from collections.abc import Callable
from typing import Literal
from einops import repeat
from torch import Tensor, nn
from torch.nn.functional import relu
from torch_uncertainty.layers import MaskedConv2d, MaskedLinear
__all__ = [
"masked_wideresnet28x10",
]
class _WideBasicBlock(nn.Module):
def __init__(
self,
in_planes: int,
planes: int,
conv_bias: bool,
dropout_rate: float,
stride: int,
num_estimators: int,
scale: float,
groups: int,
activation_fn: Callable,
normalization_layer: type[nn.Module],
) -> None:
super().__init__()
self.activation_fn = activation_fn
self.conv1 = MaskedConv2d(
in_planes,
planes,
kernel_size=3,
num_estimators=num_estimators,
padding=1,
scale=scale,
groups=groups,
bias=conv_bias,
)
self.dropout = nn.Dropout2d(p=dropout_rate)
self.bn1 = normalization_layer(planes)
self.conv2 = MaskedConv2d(
planes,
planes,
kernel_size=3,
num_estimators=num_estimators,
stride=stride,
padding=1,
scale=scale,
groups=groups,
bias=conv_bias,
)
self.bn2 = normalization_layer(planes)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != planes:
self.shortcut = nn.Sequential(
MaskedConv2d(
in_planes,
planes,
kernel_size=1,
num_estimators=num_estimators,
stride=stride,
scale=scale,
groups=groups,
bias=conv_bias,
),
)
def forward(self, x: Tensor) -> Tensor:
out = self.activation_fn(self.bn1(self.dropout(self.conv1(x))))
out = self.conv2(out)
out += self.shortcut(x)
return self.activation_fn(self.bn2(out))
class _MaskedWideResNet(nn.Module):
def __init__(
self,
depth: int,
widen_factor: int,
in_channels: int,
num_classes: int,
num_estimators: int,
conv_bias: bool,
dropout_rate: float,
scale: float = 2.0,
groups: int = 1,
style: Literal["imagenet", "cifar"] = "imagenet",
activation_fn: Callable = relu,
normalization_layer: type[nn.Module] = nn.BatchNorm2d,
repeat_strategy: Literal["legacy", "paper"] = "legacy",
) -> None:
if repeat_strategy not in ("legacy", "paper"):
raise ValueError(f"Unknown repeat_strategy. Got {repeat_strategy}.")
super().__init__()
self.num_estimators = num_estimators
self.activation_fn = activation_fn
self.in_planes = 16
self.repeat_strategy = repeat_strategy
if (depth - 4) % 6 != 0:
raise ValueError(f"Wide-resnet depth should be 6n+4. Got {depth}.")
num_blocks = (depth - 4) // 6
num_stages = [
16,
16 * widen_factor,
32 * widen_factor,
64 * widen_factor,
]
if style == "imagenet":
self.conv1 = nn.Conv2d(
in_channels,
num_stages[0],
kernel_size=7,
stride=2,
padding=3,
bias=conv_bias,
groups=1,
)
elif style == "cifar":
self.conv1 = nn.Conv2d(
in_channels,
num_stages[0],
kernel_size=3,
stride=1,
padding=1,
bias=conv_bias,
groups=1,
)
else:
raise ValueError(f"Unknown WideResNet style: {style}. ")
self.bn1 = normalization_layer(num_stages[0])
if style == "imagenet":
self.optional_pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
else:
self.optional_pool = nn.Identity()
self.layer1 = self._wide_layer(
_WideBasicBlock,
num_stages[1],
num_blocks=num_blocks,
conv_bias=conv_bias,
dropout_rate=dropout_rate,
stride=1,
num_estimators=self.num_estimators,
scale=scale,
groups=groups,
activation_fn=activation_fn,
normalization_layer=normalization_layer,
)
self.layer2 = self._wide_layer(
_WideBasicBlock,
num_stages[2],
num_blocks=num_blocks,
conv_bias=conv_bias,
dropout_rate=dropout_rate,
stride=2,
num_estimators=self.num_estimators,
scale=scale,
groups=groups,
activation_fn=activation_fn,
normalization_layer=normalization_layer,
)
