BaMix:BALANCED MIXUP LOSS FOR LONG-TAILED VISUAL RECOGNITION

Dataset

• Imbalanced CIFAR. The original data will be downloaded and converted by imbalancec_cifar.py.
• ImageNet-LT. We followed CAM to generate long-tailed ImageNet.

Main requirements

Python==3.7
pytorch==1.7.1
torchaudio==0.10.0
torchvision==0.11.1
tensorboardX==2.4
matplotlib==3.4.3
scikit-learn==1.0.1
opencv-python==4.5.3.56
seaborn==0.11.2
numpy==1.21.4
pandas==1.1.5
pillow==8.4.0
six==1.16.0
tqdm==4.62.3

file structure

BaMix
  ├── data
  │     ├── ImageNet_LT
  │     │     ├── ImageNet_LT_train.json
  │     │     ├── ImageNet_LT_val.json
  ├── lib
  ├── main
dataset
  ├── imbalance_cifar
  │     ├── cifar-10-batches-py
  │     ├── cifar-100-python
  ├── ImageNet    
  │     ├── train
  │     │     ├── n01440764
  │     │     │       ├── n01440764_18.JPEG
  │     │     │       ├── n01440764_36.JPEG
  │     │     │       ├── ......
  │     │     ├── n01443537
  │     │     │       ├── ......
  │     │     ├── ......
  │     ├── val
  │     │     ├── ILSVRC2012_val_00000001.JPEG
  │     │     ├── ILSVRC2012_val_00000002.JPEG
  │     │     ├── ......

Training

We provide several training examples with this repo:

• To train cifar10-LT with imbalance ratio of 100

python cifar_train.py --dataset cifar10 -im 0.01 --device 0

• To train cifar100-LT with imbalance ratio of 200

python cifar_train.py --dataset cifar100 -im 0.005 --device 0

• To train ImageNet-LT with architecture of resnet10 and batch size of 256.

python ImgNet_train.py -a resnet10i -b 256  --devices 0  --mixepoch 160

• To train ImageNet-LT with architecture of resnet50 and batch size of 128 on two gpus.

python ImgNet_train.py -a resnet50i -b 128  --devices 0,1 --mixepoch 160

-To train iNaturalist 2018

python iNaturalist_train.py -a resnet50i -b 128 --devices 0

Testing

Make sure that the name of the folder conforms to our naming convention, which can be automatically generated through training.

• To test cifar-10/100-LT

python test.py --resume ../checkpoint/cifar10_resnet32_STM_DRW_0_0.01_alpha_1.0_maxm_0.5_mixepoch_360/best.pth.tar