v0.5.1: adds several datasets from the WILDS benchmark

Author: KaiyangZhou

README.md

@@ -20,6 +20,7 @@ A drawback of Dassl is that it doesn't (yet? hmm) support distributed multi-GPU
 We don't provide detailed documentations for Dassl, unlike another [project](https://kaiyangzhou.github.io/deep-person-reid/) of ours. This is because Dassl is developed for research purpose and as a researcher, we think it's important to be able to read source code and we highly encourage you to do so---definitely not because we are lazy. :-)
 
 ## What's new
+- Jun 2022: Adds three datasets from the [WILDS](https://wilds.stanford.edu/) benchmark: IWildCam, FMoW and Camelyon17.
 - May 2022: A new domain generalization method [DDG](https://arxiv.org/abs/2205.13913) developed by [Zhishu Sun](https://github.com/siaimes) and to appear at IJCAI'22 is added to this repo. See [here](https://github.com/MetaVisionLab/DDG) for more details.
 - Mar 2022: A new domain generalization method [EFDM](https://arxiv.org/abs/2203.07740) developed by [Yabin Zhang (PolyU)](https://ybzh.github.io/) and to appear at CVPR'22 is added to this repo. See [here](https://github.com/KaiyangZhou/Dassl.pytorch/pull/36) for more details.
 - Feb 2022: In case you don't know, a class in the painting domain of DomainNet (the official splits) only has test images (no training images), which could affect performance. See section 4.a in our [paper](https://arxiv.org/abs/2003.07325) for more details.

configs/datasets/dg/camelyon17.yaml

@@ -0,0 +1,6 @@
+INPUT:
+  SIZE: (224, 224)
+  TRANSFORMS: ["random_resized_crop", "random_flip", "normalize"]
+
+DATASET:
+  NAME: "Camelyon17"

configs/datasets/dg/fmow.yaml

@@ -0,0 +1,6 @@
+INPUT:
+  SIZE: (224, 224)
+  TRANSFORMS: ["random_resized_crop", "random_flip", "normalize"]
+
+DATASET:
+  NAME: "FMoW"

configs/datasets/dg/iwildcam.yaml

@@ -0,0 +1,6 @@
+INPUT:
+  SIZE: (224, 224)
+  TRANSFORMS: ["random_resized_crop", "random_flip", "normalize"]
+
+DATASET:
+  NAME: "IWildCam"

dassl/__init__.py

@@ -13,6 +13,6 @@
 }
 """
 
-__version__ = "0.5.0"
+__version__ = "0.5.1"
 __author__ = "Kaiyang Zhou"
 __homepage__ = "https://kaiyangzhou.github.io/"

dassl/data/datasets/base_dataset.py

@@ -60,7 +60,6 @@ def __init__(self, train_x=None, train_u=None, val=None, test=None):
         self._train_u = train_u  # unlabeled training data (optional)
         self._val = val  # validation data (optional)
         self._test = test  # test data
-
         self._num_classes = self.get_num_classes(train_x)
         self._lab2cname, self._classnames = self.get_lab2cname(train_x)
 

dassl/data/datasets/dg/__init__.py

@@ -1,5 +1,6 @@
 from .pacs import PACS
 from .vlcs import VLCS
+from .wilds import *
 from .cifar_c import CIFAR10C, CIFAR100C
 from .digits_dg import DigitsDG
 from .digit_single import DigitSingle

dassl/data/datasets/dg/wilds/__init__.py

@@ -0,0 +1,3 @@
+from .fmow import FMoW
+from .iwildcam import IWildCam
+from .camelyon17 import Camelyon17

dassl/data/datasets/dg/wilds/camelyon17.py

@@ -0,0 +1,24 @@
+from dassl.data.datasets import DATASET_REGISTRY
+
+from .wilds_base import WILDSBase
+
+
+@DATASET_REGISTRY.register()
+class Camelyon17(WILDSBase):
+    """Tumor tissue recognition.
+
+    2 classes (whether a given region of tissue contains tumor tissue).
+
+    Reference:
+        - Bandi et al. "From detection of individual metastases to classification of lymph
+        node status at the patient level: the CAMELYON17 challenge." TMI 2021.
+        - Koh et al. "Wilds: A benchmark of in-the-wild distribution shifts." ICML 2021.
+    """
+
+    dataset_dir = "camelyon17_v1.0"
+
+    def __init__(self, cfg):
+        super().__init__(cfg)
+
+    def load_classnames(self):
+        return {0: "healthy tissue", 1: "tumor tissue"}

dassl/data/datasets/dg/wilds/fmow.py

@@ -0,0 +1,57 @@
+import os.path as osp
+
+from dassl.data.datasets import DATASET_REGISTRY
+
+from .wilds_base import WILDSBase
+
+CATEGORIES = [
+    "airport", "airport_hangar", "airport_terminal", "amusement_park",
+    "aquaculture", "archaeological_site", "barn", "border_checkpoint",
+    "burial_site", "car_dealership", "construction_site", "crop_field", "dam",
+    "debris_or_rubble", "educational_institution", "electric_substation",
+    "factory_or_powerplant", "fire_station", "flooded_road", "fountain",
+    "gas_station", "golf_course", "ground_transportation_station", "helipad",
+    "hospital", "impoverished_settlement", "interchange", "lake_or_pond",
+    "lighthouse", "military_facility", "multi-unit_residential",
+    "nuclear_powerplant", "office_building", "oil_or_gas_facility", "park",
+    "parking_lot_or_garage", "place_of_worship", "police_station", "port",
+    "prison", "race_track", "railway_bridge", "recreational_facility",
+    "road_bridge", "runway", "shipyard", "shopping_mall",
+    "single-unit_residential", "smokestack", "solar_farm", "space_facility",
+    "stadium", "storage_tank", "surface_mine", "swimming_pool", "toll_booth",
+    "tower", "tunnel_opening", "waste_disposal", "water_treatment_facility",
+    "wind_farm", "zoo"
+]
+
+
+@DATASET_REGISTRY.register()
+class FMoW(WILDSBase):
+    """Satellite imagery classification.
+
+    62 classes (building or land use categories).
+
+    Reference:
+        - Christie et al. "Functional Map of the World." CVPR 2018.
+        - Koh et al. "Wilds: A benchmark of in-the-wild distribution shifts." ICML 2021.
+    """
+
+    dataset_dir = "fmow_v1.1"
+
+    def __init__(self, cfg):
+        super().__init__(cfg)
+
+    def get_image_path(self, dataset, idx):
+        idx = dataset.full_idxs[idx]
+        image_name = f"rgb_img_{idx}.png"
+        image_path = osp.join(self.dataset_dir, "images", image_name)
+        return image_path
+
+    def get_domain(self, dataset, idx):
+        # number of regions: 5 or 6
+        # number of years: 16
+        region_id = int(dataset.metadata_array[idx][0])
+        year_id = int(dataset.metadata_array[idx][1])
+        return region_id*16 + year_id
+
+    def load_classnames(self):
+        return {i: cat for i, cat in enumerate(CATEGORIES)}