"""These layers are still work in progress."""
import math
import torch
import torch.nn.functional as F
from torch import Tensor, nn
from torch.nn.common_types import _size_2_t
def check_lpbnn_parameters_consistency(
hidden_size: int, std_factor: float, num_estimators: int
) -> None:
if hidden_size < 1:
raise ValueError(f"hidden_size must be greater than 0. Got {hidden_size}.")
if std_factor < 0:
raise ValueError(f"std_factor must be greater than 0. Got {std_factor}.")
if num_estimators < 1:
raise ValueError(f"num_estimators must be greater than 0. Got {num_estimators}.")
def _sample(mu: Tensor, logvar: Tensor, std_factor: float) -> Tensor:
eps = torch.randn_like(mu)
return eps * std_factor * torch.exp(logvar * 0.5) + mu
[docs]class LPBNNLinear(nn.Module):
__constants__ = [
"in_features",
"out_features",
"num_estimators",
"hidden_size",
]
in_features: int
out_features: int
num_estimators: int
bias: torch.Tensor | None
def __init__(
self,
in_features: int,
out_features: int,
num_estimators: int,
hidden_size: int = 32,
std_factor: float = 1e-2,
bias: bool = True,
device=None,
dtype=None,
) -> None:
"""LPBNN-style linear layer.
Args:
in_features (int): Number of input features.
out_features (int): Number of output features.
num_estimators (int): Number of models to sample from.
hidden_size (int): Size of the hidden layer. Defaults to 32.
std_factor (float): Factor to multiply the standard deviation of the
latent noise. Defaults to 1e-2.
bias (bool): If ``True``, adds a learnable bias to the output.
Defaults to ``True``.
device (torch.device): Device on which the layer is stored.
Defaults to ``None``.
dtype (torch.dtype): Data type of the layer. Defaults to ``None``.
Reference:
`Encoding the latent posterior of Bayesian Neural Networks for
uncertainty quantification <https://arxiv.org/abs/2012.02818>`_.
"""
check_lpbnn_parameters_consistency(hidden_size, std_factor, num_estimators)
factory_kwargs = {"device": device, "dtype": dtype}
super().__init__()
self.in_features = in_features
self.out_features = out_features
self.hidden_size = hidden_size
self.std_factor = std_factor
self.num_estimators = num_estimators
# for the KL Loss
self.lprior = 0
self.linear = nn.Linear(in_features, out_features, bias=False, **factory_kwargs)
self.alpha = nn.Parameter(
torch.empty((num_estimators, in_features), **factory_kwargs),
requires_grad=False,
)
self.gamma = nn.Parameter(torch.empty((num_estimators, out_features), **factory_kwargs))
self.encoder = nn.Linear(in_features, self.hidden_size, **factory_kwargs)
self.latent_mean = nn.Linear(self.hidden_size, self.hidden_size, **factory_kwargs)
self.latent_logvar = nn.Linear(self.hidden_size, self.hidden_size, **factory_kwargs)
self.decoder = nn.Linear(self.hidden_size, in_features, **factory_kwargs)
self.latent_loss = torch.zeros(1, **factory_kwargs)
if bias:
self.bias = nn.Parameter(torch.empty((num_estimators, out_features), **factory_kwargs))
else:
self.register_parameter("bias", None)
self.reset_parameters()
def reset_parameters(self):
nn.init.normal_(self.alpha, mean=1.0, std=0.1)
nn.init.normal_(self.gamma, mean=1.0, std=0.1)
self.linear.reset_parameters()
self.encoder.reset_parameters()
self.decoder.reset_parameters()
self.latent_mean.reset_parameters()
self.latent_logvar.reset_parameters()
if self.bias is not None:
fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.linear.weight)
bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0
nn.init.uniform_(self.bias, -bound, bound)
def forward(self, x: Tensor) -> Tensor:
# Draw a sample from the dist generated by the noise self.alpha
latent = F.relu(self.encoder(self.alpha))
latent_mean, latent_logvar = (
self.latent_mean(latent),
self.latent_logvar(latent),
)
z_latent = _sample(latent_mean, latent_logvar, self.std_factor)
# one sample per "model" with as many features as x
alpha_sample = self.decoder(z_latent)
# Compute the latent loss
if self.training:
mse = F.mse_loss(alpha_sample, self.alpha)
kld = -0.5 * torch.sum(1 + latent_logvar - latent_mean**2 - torch.exp(latent_logvar))
# For the KL Loss
self.lvposterior = mse + kld
# Compute the output
num_examples_per_model = int(x.size(0) / self.num_estimators)
alpha = alpha_sample.repeat((num_examples_per_model, 1))
gamma = self.gamma.repeat((num_examples_per_model, 1))
out = self.linear(x * alpha) * gamma
if self.bias is not None:
bias = self.bias.repeat((num_examples_per_model, 1))
out += bias
return out
def extra_repr(self) -> str:
return (
f"in_features={self.in_features}, "
f"out_features={self.out_features}, "
f"num_estimators={self.num_estimators}, "
f"hidden_size={self.hidden_size}, bias={self.bias is not None}"
)
[docs]class LPBNNConv2d(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
num_estimators: int,
kernel_size: _size_2_t,
stride: _size_2_t = 1,
padding: str | _size_2_t = 0,
groups: int = 1,
hidden_size: int = 32,
std_factor: float = 1e-2,
gamma: bool = True,
bias: bool = True,
padding_mode: str = "zeros",
device=None,
dtype=None,
):
"""LPBNN-style 2D convolutional layer.
