no message

Author: KaihuaTang

classification/README.md

@@ -31,7 +31,7 @@ python main.py --cfg ./config/CIFAR100_LT/causal_norm.yaml
 
 If you want to change any hyper-parameter, you can find them in the corresponding yaml config file. **IMPORTANT: if you just want to change the TDE trade-off parameter alpha, you don't need to re-train the model, you can directly use different alphas during testing, because it's not involved in training. I also have a useful trick to pick up alpha: when you are testing on a dataset without additional val set (or directly testing on val set), you can choose the alpha that makes alpha times cos approximate to 1.0 in average.**.
 
-For Long-Tailed CIFAR-10/-100, if you want to change the imbalance ratio, you can set "cifar_imb_ratio" in the corresponding yaml file, e.g., cifar_imb_ratio=0.01/0.02/0.1 means imbalance ratio = 100, 50, 10 in the paper. To compare our methods with [BBN](https://github.com/Megvii-Nanjing/BBN) using Long-tailed CIFAR-10/-100, we copy their [ResNet32](https://github.com/Megvii-Nanjing/BBN/tree/master/lib/backbone) as your backbone.
+For Long-Tailed CIFAR-10/-100, if you want to change the imbalance ratio, you can set "cifar_imb_ratio" in the corresponding yaml file, e.g., cifar_imb_ratio=0.01/0.02/0.1 means imbalance ratio = 100, 50, 10 in the paper. To compare our methods with [BBN](https://github.com/Megvii-Nanjing/BBN) using Long-tailed CIFAR-10/-100, we adopted their dataloader, [ResNet32](https://github.com/Megvii-Nanjing/BBN/tree/master/lib/backbone) backbone, the number of training epochs and the same warm-up scheduler.
 
 ### Testing
 For ImageNet_LT: 
@@ -63,6 +63,8 @@ python main.py --cfg ./config/CIFAR100_LT/causal_norm.yaml --test --model_dir ./
 
 ![alt text](imagenet-lt.png "from 'Long-Tailed Classification by Keeping the Good and Removing the Bad Momentum Causal Effect'")
 
+As to the experiment results on Long-Tailed CIFAR-10/-100, we would like to thank @ZhangMingliangAI for reminding us that our previous setting is slightly different with BBN. This experiment was added during rebuttal, so we didn't have enough time to check all the BBN's code at that time. After I thoroughly went through the BBN's project, I turned off the pretrained initialization and adopted the same warm-up scheduler and number of training epochs. Since Long-Tailed CIFAR-10/-100 would sample different subsets at each time, we run each experiment twice and reported their mean results in the following table. Although the results are changed a little bit, we still outperform previous state-of-the-arts in the BBN's setting.
+
 ![alt text](long-tailed-cifar.png "from 'Long-Tailed Classification by Keeping the Good and Removing the Bad Momentum Causal Effect'")
 
 ## Citation

classification/config/CIFAR100_LT/causal_norm.yaml

@@ -16,7 +16,7 @@ networks:
     def_file: ./models/CausalNormClassifier.py
     optim_params: {lr: 0.2, momentum: 0.9, weight_decay: 0.0005}
     scheduler_params: {coslr: false, endlr: 0.0, gamma: 0.1, step_size: 30, warmup: true, lr_step: [120, 160], lr_factor: 0.01, warm_epoch: 5}
-    params: {dataset: CIFAR100_LT, feat_dim: 2048, num_classes: 100, stage1_weights: false, use_effect: true, num_head: 2, tau: 16.0, alpha: 2.0, gamma: 0.03125}
+    params: {dataset: CIFAR100_LT, feat_dim: 2048, num_classes: 100, stage1_weights: false, use_effect: true, num_head: 2, tau: 16.0, alpha: 1.5, gamma: 0.03125}
   feat_model:
     def_file: ./models/ResNext50Feature.py
     fix: false

classification/config/CIFAR100_LT/causal_norm_32.yaml

@@ -16,7 +16,7 @@ networks:
     def_file: ./models/CausalNormClassifier.py
     optim_params: {lr: 0.2, momentum: 0.9, weight_decay: 0.0005}
     scheduler_params: {coslr: false, endlr: 0.0, gamma: 0.1, step_size: 30, warmup: true, lr_step: [120, 160], lr_factor: 0.01, warm_epoch: 5}
-    params: {dataset: CIFAR100_LT, feat_dim: 128, num_classes: 100, stage1_weights: false, use_effect: true, num_head: 2, tau: 16.0, alpha: 2.0, gamma: 0.03125}
+    params: {dataset: CIFAR100_LT, feat_dim: 128, num_classes: 100, stage1_weights: false, use_effect: true, num_head: 2, tau: 16.0, alpha: 1.5, gamma: 0.03125}
   feat_model:
     def_file: ./models/ResNet32Feature.py
     fix: false