Restructuring of vision modules

.gitignore

@@ -45,3 +45,9 @@ nas_results/
 
 .nv
 .tmp/
+tmp/
+
+.torch
+type_info.json
+*.log
+.pytest_cache

experiments/kvasir/config.yaml

@@ -20,7 +20,7 @@ defaults:
     - base_config
     - override dataset: kvasir_capsule   # Dataset configuration name
     - override features: identity        # Feature extractor configuration name (use identity for vision datasets)
-    - override model: timm_resnet152      # Neural network name (for now timm_resnet50 or timm_efficientnet_lite1)
+    - override model: timm_mobilenetv3_small_075      # Neural network name (for now timm_resnet50 or timm_efficientnet_lite1)
     - override scheduler: 1cycle         # learning rate scheduler config name
     - override optimizer: adamw          # Optimizer config name
     - override normalizer: null          # Feature normalizer (used for quantized neural networks)

experiments/kvasir_ae/config.yaml

@@ -20,7 +20,7 @@ defaults:
     - base_config
     - override dataset: kvasir_unlabeled   # Dataset configuration name
     - override features: identity        # Feature extractor configuration name (use identity for vision datasets)
-    - override model: timm_resnet152      # Neural network name (for now timm_resnet50 or timm_efficientnet_lite1)
+    - override model: timm_mobilenetv3_small_075      # Neural network name (for now timm_resnet50 or timm_efficientnet_lite1)
     - override scheduler: 1cycle         # learning rate scheduler config name
     - override optimizer: adamw          # Optimizer config name
     - override normalizer: null          # Feature normalizer (used for quantized neural networks)

hannah/conf/dataset/fake2d.yaml

@@ -21,4 +21,5 @@
 data_folder: ${hydra:runtime.cwd}/datasets/
 cls: hannah.datasets.vision.FakeDataset
 dataset: fake2d
+resolution: 320
 num_classes: 10

hannah/conf/module/augmentation/kvasir_ae_augment.yaml

@@ -0,0 +1,27 @@
+##
+## Copyright (c) 2022 University of Tübingen.
+##
+## This file is part of hannah.
+## See https://atreus.informatik.uni-tuebingen.de/ties/ai/hannah/hannah for further info.
+##
+## Licensed under the Apache License, Version 2.0 (the "License");
+## you may not use this file except in compliance with the License.
+## You may obtain a copy of the License at
+##
+##     http://www.apache.org/licenses/LICENSE-2.0
+##
+## Unless required by applicable law or agreed to in writing, software
+## distributed under the License is distributed on an "AS IS" BASIS,
+## WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+## See the License for the specific language governing permissions and
+## limitations under the License.
+##
+batch_augment:
+  transforms:
+    RandomRotation:
+      degrees: 180
+      p: 1.0
+    RandomVerticalFlip:
+      p: 0.5
+    RandomHorizontalFlip:
+      p: 0.5

hannah/conf/module/augmentation/kvasir_ssl_augment.yaml

@@ -0,0 +1,33 @@
+##
+## Copyright (c) 2022 University of Tübingen.
+##
+## This file is part of hannah.
+## See https://atreus.informatik.uni-tuebingen.de/ties/ai/hannah/hannah for further info.
+##
+## Licensed under the Apache License, Version 2.0 (the "License");
+## you may not use this file except in compliance with the License.
+## You may obtain a copy of the License at
+##
+##     http://www.apache.org/licenses/LICENSE-2.0
+##
+## Unless required by applicable law or agreed to in writing, software
+## distributed under the License is distributed on an "AS IS" BASIS,
+## WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+## See the License for the specific language governing permissions and
+## limitations under the License.
+##
+batch_augment:
+  transforms:
+    RandomGaussianNoise:
+      p: 0.5
+    RandomPosterize:
+      p: 0.3
+    RandomSharpness:
+      p: 0.3
+    RandomRotation:
+      degrees: 180
+      p: 1.0
+    RandomVerticalFlip:
+      p: 0.5
+    RandomHorizontalFlip:
+      p: 0.5

