from typing import Literal
import torch.nn.functional as F
from einops import repeat
from torch import Tensor, nn
from torch_uncertainty.layers import MaskedConv2d, MaskedLinear
from .utils import get_resnet_num_blocks
__all__ = [
"masked_resnet",
]
class _BasicBlock(nn.Module):
expansion = 1
def __init__(
self,
in_planes: int,
planes: int,
stride: int,
num_estimators: int,
scale: float,
conv_bias: bool,
dropout_rate: float,
groups: int,
normalization_layer: type[nn.Module],
) -> None:
super().__init__()
self.conv1 = MaskedConv2d(
in_planes,
planes,
kernel_size=3,
num_estimators=num_estimators,
scale=scale,
groups=groups,
stride=stride,
padding=1,
bias=conv_bias,
)
self.bn1 = normalization_layer(planes)
self.conv2 = MaskedConv2d(
planes,
planes,
kernel_size=3,
num_estimators=num_estimators,
scale=scale,
stride=1,
padding=1,
groups=groups,
bias=conv_bias,
)
self.dropout = nn.Dropout2d(p=dropout_rate)
self.bn2 = normalization_layer(planes)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion * planes:
self.shortcut = nn.Sequential(
MaskedConv2d(
in_planes,
self.expansion * planes,
kernel_size=1,
num_estimators=num_estimators,
scale=scale,
stride=stride,
groups=groups,
bias=conv_bias,
),
normalization_layer(self.expansion * planes),
)
def forward(self, x: Tensor) -> Tensor:
out = F.relu(self.dropout(self.bn1(self.conv1(x))))
out = self.bn2(self.conv2(out))
out += self.shortcut(x)
return F.relu(out)
class _Bottleneck(nn.Module):
expansion = 4
def __init__(
self,
in_planes: int,
planes: int,
stride: int,
num_estimators: int,
scale: float,
conv_bias: bool,
dropout_rate: float,
groups: int,
normalization_layer: type[nn.Module],
) -> None:
super().__init__()
self.conv1 = MaskedConv2d(
in_planes,
planes,
kernel_size=1,
num_estimators=num_estimators,
scale=scale,
groups=groups,
bias=conv_bias,
)
self.bn1 = normalization_layer(planes)
self.conv2 = MaskedConv2d(
planes,
planes,
kernel_size=3,
num_estimators=num_estimators,
scale=scale,
stride=stride,
padding=1,
groups=groups,
bias=conv_bias,
)
self.dropout = nn.Dropout2d(p=dropout_rate)
self.bn2 = normalization_layer(planes)
self.conv3 = MaskedConv2d(
planes,
self.expansion * planes,
kernel_size=1,
num_estimators=num_estimators,
scale=scale,
groups=groups,
bias=conv_bias,
)
self.bn3 = normalization_layer(self.expansion * planes)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion * planes:
self.shortcut = nn.Sequential(
MaskedConv2d(
in_planes,
self.expansion * planes,
kernel_size=1,
num_estimators=num_estimators,
scale=scale,
stride=stride,
groups=groups,
bias=conv_bias,
),
normalization_layer(self.expansion * planes),
)
def forward(self, x: Tensor) -> Tensor:
out = F.relu(self.bn1(self.conv1(x)))
out = F.relu(self.dropout(self.bn2(self.conv2(out))))
out = self.bn3(self.conv3(out))
out += self.shortcut(x)
return F.relu(out)
class _MaskedResNet(nn.Module):
def __init__(
self,
block: type[_BasicBlock | _Bottleneck],
num_blocks: list[int],
in_channels: int,
num_classes: int,
num_estimators: int,
dropout_rate: float,
scale: float = 2.0,
conv_bias: bool = True,
groups: int = 1,
style: Literal["imagenet", "cifar"] = "imagenet",
in_planes: int = 64,
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.in_channels = in_channels
self.in_planes = in_planes
block_planes = self.in_planes
self.num_estimators = num_estimators
self.repeat_strategy = repeat_strategy
if style == "imagenet":
self.conv1 = nn.Conv2d(
self.in_channels,
block_planes,
kernel_size=7,
stride=2,
padding=3,
groups=groups,
bias=False,
)
elif style == "cifar":
self.conv1 = nn.Conv2d(
self.in_channels,
block_planes,
kernel_size=3,
stride=1,
padding=1,
groups=groups,
bias=False,
)
else:
raise ValueError(f"Unknown style. Got {style}.")
