
    def __init__(self, *means):
        super().__init__()

        if isinstance(means, tuple):
            if len(means) == 1 and isinstance(means[0], list):
                means = means[0]
            else:
                means = list(means)
        elif not isinstance(means, list):
            means = [means]
        if len(means) == 0:
            raise ValueError("must pass at least one mean")
        for i, mean in enumerate(means):
            if not issubclass(type(mean), Mean):
                raise ValueError("must pass means")
            elif isinstance(mean, MultiOutputMean):
                raise ValueError("can not nest MultiOutputMeans")

        self.output_dims = len(means)
        self.means = means

    def name(self):
        names = [mean.name() for mean in self.means]
        return '[%s]' % (','.join(names),)

    def _channel_indices(self, X):
        c = X[:,0].long()
        m = [c==j for j in range(self.output_dims)]
        r = [torch.nonzero(m[i], as_tuple=False).reshape(-1) for i in range(self.output_dims)]
        return r

    def mean(self, X):
        r = self._channel_indices(X)
        res = torch.empty(X.shape[0], 1, device=config.device, dtype=config.dtype)  # Nx1
        for i in range(self.output_dims):
            res[r[i]] = self.means[i].mean(X[r[i],1:])
        return res

Ancestors

  • Mean
  • torch.nn.modules.module.Module

Methods

def name(self)
Expand source code Browse git 
def name(self):
    names = [mean.name() for mean in self.means]
    return '[%s]' % (','.join(names),)

Inherited members