hannah/conf/module/augmentation/mixup.yaml

@@ -0,0 +1,25 @@
+##
+## Copyright (c) 2022 University of Tübingen.
+##
+## This file is part of hannah.
+## See https://atreus.informatik.uni-tuebingen.de/ties/ai/hannah/hannah for further info.
+##
+## Licensed under the Apache License, Version 2.0 (the "License");
+## you may not use this file except in compliance with the License.
+## You may obtain a copy of the License at
+##
+##     http://www.apache.org/licenses/LICENSE-2.0
+##
+## Unless required by applicable law or agreed to in writing, software
+## distributed under the License is distributed on an "AS IS" BASIS,
+## WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+## See the License for the specific language governing permissions and
+## limitations under the License.
+##
+mixup:
+    mixup_alpha: 0.1       # mixup alpha value, mixup is active if > 0.
+    cutmix_alpha: 0.1      # cutmix alpha value, cutmix is active if > 0.
+    prob: 1.0              # Probability of performing mixup or cutmix when either/both is enabled
+    switch_prob: 0.0       # Probability of switching to cutmix when both mixup and cutmix enabled
+    mode: pair             # batch, pair or elem - wise
+    label_smoothing : 0.0  # Apply label smoothing

hannah/conf/module/augmentation/rand_augment.yaml

@@ -0,0 +1,36 @@
+##
+## Copyright (c) 2022 University of Tübingen.
+##
+## This file is part of hannah.
+## See https://atreus.informatik.uni-tuebingen.de/ties/ai/hannah/hannah for further info.
+##
+## Licensed under the Apache License, Version 2.0 (the "License");
+## you may not use this file except in compliance with the License.
+## You may obtain a copy of the License at
+##
+##     http://www.apache.org/licenses/LICENSE-2.0
+##
+## Unless required by applicable law or agreed to in writing, software
+## distributed under the License is distributed on an "AS IS" BASIS,
+## WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+## See the License for the specific language governing permissions and
+## limitations under the License.
+##
+rand_augment:
+    N: 1
+    M: 1
+    probs:
+        Identity : 0.5
+        ShearX: 0.5
+        ShearY: 0.5
+        TranslateX: 0.5
+        TranslateY : 0.5
+        Rotate : 0.5
+        Brightness : 0.0
+        Color : 0.5
+        Contrast : 0.0
+        Sharpness : 0.5
+        Posterize : 0.0
+        Solarize : 0.0
+        AutoContrast : 0.0
+        Equalize : 0.1

hannah/datasets/base.py

@@ -123,22 +123,6 @@ def mean(self) -> Optional[Tuple[int, ...]]:
         """Returns channel-wise means for dataset if applicable"""
         return None
 
-    def get_mixup_fn(
-        self,
-    ) -> Optional[
-        Callable[[torch.Tensor, torch.Tensor], Tuple[torch.Tensor, torch.Tensor]]
-    ]:
-        """Returns an optional dataset specific mixup function"""
-        return None
-
-    def get_batch_augment_fn(
-        self,
-    ) -> Optional[
-        Callable[[torch.Tensor, torch.Tensor], Tuple[torch.Tensor, torch.Tensor]]
-    ]:
-        """Returns an optional dataset specific batch augment function"""
-        return None
-
     @property
     def weights(self) -> Optional[List[float]]:
         """Class weights for weighted sampling"""

hannah/datasets/vision/base.py

@@ -23,6 +23,7 @@
 from collections import Counter, namedtuple
 from typing import Dict, List
 
+import albumentations as A
 import cv2
 import numpy as np
 import pandas as pd
@@ -32,7 +33,6 @@
 from sklearn.model_selection import train_test_split
 from timm.data.mixup import Mixup
 
-import albumentations as A
 from hannah.modules.augmentation import rand_augment
 from hannah.modules.augmentation.batch_augmentation import BatchAugmentationPipeline
 
@@ -41,38 +41,18 @@
 logger = logging.getLogger(__name__)
 
 
-class AugmentationMixin:
-    def get_mixup_fn(self):
-        mixup = None
-
-        augmentation_config = self.config.get("augmentation", None)
-        if augmentation_config is not None:
-            mixup_config = augmentation_config.get("mixup", None)
-            if mixup_config is not None:
-                mixup_config = Mixup(**mixup_config)
-
-        return mixup
-
-    def get_batch_augment_fn(self):
-
-        batch_augment = None
-
-        augmentation_config = self.config.get("augmentation", None)
-        if augmentation_config is not None:
-            batch_augment_config = self.config.get("batch_augment", None)
-            if batch_augment_config is not None:
-
-                batch_augment = BatchAugmentationPipeline(
-                    transforms=batch_augment_config.get("batch_transforms", [])
-                )
-
-        return batch_augment
-
-
-class VisionDatasetBase(AbstractDataset, AugmentationMixin):
+class VisionDatasetBase(AbstractDataset):
     def __init__(self, config):
         self.config = config
 
