add cifar-N to docta

Author: zwzhu-d

.gitignore

@@ -3,3 +3,8 @@ dist
 docta.ai.egg-info
 *.pyc
 results/Tabular_train/label_error_Tabular_train_diagnose_report.pt
+data/cifar/cifar-10-python.tar.gz
+data/cifar/cifar-10-batches-py/*
+data/cifar/cifar-100-python.tar.gz
+data/cifar/cifar-100-python/*
+test_cifar_N.py

config/cifar100.py

@@ -0,0 +1,13 @@
+# dataset settings
+seed = 0
+dataset_type = 'CIFAR'
+modality = 'image' # image, text, tabular
+num_classes = 100
+data_root = './data/cifar/'
+label_sel = 1 # which label/attribute we want to diagnose
+train_label_sel = label_sel # 1 for noisy
+test_label_sel = train_label_sel
+
+file_name = 'c100'
+dataset_type += '_' + file_name
+save_path = f'./results/{dataset_type}/'

data/cifar/CIFAR-100_human.pt

@@ -1,10 +1,10 @@
 # different dataset builder
-from .cifar import Cifar10_noisy, Cifar100_noisy, Cifar10_clean, Cifar100_clean
+from .cifar import Cifar10_noisy, Cifar100_noisy, Cifar10_clean, Cifar100_clean, Cifar10N, Cifar100N
 from .hh_rlhf import HH_RLHF
 from .customize import CustomizedDataset
 from .customize_img_folder import Customize_Image_Folder
 from .csv_loder import TabularDataset
 
 __all__ = [
-    'Cifar10_noisy', 'Cifar100_noisy', 'HH_RLHF', 'CustomizedDataset', 'Customize_Image_Folder', 'Cifar10_clean', 'Cifar100_clean', 'TabularDataset'
+    'Cifar10_noisy', 'Cifar100_noisy', 'HH_RLHF', 'CustomizedDataset', 'Customize_Image_Folder', 'Cifar10_clean', 'Cifar100_clean', 'TabularDataset', 'Cifar10N', 'Cifar100N'
 ]

docta/datasets/__init__.py

@@ -68,6 +68,31 @@ def __getitem__(self, index: int):
         return img, label, index
 
 
+class Cifar10N(Cifar10_noisy):
+    
+    train_transform = transforms.Compose([
+        transforms.RandomCrop(32, padding=4),
+        transforms.RandomHorizontalFlip(),
+        transforms.ToTensor(),
+        transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
+    ])
+
+    test_transform = transforms.Compose([
+        transforms.ToTensor(),
+        transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
+    ])
+
+    def __init__(self, cfg, train=True, preprocess=None, noisy_label_key=None, clean_label_key=None) -> None:
+        print(f"CIFAR-10N includes five sets of noisy labels: random_label1, random_label2, random_label3, aggre_label, worse_label.\nPlease set cfg.noisy_label_key to one of them.")
+        if preprocess is None:
+            preprocess = self.train_transform if train else self.test_transform
+        self.cfg = cfg
+        self.cfg.label_path = "./data/cifar/CIFAR-10_human.pt"
+        if noisy_label_key is not None:
+            self.cfg.noisy_label_key = noisy_label_key
+        if clean_label_key is not None:
+            self.cfg.clean_label_key = clean_label_key
+        super(Cifar10N, self).__init__(cfg, train, preprocess)
 
 class Cifar10_clean(CIFAR10):
 
@@ -159,4 +184,16 @@ class Cifar100_clean(Cifar10_clean):
     test_transform = transforms.Compose([
         transforms.ToTensor(),
         transforms.Normalize((0.5071, 0.4867, 0.4408), (0.2675, 0.2565, 0.2761)),
-    ])
+    ])
+
+
+class Cifar100N(Cifar100_noisy):
+
+    def __init__(self, cfg, train=True, preprocess=None) -> None:
+        if preprocess is None:
+            preprocess = self.train_transform if train else self.test_transform
+        self.cfg = cfg
+        self.cfg.label_path = "./data/cifar/CIFAR-100_human.pt"
+        self.cfg.noisy_label_key = "noisy_label"
+        self.cfg.clean_label_key = "clean_label"
+        super(Cifar100N, self).__init__(cfg, train, preprocess)