Args:
in_channels (int): Number of input channels.
out_channels (int): Number of output channels.
num_estimators (int): Number of models to sample from.
kernel_size (int or tuple): Size of the convolving kernel.
stride (int or tuple, optional): Stride of the convolution. Default: 1.
padding (int or tuple, optional): Zero-padding added to both sides of the input. Default: 0.
groups (int, optional): Number of blocked connections from input channels to output channels. Default: 1.
hidden_size (int): Size of the hidden layer. Defaults to 32.
std_factor (float): Factor to multiply the standard deviation of the
latent noise. Defaults to 1e-2.
gamma (bool): If ``True``, adds a learnable gamma to the output.
Defaults to ``True``.
bias (bool): If ``True``, adds a learnable bias to the output.
Defaults to ``True``.
padding_mode (str): 'zeros', 'reflect', 'replicate' or 'circular'. Default: 'zeros'.
device (torch.device): Device on which the layer is stored.
Defaults to ``None``.
dtype (torch.dtype): Data type of the layer. Defaults to ``None``.
Reference:
`Encoding the latent posterior of Bayesian Neural Networks for
uncertainty quantification <https://arxiv.org/abs/2012.02818>`_.
"""
check_lpbnn_parameters_consistency(hidden_size, std_factor, num_estimators)
factory_kwargs = {"device": device, "dtype": dtype}
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.hidden_size = hidden_size
self.std_factor = std_factor
self.num_estimators = num_estimators
# for the KL Loss
self.lprior = 0
self.conv = nn.Conv2d(
in_channels,
out_channels,
kernel_size,
stride=stride,
padding=padding,
groups=groups,
bias=False,
padding_mode=padding_mode,
)
self.alpha = nn.Parameter(
torch.empty(num_estimators, in_channels, **factory_kwargs),
requires_grad=False,
)
self.encoder = nn.Linear(in_channels, self.hidden_size, **factory_kwargs)
self.decoder = nn.Linear(self.hidden_size, in_channels, **factory_kwargs)
self.latent_mean = nn.Linear(self.hidden_size, self.hidden_size, **factory_kwargs)
self.latent_logvar = nn.Linear(self.hidden_size, self.hidden_size, **factory_kwargs)
self.latent_loss = torch.zeros(1, **factory_kwargs)
if gamma:
self.gamma = nn.Parameter(torch.empty((num_estimators, out_channels), **factory_kwargs))
else:
self.register_parameter("gamma", None)
if bias:
self.bias = nn.Parameter(torch.empty((num_estimators, out_channels), **factory_kwargs))
else:
self.register_parameter("bias", None)
self.reset_parameters()
def reset_parameters(self):
nn.init.normal_(self.alpha, mean=1.0, std=0.1)
if self.gamma is not None:
nn.init.normal_(self.gamma, mean=1.0, std=0.1)
self.conv.reset_parameters()
self.encoder.reset_parameters()
self.decoder.reset_parameters()
self.latent_mean.reset_parameters()
self.latent_logvar.reset_parameters()
if self.bias is not None:
fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.conv.weight)
bound = 1 / math.sqrt(fan_in)
nn.init.uniform_(self.bias, -bound, bound)
def forward(self, x: Tensor) -> Tensor:
# Draw a sample from the dist generated by the latent noise self.alpha
latent = F.relu(self.encoder(self.alpha))
latent_mean, latent_logvar = (
self.latent_mean(latent),
self.latent_logvar(latent),
)
z_latent = _sample(latent_mean, latent_logvar, self.std_factor)
# one sample per "model" with as many features as x
alpha_sample = self.decoder(z_latent)
# Compute the latent loss
if self.training:
mse = F.mse_loss(alpha_sample, self.alpha)
kld = -0.5 * torch.sum(1 + latent_logvar - latent_mean.pow(2) - latent_logvar.exp())
# for the KL Loss
self.lvposterior = mse + kld
num_examples_per_model = int(x.size(0) / self.num_estimators)
# Compute the output
alpha = alpha_sample.repeat((num_examples_per_model, 1)).unsqueeze(-1).unsqueeze(-1)
if self.gamma is not None:
gamma = self.gamma.repeat((num_examples_per_model, 1)).unsqueeze(-1).unsqueeze(-1)
out = self.conv(x * alpha) * gamma
else:
out = self.conv(x * alpha)
if self.bias is not None:
bias = self.bias.repeat((num_examples_per_model, 1)).unsqueeze(-1).unsqueeze(-1)
out += bias
return out
def extra_repr(self) -> str:
return (
f"in_channels={self.in_channels}, "
f"out_channels={self.out_channels}, "
f"num_estimators={self.num_estimators}, "
f"hidden_size={self.hidden_size}, "
f"gamma={self.gamma is not None}, "
f"bias={self.bias is not None}"
)