+    @property
+    def std(self):
+        pass
+
+    @property
+    def mean(self):
+        pass
+
 
 class TorchvisionDatasetBase(VisionDatasetBase):
     """Wrapper around torchvision classification datasets"""
@@ -88,7 +68,7 @@ def class_counts(self):
 
     def __getitem__(self, index):
         data, target = self.dataset[index]
-        data = np.array(data)
+        data = np.array(data) / 255
         if self.transform:
             data = self.transform(image=data)["image"]
         return data, target
@@ -115,19 +95,14 @@ def __init__(self, X, y, classes, transform=None):
         self.X = X
         self.y = y
         self.classes = classes
-        self.transform = (
-            transform if transform else A.Compose([A.Normalize(), ToTensorV2()])
-        )
+        self.transform = transform if transform else A.Compose([ToTensorV2()])
         self.label_to_index = {k: v for v, k in enumerate(classes)}
 
     def __getitem__(self, index):
         image = cv2.imread(str(self.X[index]))
-        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB).astype(np.float32) / 255
         label = self.y[index]
-        if self.transform:
-            data = self.transform(image=image)["image"]
-        else:
-            data = image
+        data = self.transform(image=image)["image"]
         target = self.label_to_index[label]
         return {"data": data, "labels": target}
 

hannah/datasets/vision/cifar.py

@@ -58,30 +58,6 @@ def class_names(self):
     def splits(cls, config):
         data_folder = config.data_folder
         root_folder = os.path.join(data_folder, "CIFAR10")
-
-        # print(loaded_transform)
-        train_transform = A.Compose(
-            [
-                A.SmallestMaxSize(max_size=32),
-                A.ShiftScaleRotate(
-                    shift_limit=0.05, scale_limit=0.05, rotate_limit=15, p=0.5
-                ),
-                A.RandomCrop(height=32, width=32),
-                A.RGBShift(r_shift_limit=15, g_shift_limit=15, b_shift_limit=15, p=0.5),
-                A.RandomBrightnessContrast(p=0.5),
-                A.Normalize(mean=(0.4914, 0.4822, 0.4465), std=(0.247, 0.243, 0.261)),
-                ToTensorV2(),
-            ]
-        )
-
-        val_transform = A.Compose(
-            [
-                A.SmallestMaxSize(max_size=32),
-                A.Normalize(mean=(0.4914, 0.4822, 0.4465), std=(0.247, 0.243, 0.261)),
-                ToTensorV2(),
-            ]
-        )
-
         test_set = torchvision.datasets.CIFAR10(
             root_folder, train=False, download=False
         )
@@ -97,7 +73,7 @@ def splits(cls, config):
         train_set, val_set = data.random_split(train_val_set, split_sizes)
 
         return (
-            cls(config, train_set, train_transform),
-            cls(config, val_set, val_transform),
-            cls(config, test_set, val_transform),
+            cls(config, train_set),
+            cls(config, val_set),
+            cls(config, test_set),
         )

hannah/datasets/vision/kvasir.py

@@ -23,14 +23,14 @@
 from collections import Counter, namedtuple
 from typing import Dict, List
 
-import albumentations as A
 import cv2
 import pandas as pd
 import requests
 import torchvision
 from albumentations.pytorch import ToTensorV2
 from sklearn.model_selection import train_test_split
 
+import albumentations as A
 from hannah.modules.augmentation import rand_augment
 
 from .base import ImageDatasetBase
@@ -104,27 +104,6 @@ def splits(cls, config):
         if isinstance(resolution, int):
             resolution = (resolution, resolution)
 
-        res_x, res_y = resolution
-
-        # FIXME(gerum):  add back rand augment
-        train_transform = A.Compose(
-            [
-                # A.Resize(int(res_x*1.1), int(res_y*1.1)),
-                # A.RandomResizedCrop(res_x, res_y),
-                A.Resize(res_x, res_y),
-                A.Normalize(mean=config.normalize.mean, std=config.normalize.std),
-                ToTensorV2(),
-            ]
-        )
-
-        test_transform = A.Compose(
-            [
-                A.Resize(res_x, res_y),
-                A.Normalize(mean=config.normalize.mean, std=config.normalize.std),
-                ToTensorV2(),
-            ]
-        )
-
         label_to_folder = {
             "Angiectasia": "Angiectasia",
             "Pylorus": "Pylorus",
@@ -232,18 +211,15 @@ def relable_anomaly(X):
                 train_images,
                 train_labels,
                 classes,
-                transform=train_transform,
             ),
             cls(
                 val_images,
                 val_labels,
                 classes,
-                transform=test_transform,
             ),
             cls(
                 test_images,
                 test_labels,
                 classes,
-                transform=test_transform,
             ),
         )