✨ Everything as a second level import

tests/baselines/test_standard.py

@@ -163,15 +163,3 @@ def test_standard(self):
             separable=True,
         ).eval()
         _ = net(torch.rand(1, 3, 32, 32))
-
-    def test_errors(self):
-        with pytest.raises(ValueError):
-            DeepLabBaseline(
-                num_classes=10,
-                loss=nn.CrossEntropyLoss(),
-                version="test",
-                style="v3",
-                output_stride=16,
-                arch=50,
-                separable=True,
-            )

tests/models/test_deeplab.py

@@ -1,29 +1,25 @@
 import pytest
 import torch
 
-from torch_uncertainty.models.segmentation.deeplab import (
-    _DeepLabV3,
-    deep_lab_v3_resnet50,
-    deep_lab_v3_resnet101,
-)
+from torch_uncertainty.models.segmentation.deeplab import _DeepLabV3, deep_lab_v3_resnet
 
 
 class TestDeeplab:
     """Testing the Deeplab class."""
 
     @torch.no_grad()
     def test_main(self):
-        model = deep_lab_v3_resnet50(10, "v3", 16, True, False).eval()
+        model = deep_lab_v3_resnet(10, 50, "v3", 16, True, False).eval()
         model(torch.randn(1, 3, 32, 32))
-        model = deep_lab_v3_resnet50(10, "v3", 16, False, False).eval()
-        model = deep_lab_v3_resnet101(10, "v3+", 8, True, False).eval()
+        model = deep_lab_v3_resnet(10, 50, "v3", 16, False, False).eval()
+        model = deep_lab_v3_resnet(10, 101, "v3+", 8, True, False).eval()
         model(torch.randn(1, 3, 32, 32))
-        model = deep_lab_v3_resnet101(10, "v3+", 8, False, False).eval()
+        model = deep_lab_v3_resnet(10, 101, "v3+", 8, False, False).eval()
 
     def test_errors(self):
         with pytest.raises(ValueError, match="Unknown backbone:"):
             _DeepLabV3(10, "other", "v3", 16, True, False)
         with pytest.raises(ValueError, match="output_stride: "):
-            deep_lab_v3_resnet50(10, "v3", 15, True, False)
+            deep_lab_v3_resnet(10, 50, "v3", 15, True, False)
         with pytest.raises(ValueError, match="Unknown style: "):
-            deep_lab_v3_resnet50(10, "v2", 16, True, False)
+            deep_lab_v3_resnet(10, 50, "v2", 16, True, False)

torch_uncertainty/baselines/segmentation/deeplab.py

@@ -2,21 +2,11 @@
 
 from torch import nn
 
-from torch_uncertainty.models.segmentation.deeplab import (
-    deep_lab_v3_resnet50,
-    deep_lab_v3_resnet101,
-)
+from torch_uncertainty.models.segmentation import deep_lab_v3_resnet
 from torch_uncertainty.routines.segmentation import SegmentationRoutine
 
 
 class DeepLabBaseline(SegmentationRoutine):
-    single = ["std"]
-    versions = {
-        "std": [
-            deep_lab_v3_resnet50,
-            deep_lab_v3_resnet101,
-        ]
-    }
     archs = [50, 101]
 
     def __init__(
@@ -35,6 +25,7 @@ def __init__(
     ) -> None:
         params = {
             "num_classes": num_classes,
+            "arch": arch,
             "style": style,
             "output_stride": output_stride,
             "separable": separable,
@@ -43,10 +34,7 @@ def __init__(
 
         format_batch_fn = nn.Identity()
 
-        if version not in self.versions:
-            raise ValueError(f"Unknown version {version}")
-
-        model = self.versions[version][self.archs.index(arch)](**params)
+        model = deep_lab_v3_resnet(**params)
         super().__init__(
             num_classes=num_classes,
             model=model,

torch_uncertainty/datamodules/__init__.py

@@ -11,6 +11,7 @@
     OnlineShoppersDataModule,
     SpamBaseDataModule,
     TinyImageNetDataModule,
+    UCIClassificationDataModule,
 )
-from .segmentation import CamVidDataModule, CityscapesDataModule
+from .segmentation import CamVidDataModule, CityscapesDataModule, MUADDataModule
 from .uci_regression import UCIRegressionDataModule

torch_uncertainty/datamodules/classification/__init__.py

@@ -10,4 +10,5 @@
     HTRU2DataModule,
     OnlineShoppersDataModule,
     SpamBaseDataModule,
+    UCIClassificationDataModule,
 )

torch_uncertainty/datasets/__init__.py

@@ -1,7 +1,31 @@
 # ruff: noqa: F401
 from .aggregated_dataset import AggregatedDataset
+from .classification import (
+    CIFAR10C,
+    CIFAR10H,
+    CIFAR10N,
+    CIFAR100C,
+    CIFAR100N,
+    CUB,
+    HTRU2,
+    MNISTC,
+    BankMarketing,
+    DOTA2Games,
+    ImageNetA,
+    ImageNetC,
+    ImageNetO,
+    ImageNetR,
+    NotMNIST,
+    OnlineShoppers,
+    OpenImageO,
+    SpamBase,
+    TinyImageNet,
+    TinyImageNetC,
+)
 from .fractals import Fractals
 from .frost import FrostImages
 from .kitti import KITTIDepth
 from .muad import MUAD
 from .nyu import NYUv2
+from .regression import UCIRegression
+from .segmentation import CamVid, Cityscapes

