update

MediaBrain-SJTU · May 30, 2022 · 6a943c7 · 6a943c7
1 parent 32edab5
commit 6a943c7
Show file tree

Hide file tree

Showing 2 changed files with 20 additions and 5 deletions.
diff --git a/README.md b/README.md
@@ -104,17 +104,17 @@ python train.py SimCLR --lr 0.5 --epochs 2000 --temperature 0.2 --weight_decay 5
 
 - BCL-I
 ```train BCL-I
-python train.py BCL_I --bcl --lr 0.5 --epochs 2000 --temperature 0.2 --weight_decay 5e-4 --data_folder ${data_folder} --trainSplit cifar100_imbSub_with_subsets/cifar100_split1_D_i.npy 
+python train.py BCL_I --bcl --rand_k 1 --lr 0.5 --epochs 2000 --temperature 0.2 --weight_decay 5e-4 --data_folder ${data_folder} --trainSplit cifar100_imbSub_with_subsets/cifar100_split1_D_i.npy 
 ```
 
 - SDCLR
-```train SimCLR
+```train SDCLR
 python train_sdclr.py SDCLR --lr 0.5 --epochs 2000 --temperature 0.2 --weight_decay 1e-4 --data_folder ${data_folder} --trainSplit cifar100_imbSub_with_subsets/cifar100_split1_D_i.npy 
 ```
 
 - BCL-D
 ```train BCL-D
-python train_sdclr.py BCL_D --bcl --lr 0.5 --epochs 2000 --temperature 0.2 --weight_decay 1e-4 --data_folder ${data_folder} --trainSplit cifar100_imbSub_with_subsets/cifar100_split1_D_i.npy 
+python train_sdclr.py BCL_D --bcl --rand_k 2 --lr 0.5 --epochs 2000 --temperature 0.2 --weight_decay 1e-4 --data_folder ${data_folder} --trainSplit cifar100_imbSub_with_subsets/cifar100_split1_D_i.npy 
 ```
 
 Pretrained checkpoints will be saved in 'checkpoints/'.
@@ -135,6 +135,19 @@ python test.py --checkpoint ${checkpoint_pretrain} --prune --test_fullshot --tes
 
 The code will output the results of full-shot/100-shot/50-shot linear probing evaluation.
 
+### Results and Pretrained Models
+
+We provide the full-shot/100-shot/50-shot results(demo) pretrained on 'cifar100_split1_D_i.npy' with the corresponding checkpoint weights.
+
+  | Method                 | Full-shot  | 100-shot |  50-shot    | Model  |
+  | ---------------------- |:-----:|:------:|:------: |:------:|
+  | SimCLR                 | 50.7  |  46.3  |  42.4	  | [ResNet18](https://drive.google.com/file/d/1gcVzeLDY1fCR1uUjwlrYyatByb3AP74W/view?usp=sharing)|
+  | SDCLR                  | 55.0  |  49.7	|  45.6   | [ResNet18](https://drive.google.com/file/d/1MHH8K8mmCCPL6GOwR9g0TtjSJPPouRBZ/view?usp=sharing)|
+  | BCL-I                  |   |  	|     | [ResNet18]
+  | BCL-D                  |   |    |  	  | [ResNet18]
+
+After downloading the checkpoints, you could run evaluation by the instructions in the evaluating section.
+
 ### Extensions
 
 **Steps to Implement Your Own Model**

diff --git a/data/memoboosted_cifar100.py b/data/memoboosted_cifar100.py
@@ -34,14 +34,16 @@ def __getitem__(self, idx):
         img = self.data[idx]
         img = Image.fromarray(img).convert('RGB')
 
+        min_strength = 10 # training stability
         memo_boosted_aug = transforms.Compose([
                 transforms.RandomResizedCrop(32, scale=(0.1, 1.0), interpolation=3),
                 transforms.RandomHorizontalFlip(),
                 transforms.RandomApply([transforms.ColorJitter(0.4, 0.4, 0.4, 0.1)], p=0.8),
                 transforms.RandomGrayscale(p=0.2),
+                # RandAugment_prob(1, 23, 0.5),
                 # RandAugment_prob(1, 30*self.momentum_weight[idx]*np.random.rand(1), 1.0*self.momentum_weight[idx]),
-                RandAugment_prob(1, 5+25*self.momentum_weight[idx]*np.random.rand(1), 1.0*self.momentum_weight[idx]),
-                # RandAugment_prob(self.args.rand_k, self.args.rand_strength*self.momentum_weight[idx]*np.random.rand(1), 1.0*self.momentum_weight[idx]),
+                # RandAugment_prob(1, 10+20*self.momentum_weight[idx]*np.random.rand(1), 1.0*self.momentum_weight[idx]),
+                RandAugment_prob(self.args.rand_k, min_strength + (self.args.rand_strength - min_strength)*self.momentum_weight[idx]*np.random.rand(1), 1.0*self.momentum_weight[idx]),
                 transforms.ToTensor(),
             ])