self.layer3 = self._wide_layer(
_WideBasicBlock,
num_stages[3],
num_blocks=num_blocks,
conv_bias=conv_bias,
dropout_rate=dropout_rate,
stride=2,
num_estimators=self.num_estimators,
scale=scale,
groups=groups,
activation_fn=activation_fn,
normalization_layer=normalization_layer,
)
self.final_dropout = nn.Dropout(p=dropout_rate)
self.pool = nn.AdaptiveAvgPool2d(output_size=1)
self.flatten = nn.Flatten(1)
self.linear = MaskedLinear(num_stages[3], num_classes, num_estimators, scale=scale)
def _wide_layer(
self,
block: type[nn.Module],
planes: int,
num_blocks: int,
conv_bias: bool,
dropout_rate: float,
stride: int,
num_estimators: int,
scale: float,
groups: int,
activation_fn: Callable,
normalization_layer: type[nn.Module],
) -> nn.Module:
strides = [stride] + [1] * (int(num_blocks) - 1)
layers = []
for stride in strides:
layers.append(
block(
in_planes=self.in_planes,
planes=planes,
stride=stride,
num_estimators=num_estimators,
conv_bias=conv_bias,
dropout_rate=dropout_rate,
scale=scale,
groups=groups,
activation_fn=activation_fn,
normalization_layer=normalization_layer,
)
)
self.in_planes = planes
return nn.Sequential(*layers)
def feats_forward(self, x: Tensor) -> Tensor:
if not self.training or self.repeat_strategy == "legacy":
x = repeat(x, "b ... -> (m b) ...", m=self.num_estimators)
out = self.activation_fn(self.bn1(self.conv1(x)))
out = self.optional_pool(out)
out = self.layer1(out)
out = self.layer2(out)
out = self.layer3(out)
out = self.pool(out)
return self.final_dropout(self.flatten(out))
def forward(self, x: Tensor) -> Tensor:
return self.linear(self.feats_forward(x))
[docs]
def masked_wideresnet28x10(
in_channels: int,
num_classes: int,
num_estimators: int,
scale: float,
conv_bias: bool = True,
dropout_rate: float = 0.3,
groups: int = 1,
style: Literal["imagenet", "cifar"] = "imagenet",
activation_fn: Callable = relu,
normalization_layer: type[nn.Module] = nn.BatchNorm2d,
repeat_strategy: Literal["legacy", "paper"] = "paper",
) -> _MaskedWideResNet:
"""Masksembles of Wide-ResNet-28x10.
Args:
in_channels (int): Number of input channels.
num_classes (int): Number of classes to predict.
num_estimators (int): Number of estimators in the ensemble.
scale (float): Expansion factor affecting the width of the estimators.
conv_bias (bool): Whether to use bias in convolutions. Defaults to
``True``.
dropout_rate (float, optional): Dropout rate. Defaults to ``0.3``.
groups (int): Number of groups within each estimator. Defaults to
``1``.
style (bool, optional): Whether to use the ImageNet
structure. Defaults to ``True``.
activation_fn (Callable, optional): Activation function. Defaults to
``torch.nn.functional.relu``.
normalization_layer (nn.Module, optional): Normalization layer.
Defaults to ``torch.nn.BatchNorm2d``.
repeat_strategy ("legacy"|"paper", optional): The repeat
strategy to use during training:
- "legacy": Repeat inputs for each estimator during both
training and evaluation.
- "paper"(default): Repeat inputs for each estimator only during
evaluation.
Returns:
_MaskedWideResNet: A Masksembles-style Wide-ResNet-28x10.
"""
return _MaskedWideResNet(
in_channels=in_channels,
num_classes=num_classes,
depth=28,
widen_factor=10,
conv_bias=conv_bias,
dropout_rate=dropout_rate,
num_estimators=num_estimators,
scale=scale,
groups=groups,
style=style,
activation_fn=activation_fn,
normalization_layer=normalization_layer,
repeat_strategy=repeat_strategy,
)