torch_uncertainty/layers/__init__.py

@@ -2,6 +2,19 @@
 from .batch_ensemble import BatchConv2d, BatchLinear
 from .bayesian import BayesConv1d, BayesConv2d, BayesConv3d, BayesLinear
 from .channel_layer_norm import ChannelLayerNorm
+from .distributions import (
+    CauchyConvNd,
+    CauchyLinear,
+    LaplaceConvNd,
+    LaplaceLinear,
+    NormalConvNd,
+    NormalInverseGammaConvNd,
+    NormalInverseGammaLinear,
+    NormalLinear,
+    StudentTConvNd,
+    StudentTLinear,
+)
+from .filter_response_norm import FilterResponseNorm1d, FilterResponseNorm2d, FilterResponseNorm3d
 from .masksembles import MaskedConv2d, MaskedLinear
 from .modules import Identity
 from .packed import (

torch_uncertainty/models/__init__.py

@@ -1,4 +1,24 @@
 # ruff: noqa: F401
+from .classification import (
+    batched_resnet,
+    batched_wideresnet28x10,
+    batchensemble_lenet,
+    bayesian_lenet,
+    lenet,
+    lpbnn_resnet,
+    masked_resnet,
+    masked_wideresnet28x10,
+    mimo_resnet,
+    mimo_wideresnet28x10,
+    packed_lenet,
+    packed_resnet,
+    packed_vgg,
+    packed_wideresnet28x10,
+    resnet,
+    vgg,
+    wideresnet28x10,
+)
+from .depth import bts_resnet
 from .wrappers import (
     EMA,
     EPOCH_UPDATE_MODEL,
@@ -9,6 +29,7 @@
     CheckpointCollector,
     MCDropout,
     StochasticModel,
+    Zero,
     batch_ensemble,
     deep_ensembles,
     mc_dropout,

torch_uncertainty/models/classification/__init__.py

@@ -1,5 +1,11 @@
-# ruff: noqa: F401, F403
-from .lenet import *
-from .resnet import *
-from .vgg import *
-from .wideresnet import *
+# ruff: noqa: F401
+from .lenet import batchensemble_lenet, bayesian_lenet, lenet, packed_lenet
+from .resnet import batched_resnet, lpbnn_resnet, masked_resnet, mimo_resnet, packed_resnet, resnet
+from .vgg import packed_vgg, vgg
+from .wideresnet import (
+    batched_wideresnet28x10,
+    masked_wideresnet28x10,
+    mimo_wideresnet28x10,
+    packed_wideresnet28x10,
+    wideresnet28x10,
+)

torch_uncertainty/models/classification/lenet.py

@@ -8,8 +8,8 @@
 from torch_uncertainty.layers.bayesian import BayesConv2d, BayesLinear
 from torch_uncertainty.layers.mc_batch_norm import MCBatchNorm2d
 from torch_uncertainty.layers.packed import PackedConv2d, PackedLinear
-from torch_uncertainty.models import StochasticModel
 from torch_uncertainty.models.wrappers.batch_ensemble import BatchEnsemble
+from torch_uncertainty.models.wrappers.stochastic import StochasticModel
 
 __all__ = ["batchensemble_lenet", "bayesian_lenet", "lenet", "packed_lenet"]
 

torch_uncertainty/models/classification/resnet/__init__.py

@@ -1,4 +1,4 @@
-# ruff: noqa: F401, F403
+# ruff: noqa: F401
 from .batched import batched_resnet
 from .lpbnn import lpbnn_resnet
 from .masked import masked_resnet

torch_uncertainty/models/classification/vgg/__init__.py

@@ -1,3 +1,3 @@
-# ruff: noqa: F401, F403
+# ruff: noqa: F401
 from .packed import packed_vgg
 from .std import vgg

torch_uncertainty/models/classification/wideresnet/__init__.py

@@ -1,6 +1,6 @@
-# ruff: noqa: F401, F403
-from .batched import *
-from .masked import *
-from .mimo import *
-from .packed import *
-from .std import *
+# ruff: noqa: F401
+from .batched import batched_wideresnet28x10
+from .masked import masked_wideresnet28x10
+from .mimo import mimo_wideresnet28x10
+from .packed import packed_wideresnet28x10
+from .std import wideresnet28x10

torch_uncertainty/models/depth/__init__.py

@@ -0,0 +1,2 @@
+# ruff: noqa: F401
+from .bts import bts_resnet

torch_uncertainty/models/depth/bts.py

@@ -576,7 +576,8 @@ def _bts(
     return _BTS(backbone_name, max_depth, bts_size, dist_family, pretrained_backbone)
 
 
-def bts_resnet50(
+def bts_resnet(
+    arch: int,
     max_depth: float,
     bts_size: int = 512,
     dist_family: str | None = None,
@@ -585,36 +586,14 @@ def bts_resnet50(
     """BTS model with ResNet-50 backbone.
 
