flowers102 one class

common/common.py

@@ -7,7 +7,8 @@ def parse_args(default=False):
     parser = ArgumentParser(description='Pytorch implementation of CSI')
 
     parser.add_argument('--dataset', help='Dataset',
-                        choices=['cifar10', 'cifar100', 'imagenet', 'flowers102', 'fashionmnist', 'mnist'], type=str)
+                        choices=['cifar10', 'cifar100', 'imagenet', 'flowers102', 'fashionmnist',
+                                 'mnist', 'oxford102flower'], type=str)
     parser.add_argument('--one_class_idx', help='None: multi-class, Not None: one-class',
                         default=None, type=int)
     parser.add_argument('--model', help='Model',

datasets/datasets.py

@@ -136,7 +136,7 @@ def get_transform_imagenet():
 
 
 def get_dataset(P, dataset, test_only=False, image_size=None, download=True, eval=False):
-    if dataset in ['imagenet', 'cub', 'stanford_dogs', 'flowers102',
+    if dataset in ['imagenet', 'cub', 'stanford_dogs', 'flowers102', 'oxford102flower', 
                    'places365', 'food_101', 'caltech_256', 'dtd', 'pets']:
         if eval:
             train_transform, test_transform = get_simclr_eval_transform_imagenet(P.ood_samples,
@@ -168,6 +168,13 @@ def get_dataset(P, dataset, test_only=False, image_size=None, download=True, eva
         train_set = datasets.MNIST(DATA_PATH, train=True, download=download, transform=train_transform)
         test_set = datasets.MNIST(DATA_PATH, train=False, download=download, transform=test_transform)
 
+    elif dataset == 'oxford102flower':
+        image_size = (224, 224, 3)
+        n_classes = 102
+        base_dir = os.path.join(DATA_PATH, 'oxford102flower')
+        train_set = datasets.ImageFolder(base_dir + "/train", transform=train_transform)
+        test_set = datasets.ImageFolder(base_dir + "/valid", transform=test_transform)
+
     elif dataset == 'cifar100':
         image_size = (32, 32, 3)
         n_classes = 100
@@ -275,6 +282,8 @@ def get_superclass_list(dataset):
         return list(range(10))
     elif dataset == 'mnist':
         return list(range(10))
+    elif dataset == 'oxford102flower':
+        return list(range(1, 103))
     else:
         raise NotImplementedError()
 

training/unsup/simclr_CSI.py

@@ -3,7 +3,7 @@
 import torch.optim
 
 import models.transform_layers as TL
-from training.contrastive_loss import get_similarity_matrix, NT_xent
+from training.contrastive_loss import NT_xent, get_similarity_matrix
 from utils.utils import AverageMeter, normalize
 
 device = torch.device(f"cuda" if torch.cuda.is_available() else "cpu")
@@ -39,13 +39,14 @@ def train(P, epoch, model, criterion, optimizer, scheduler, loader, logger=None,
         check = time.time()
 
         ### SimCLR loss ###
-        if P.dataset != 'imagenet':
+        if P.dataset == 'imagenet' or P.dataset == 'oxford102flower':
+            batch_size = images[0].size(0)
+            images1, images2 = images[0].to(device), images[1].to(device)
+        else:
             batch_size = images.size(0)
             images = images.to(device)
             images1, images2 = hflip(images.repeat(2, 1, 1, 1)).chunk(2)  # hflip
-        else:
-            batch_size = images[0].size(0)
-            images1, images2 = images[0].to(device), images[1].to(device)
+
         labels = labels.to(device)
 
         images1 = torch.cat([P.shift_trans(images1, k) for k in range(P.K_shift)])
@@ -110,4 +111,3 @@ def train(P, epoch, model, criterion, optimizer, scheduler, loader, logger=None,
         logger.scalar_summary('train/loss_sim', losses['sim'].average, epoch)
         logger.scalar_summary('train/loss_shift', losses['shift'].average, epoch)
         logger.scalar_summary('train/batch_time', batch_time.average, epoch)
-