self.bn1 = normalization_layer(block_planes)
if style == "imagenet":
self.optional_pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
else:
self.optional_pool = nn.Identity()
self.layer1 = self._make_layer(
block,
block_planes,
num_blocks[0],
stride=1,
num_estimators=num_estimators,
conv_bias=conv_bias,
dropout_rate=dropout_rate,
scale=scale,
groups=groups,
normalization_layer=normalization_layer,
)
self.layer2 = self._make_layer(
block,
block_planes * 2,
num_blocks[1],
stride=2,
num_estimators=num_estimators,
conv_bias=conv_bias,
dropout_rate=dropout_rate,
scale=scale,
groups=groups,
normalization_layer=normalization_layer,
)
self.layer3 = self._make_layer(
block,
block_planes * 4,
num_blocks[2],
stride=2,
num_estimators=num_estimators,
conv_bias=conv_bias,
dropout_rate=dropout_rate,
scale=scale,
groups=groups,
normalization_layer=normalization_layer,
)
if len(num_blocks) == 4:
self.layer4 = self._make_layer(
block,
block_planes * 8,
num_blocks[3],
stride=2,
num_estimators=num_estimators,
conv_bias=conv_bias,
dropout_rate=dropout_rate,
scale=scale,
groups=groups,
normalization_layer=normalization_layer,
)
linear_multiplier = 8
else:
self.layer4 = nn.Identity()
linear_multiplier = 4
self.final_dropout = nn.Dropout(p=dropout_rate)
self.pool = nn.AdaptiveAvgPool2d(output_size=1)
self.flatten = nn.Flatten(1)
self.linear = MaskedLinear(
block_planes * linear_multiplier * block.expansion,
num_classes,
num_estimators,
scale=scale,
)
def _make_layer(
self,
block: type[_BasicBlock | _Bottleneck],
planes: int,
num_blocks: int,
stride: int,
num_estimators: int,
conv_bias: bool,
dropout_rate: float,
scale: float,
groups: int,
normalization_layer: type[nn.Module],
) -> nn.Module:
strides = [stride] + [1] * (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,
normalization_layer=normalization_layer,
)
)
self.in_planes = planes * block.expansion
return nn.Sequential(*layers)
def 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 = F.relu(self.bn1(self.conv1(x)))
out = self.optional_pool(out)
out = self.layer1(out)
out = self.layer2(out)
out = self.layer3(out)
out = self.layer4(out)
out = self.pool(out)
out = self.final_dropout(self.flatten(out))
return self.linear(out)
[docs]
def masked_resnet(
in_channels: int,
num_classes: int,
arch: int,
num_estimators: int,
scale: float,
width_multiplier: float = 1.0,
groups: int = 1,
conv_bias: bool = True,
dropout_rate: float = 0,
style: Literal["imagenet", "cifar"] = "imagenet",
normalization_layer: type[nn.Module] = nn.BatchNorm2d,
repeat_strategy: Literal["legacy", "paper"] = "paper",
) -> _MaskedResNet:
"""Masksembles of ResNet.
Args:
in_channels (int): Number of input channels.
num_classes (int): Number of classes to predict.
arch (int): The architecture of the ResNet.
num_estimators (int): Number of estimators in the ensemble.
scale (float): The scale of the mask.
width_multiplier (float): Width multiplier. Defaults to 1.
groups (int): Number of groups within each estimator. Defaults to 1.
conv_bias (bool): Whether to use bias in convolutions. Defaults to
``True``.
dropout_rate (float): Dropout rate. Defaults to ``0``.
style (str, optional): The style of the model. Defaults to ``"imagenet"``.
normalization_layer (nn.Module, optional): Normalization layer.
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:
_MaskedResNet: A Masksembles-style ResNet.
"""
block = _BasicBlock if arch in [18, 20, 34, 44, 56, 110, 1202] else _Bottleneck
in_planes = 16 if arch in [20, 44, 56, 110, 1202] else 64
return _MaskedResNet(
block=block,
num_blocks=get_resnet_num_blocks(arch),
num_classes=num_classes,
in_channels=in_channels,
num_estimators=num_estimators,
scale=scale,
groups=groups,
conv_bias=conv_bias,
dropout_rate=dropout_rate,
style=style,
in_planes=int(in_planes * width_multiplier),
normalization_layer=normalization_layer,
repeat_strategy=repeat_strategy,
)