     Args:
+        arch (int): The number of layers of the underlying ResNet model: 50 or 101.
         max_depth (float): Maximum predicted depth.
         bts_size (int): BTS feature size. Defaults to 512.
         dist_family (str): Distribution family name. Defaults to None.
         pretrained_backbone (bool): Use a pretrained backbone. Defaults to True.
     """
     return _bts(
-        "resnet50",
-        max_depth,
-        bts_size=bts_size,
-        dist_family=dist_family,
-        pretrained_backbone=pretrained_backbone,
-    )
-
-
-def bts_resnet101(
-    max_depth: float,
-    bts_size: int = 512,
-    dist_family: str | None = None,
-    pretrained_backbone: bool = True,
-) -> _BTS:
-    """BTS model with ResNet-101 backbone.
-
-    Args:
-        max_depth (float): Maximum predicted depth.
-        bts_size (int): BTS feature size. Defaults to 512.
-        dist_family (str): Distribution family name. Defaults to None.
-        pretrained_backbone (bool): Use a pretrained backbone. Defaults to True.
-    """
-    return _bts(
-        "resnet101",
+        f"resnet{arch}",
         max_depth,
         bts_size=bts_size,
         dist_family=dist_family,

torch_uncertainty/models/segmentation/__init__.py

@@ -0,0 +1,2 @@
+# ruff: noqa: F401
+from .deeplab import deep_lab_v3_resnet

torch_uncertainty/models/segmentation/deeplab.py

@@ -352,17 +352,19 @@ def forward(self, x: Tensor) -> Tensor:
         )
 
 
-def deep_lab_v3_resnet50(
+def deep_lab_v3_resnet(
     num_classes: int,
+    arch: int,
     style: Literal["v3", "v3+"],
     output_stride: int = 16,
     separable: bool = False,
     pretrained_backbone: bool = True,
 ) -> _DeepLabV3:
-    """DeepLab V3(+) model with ResNet-50 backbone.
+    """DeepLab V3(+) model with ResNet-50/101 backbone.
 
     Args:
         num_classes (int): Number of classes.
+        arch (int): Number of layers of the underlying ResNet model: 50 or 101.
         style (Literal["v3", "v3+"]): Whether to use a DeepLab V3 or V3+ model.
         output_stride (int, optional): Output stride. Defaults to 16.
         separable (bool, optional): Use separable convolutions. Defaults to
@@ -372,33 +374,7 @@ def deep_lab_v3_resnet50(
     """
     return _DeepLabV3(
         num_classes,
-        "resnet50",
-        style,
-        output_stride=output_stride,
-        separable=separable,
-        pretrained_backbone=pretrained_backbone,
-    )
-
-
-def deep_lab_v3_resnet101(
-    num_classes: int,
-    style: Literal["v3", "v3+"],
-    output_stride: int = 16,
-    separable: bool = False,
-    pretrained_backbone: bool = True,
-) -> _DeepLabV3:
-    """DeepLab V3(+) model with ResNet-50 backbone.
-
-    Args:
-        num_classes (int): Number of classes.
-        style (Literal["v3", "v3+"]): Whether to use a DeepLab V3 or V3+ model.
-        output_stride (int, optional): Output stride. Defaults to 16.
-        separable (bool, optional): Use separable convolutions. Defaults to False.
-        pretrained_backbone (bool, optional): Use pretrained backbone. Defaults to True.
-    """
-    return _DeepLabV3(
-        num_classes,
-        "resnet101",
+        f"resnet{arch}",
         style,
         output_stride=output_stride,
         separable=separable,

torch_uncertainty/transforms/__init__.py

@@ -1,5 +1,23 @@
 # ruff: noqa: F401
 from .batch import MIMOBatchFormat, RepeatTarget
+from .corruption import Brightness as BrightnessCorruption
+from .corruption import Contrast as ContrastCorruption
+from .corruption import (
+    DefocusBlur,
+    Elastic,
+    Fog,
+    Frost,
+    GaussianNoise,
+    GlassBlur,
+    ImpulseNoise,
+    JPEGCompression,
+    MotionBlur,
+    Pixelate,
+    ShotNoise,
+    Snow,
+    ZoomBlur,
+    corruption_transforms,
+)
 from .cutout import Cutout
 from .image import (
     AutoContrast,

torch_uncertainty/utils/__init__.py

@@ -1,7 +1,10 @@
 # ruff: noqa: F401
 from .checkpoints import get_version
 from .cli import TULightningCLI
+from .distributions import NormalInverseGamma, get_dist_class, get_dist_estimate
+from .evaluation_loop import TUEvaluationLoop
 from .hub import load_hf
-from .misc import csv_writer, plot_hist
+from .misc import csv_writer
+from .plotting import plot_hist, show
 from .trainer import TUTrainer
 from .transforms import interpolation_modes_from_str