RecursiveMix (RM), which uses the historical input-prediction-label triplet to enhance the generalization of Deep Vision Models. Paper Link Here.
Experiment Environment
- python 3.6
- pytorch 1.7.1+cu101
- torchvision 0.8.2
- mmcv-full 1.4.1
- mmdet 2.19.1
- mmsegmentation 0.20.2
For example, to reproduce the results of RM in CIFAR-10 (97.65% Top-1 acc in averaged 3 runs, logs are provided in logs/):
CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --master_port 29500 main.py \
--name 'your_experiment_log_path' \
--model_file 'pyramidnet' \
--model_name 'pyramidnet_200_240' \
--data 'cifar10' \
--data_dir '/path/to/CIFAR10' \
--epoch 300 \
--batch_size 64 \
--lr 0.25 \
--scheduler 'step' \
--schedule 150 225 \
--weight_decay 1e-4 \
--nesterov \
--num_workers 8 \
--save_model \
--aug 'recursive_mix' \
--aug_alpha 0.5 \
--aug_omega 0.1
RM in ImageNet (79.20% Top-1 acc)
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -m torch.distributed.launch --nproc_per_node=8 --master_port 29500 main.py \
--name 'your_experiment_log_path' \
--model_file 'resnet' \
--model_name 'resnet50' \
--data 'imagenet' \
--epoch 300 \
--batch_size 512 \
--lr 0.2 \
--warmup 5 \
--weight_decay 1e-4 \
--aug_plus \
--num_workers 32 \
--save_model \
--aug 'recursive_mix' \
--aug_alpha 0.5 \
--aug_omega 0.5
CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --master_port 29500 main.py \
--name 'your_experiment_log_path' \
--batch_size 64 \
--model_file 'pyramidnet' \
--model_name 'pyramidnet_200_240' \
--data 'cifar10' \
--data_dir '/path/to/CIFAR10' \
--num_workers 8 \
--evaluate \
--resume 'best'
- ImageNet-1K (300 epoch)
Backbone | Size | Params (M) | Acc@1 | Log | Download |
---|---|---|---|---|---|
ResNet-50 | 224 | 25.56 | 76.32 | log | [Google] [GitHub] |
+ Mixup | 224 | 25.56 | 77.42 | log | [Google] [GitHub] |
+ CutMix | 224 | 25.56 | 78.60 | log | [Google] [GitHub] |
+ RecursiveMix | 224 | 25.56 | 79.20 | log | [Google] [GitHub] |
- COCO (1x schedule)
Backbone | Lr schd | Mem (GB) | Inf time (fps) | box AP | Log | Download |
---|---|---|---|---|---|---|
ResNet-50 | 1x | 3.7 | 19.7 | 39.4 | log | [Google] [GitHub] |
+ CutMix | 1x | 3.7 | 19.7 | 40.1 | log | [Google] [GitHub] |
+ RecursiveMix | 1x | 3.7 | 19.7 | 41.5 | log | [Google] [GitHub] |
- ADE20K (80k iteration)
Backbone | Crop Size | Lr schd | Mem (GB) | Inf time (fps) | mIoU | Log | download |
---|---|---|---|---|---|---|---|
ResNet-50 | 512x512 | 80000 | 8.1 | 23.40 | 40.40 | log | [Google] [GitHub] |
+ CutMix | 512x512 | 80000 | 8.1 | 23.40 | 41.24 | log | [Google] [GitHub] |
+ RecursiveMix | 512x512 | 80000 | 8.1 | 23.40 | 42.30 | log | [Google] [GitHub] |