progress model-to-debug

tutorials/tiny_image_net_resnet18_train.py

@@ -39,12 +39,14 @@
 
 regular_transforms = transforms.Compose(
         [
+            transforms.ToTensor(),
             transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
         ]
     )
 backdoor_transforms = transforms.Compose(
         [
             transforms.Resize((64, 64)),
+            transforms.ToTensor(),
             transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
         ]
     )
@@ -86,16 +88,16 @@ def get_all_descendants(target):
 
 
 class_to_group = {id_dict[k]: i for i, k in enumerate(id_dict) if k not in dogs.union(cats)}
-n_classes = len(class_to_group) + 2
+new_n_classes = len(class_to_group) + 2
 class_to_group.update({id_dict[k]: len(class_to_group) for k in dogs})
 class_to_group.update({id_dict[k]: len(class_to_group) for k in cats})
 
 
 # function to add a yellow square to an image in torchvision
 def add_yellow_square(img):
-    img[0, 10:13, 10:13] = 1
-    img[1, 10:13, 10:13] = 1
-    img[2, 10:13, 10:13] = 0
+    #img[0, 10:13, 10:13] = 1
+    #img[1, 10:13, 10:13] = 1
+    #img[2, 10:13, 10:13] = 0
     return img
 
 
@@ -107,26 +109,26 @@ def backdoored_dataset(dataset1, backdoor_samples, backdoor_label):
     return dataset1
 
 
-def flipped_group_dataset(train_set, n_classes, regular_transforms, seed, class_to_group, shortcut_fn, p_shortcut,
+def flipped_group_dataset(train_set, n_classes, new_n_classes, regular_transforms, seed, class_to_group, shortcut_fn, p_shortcut,
                           p_flipping, backdoor_dataset, backdoor_label):
     group_dataset = LabelGroupingDataset(
         dataset=train_set,
         n_classes=n_classes,
-        dataset_transform=regular_transforms,
+        dataset_transform=None,
         class_to_group=class_to_group,
         seed=seed,
     )
     flipped = LabelFlippingDataset(
         dataset=group_dataset,
-        n_classes=n_classes,
-        dataset_transform=regular_transforms,
+        n_classes=new_n_classes,
+        dataset_transform=None,
         p=p_flipping,
         seed=seed,
     )
 
     sc_dataset = SampleTransformationDataset(
         dataset=flipped,
-        n_classes=n_classes,
+        n_classes=new_n_classes,
         dataset_transform=regular_transforms,
         cls_idx=None,
         p=p_shortcut,
@@ -137,13 +139,13 @@ def flipped_group_dataset(train_set, n_classes, regular_transforms, seed, class_
     return backdoored_dataset(sc_dataset, backdoor_dataset, backdoor_label)
 
 
-train_set = TrainTinyImageNetDataset(local_path=local_path, transforms=regular_transforms)
+train_set = TrainTinyImageNetDataset(local_path=local_path, transforms=None)
 goldfish_dataset = SingleClassVisionDataset(path=goldfish_sketch_path, transforms=backdoor_transforms)
 # split goldfish dataset into train (100) and val (100)
 goldfish_set, _ = torch.utils.data.random_split(goldfish_dataset, [200, len(goldfish_dataset)-200], generator=rng)
 
 
-train_set = flipped_group_dataset(train_set, n_classes, regular_transforms, seed=42,
+train_set = flipped_group_dataset(train_set, n_classes, new_n_classes, regular_transforms, seed=42,
                                   class_to_group=class_to_group, shortcut_fn=add_yellow_square,
                                   p_shortcut=0.1, p_flipping=0.1, backdoor_dataset=goldfish_set,
                                   backdoor_label=1)

tutorials/tiny_imagenet_dataset.py

@@ -16,12 +16,14 @@
 from torch.optim import lr_scheduler
 from torchmetrics.functional import accuracy
 import glob
-from torchvision.io import read_image, ImageReadMode
+from PIL import Image
 
 import os
 import nltk
 from nltk.corpus import wordnet as wn
 
+from quanda.utils.datasets.transformed import LabelGroupingDataset, LabelFlippingDataset, SampleTransformationDataset
+
 # Download WordNet data if not already available
 nltk.download('wordnet')
 
@@ -46,9 +48,9 @@ def __len__(self):
 
     def __getitem__(self, idx):
         img_path = self.filenames[idx]
-        image = read_image(img_path)
-        if image.shape[0] == 1:
-          image = read_image(img_path,ImageReadMode.RGB)
+        image = Image.open(img_path).convert('RGB')
+        #if image.shape[0] == 1:
+          #image = read_image(img_path,ImageReadMode.RGB)
         label = self.id_dict[img_path.split('/')[-3]]
         if self.transforms:
             image = self.transforms(image.float())
@@ -73,9 +75,9 @@ def __len__(self):
 
     def __getitem__(self, idx):
         img_path = self.filenames[idx]
-        image = read_image(img_path)
-        if image.shape[0] == 1:
-          image = read_image(img_path,ImageReadMode.RGB)
+        image = Image.open(img_path).convert('RGB')
+        #if image.shape[0] == 1:
+          #image = read_image(img_path,ImageReadMode.RGB)
         label = self.cls_dic[img_path.split('/')[-1]]
         if self.transform:
             image = self.transform(image.float())
@@ -121,7 +123,30 @@ def get_all_descendants(target):
     id_dict = {line.strip(): i for i, line in enumerate(f)}
 
 
+name_dict = {}
+with open(local_path + '/wnids.txt', 'r') as f:
+    name_dict = {id_dict[line.strip()]: wn.synset_from_pos_and_offset('n', int(line.strip()[1:])) for i, line in enumerate(f)}
+
 class_to_group = {id_dict[k]: i for i, k in enumerate(id_dict) if k not in dogs.union(cats)}
 class_to_group.update({id_dict[k]: len(class_to_group) for k in dogs})
 class_to_group.update({id_dict[k]: len(class_to_group) for k in cats})
 
+# single-class vision dataset
+class SingleClassVisionDataset(torch.utils.data.Dataset):
+    def __init__(self, path:str, transforms:transforms.Compose, class_idx:int = 0):
+        self.filenames = glob.glob(path + "/train/*/*/*.JPEG")
+        self.transforms = transforms
+        self.class_idx = class_idx
+
+    def __len__(self):
+        return len(self.filenames)
+
+    def __getitem__(self, idx):
+        img_path = self.filenames[idx]
+        image = Image.open(img_path).convert('RGB')
+        #if image.shape[0] == 1:
+          #image = read_image(img_path,ImageReadMode.RGB)
+        label = self.class_idx
+        if self.transforms:
+            image = self.transforms(image.float())
+        return image, label