from collections.abc import Callable
from typing import Literal
from torch import Tensor, nn, relu
from torch_uncertainty.layers.bayesian.lpbnn import LPBNNConv2d, LPBNNLinear
from .utils import get_resnet_num_blocks
__all__ = [
"lpbnn_resnet",
]
class _BasicBlock(nn.Module):
expansion = 1
def __init__(
self,
in_planes: int,
planes: int,
stride: int,
dropout_rate: float,
num_estimators: int,
groups: int,
activation_fn: Callable,
normalization_layer: type[nn.Module],
conv_bias: bool,
) -> None:
super().__init__()
self.activation_fn = activation_fn
self.conv1 = LPBNNConv2d(
in_planes,
planes,
kernel_size=3,
num_estimators=num_estimators,
groups=groups,
stride=stride,
padding=1,
bias=conv_bias,
)
self.bn1 = normalization_layer(planes)
self.dropout = nn.Dropout2d(p=dropout_rate)
self.conv2 = LPBNNConv2d(
planes,
planes,
kernel_size=3,
num_estimators=num_estimators,
groups=groups,
stride=1,
padding=1,
bias=conv_bias,
)
self.bn2 = normalization_layer(planes)
self.shortcut = nn.Sequential()
if stride != 1 or in_planes != self.expansion * planes:
self.shortcut = nn.Sequential(
nn.Conv2d(
in_planes,
self.expansion * planes,
groups=groups,
kernel_size=1,
stride=stride,
bias=conv_bias,
),
normalization_layer(self.expansion * planes),
)
def forward(self, inputs: Tensor) -> Tensor:
out = self.activation_fn(self.dropout(self.bn1(self.conv1(inputs))))
out = self.bn2(self.conv2(out))
out += self.shortcut(inputs)
return self.activation_fn(out)
class _Bottleneck(nn.Module):
expansion = 4
def __init__(
self,
in_planes: int,
planes: int,
stride: int,
num_estimators: int,
dropout_rate: float,
groups: int,
activation_fn: Callable,
normalization_layer: type[nn.Module],
conv_bias: bool,
) -> None:
super().__init__()
self.activation_fn = activation_fn
self.conv1 = LPBNNConv2d(
in_planes,
planes,
kernel_size=1,
num_estimators=num_estimators,
groups=groups,
bias=conv_bias,
)
self.bn1 = normalization_layer(planes)
self.conv2 = LPBNNConv2d(
planes,
planes,
kernel_size=3,
num_estimators=num_estimators,
groups=groups,
stride=stride,
padding=1,
bias=conv_bias,
)
self.bn2 = normalization_layer(planes)
self.dropout = nn.Dropout2d(p=dropout_rate)
self.conv3 = LPBNNConv2d(
planes,
self.expansion * planes,
num_estimators=num_estimators,
groups=groups,
kernel_size=1,
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(
LPBNNConv2d(
in_planes,
self.expansion * planes,
kernel_size=1,
num_estimators=num_estimators,
groups=groups,
stride=stride,
bias=conv_bias,
),
normalization_layer(self.expansion * planes),
)
def forward(self, x: Tensor) -> Tensor:
out = self.activation_fn(self.bn1(self.conv1(x)))
out = self.activation_fn(self.dropout(self.bn2(self.conv2(out))))
out = self.bn3(self.conv3(out))
out += self.shortcut(x)
return self.activation_fn(out)
class _LPBNNResNet(nn.Module):
def __init__(
self,
block: type[_BasicBlock | _Bottleneck],
num_blocks: list[int],
in_channels: int,
num_estimators: int,
num_classes: int,
conv_bias: bool,
dropout_rate: float,
groups: int,
style: Literal["imagenet", "cifar"] = "imagenet",
in_planes: int = 64,
activation_fn: Callable = relu,
normalization_layer: type[nn.Module] = nn.BatchNorm2d,
) -> None:
super().__init__()
self.in_planes = in_planes
block_planes = in_planes
self.dropout_rate = dropout_rate
self.activation_fn = activation_fn
self.num_estimators = num_estimators
if style == "imagenet":
self.conv1 = LPBNNConv2d(
in_channels,
block_planes,
kernel_size=7,
stride=2,
padding=3,
num_estimators=num_estimators,
groups=groups,
bias=conv_bias,
)
elif style == "cifar":
self.conv1 = LPBNNConv2d(
in_channels,
block_planes,
kernel_size=3,
stride=1,
padding=1,
num_estimators=num_estimators,
groups=groups,
bias=conv_bias,
)
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,
dropout_rate=dropout_rate,
groups=groups,
activation_fn=activation_fn,
normalization_layer=normalization_layer,
conv_bias=conv_bias,
num_estimators=num_estimators,
)
self.layer2 = self._make_layer(
block,
block_planes * 2,
num_blocks[1],
stride=2,
dropout_rate=dropout_rate,
groups=groups,
activation_fn=activation_fn,
normalization_layer=normalization_layer,
conv_bias=conv_bias,
num_estimators=num_estimators,
)
self.layer3 = self._make_layer(
block,
block_planes * 4,
num_blocks[2],
stride=2,
dropout_rate=dropout_rate,
groups=groups,
activation_fn=activation_fn,
normalization_layer=normalization_layer,
conv_bias=conv_bias,
num_estimators=num_estimators,
)
if len(num_blocks) == 4:
self.layer4 = self._make_layer(
block,
block_planes * 8,
num_blocks[3],
stride=2,
dropout_rate=dropout_rate,
groups=groups,
activation_fn=activation_fn,
normalization_layer=normalization_layer,
conv_bias=conv_bias,
num_estimators=num_estimators,
)
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 = LPBNNLinear(
block_planes * linear_multiplier * block.expansion,
num_classes,
num_estimators=num_estimators,
)
def _make_layer(
self,
block: type[_BasicBlock | _Bottleneck],
planes: int,
num_blocks: int,
stride: int,
num_estimators: int,
dropout_rate: float,
groups: int,
activation_fn: Callable,
normalization_layer: type[nn.Module],
conv_bias: bool,
) -> 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,
dropout_rate=dropout_rate,
groups=groups,
activation_fn=activation_fn,
normalization_layer=normalization_layer,
conv_bias=conv_bias,
num_estimators=num_estimators,
)
)
self.in_planes = planes * block.expansion
return nn.Sequential(*layers)
def feats_forward(self, x: Tensor) -> Tensor:
out = x.repeat(self.num_estimators, 1, 1, 1)
out = self.activation_fn(self.bn1(self.conv1(out)))
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)
return self.final_dropout(self.flatten(out))
def forward(self, x: Tensor) -> Tensor:
return self.linear(self.feats_forward(x))
[docs]
def lpbnn_resnet(
in_channels: int,
num_classes: int,
arch: int,
num_estimators: int,
dropout_rate: float = 0,
conv_bias: bool = True,
width_multiplier: float = 1.0,
groups: int = 1,
style: Literal["imagenet", "cifar"] = "imagenet",
) -> _LPBNNResNet:
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 _LPBNNResNet(
block=block,
num_blocks=get_resnet_num_blocks(arch),
in_channels=in_channels,
num_estimators=num_estimators,
num_classes=num_classes,
dropout_rate=dropout_rate,
conv_bias=conv_bias,
groups=groups,
style=style,
in_planes=int(in_planes * width_multiplier),
)
