- python 3.7.7
- pytorch 1.6.0
- torchvision 0.8.1
Train ResNet-101 on ImageNet
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python imagenet_DDP.py PATH_TO_DATASET --arch resnet --net resnet101 --batch-size 1024 --lr 0.4 --epochs 90 --workers 32 --dist-url 'tcp://127.0.0.1:12345' --dist-backend 'nccl' --multiprocessing-distributed --world-size 1 --rank 0 --ixx_r 5 --ixy_r 0.75
Train ResNet-152 on ImageNet
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python imagenet_DDP.py PATH_TO_DATASET --arch resnet --net resnet152 --batch-size 1024 --lr 0.4 --epochs 90 --workers 32 --dist-url 'tcp://127.0.0.1:12345' --dist-backend 'nccl' --multiprocessing-distributed --world-size 1 --rank 0 --ixx_r 5 --ixy_r 1
Train ResNeXt-101, 32×8d on ImageNet
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python imagenet_DDP.py PATH_TO_DATASET --arch resnet --net resnext101_32x8d --batch-size 1024 --lr 0.4 --epochs 90 --workers 32 --dist-url 'tcp://127.0.0.1:12345' --dist-backend 'nccl' --multiprocessing-distributed --world-size 1 --rank 0 --ixx_r 5 --ixy_r 0.75
Evaluate pre-trained model on ImageNet
CUDA_VISIBLE_DEVICES=0,1 python imagenet_DDP.py PATH_TO_DATASET --net MODEL_NAME --pre-train PATH_TO_PRE_TRAINED_MODELS -e --batch-size 512 --lr 0.4 --epochs 90 --workers 32 --dist-url 'tcp://127.0.0.1:12345' --dist-backend 'nccl' --multiprocessing-distributed --world-size 1 --rank 0
- Measured by Top-1 error.
| Model | Top-1 error | Top-5 error | Link |
|---|---|---|---|
| ResNet-101 (InfoPro*, K=2) | 21.85 | 5.89 | Tsinghua Cloud / Google Drive |
| ResNet-152 (InfoPro*, K=2) | 21.45 | 5.84 | Tsinghua Cloud / Google Drive |
| ResNeXt101, 32x8d (InfoPro*, K = 2) | 20.35 | 5.28 | Tsinghua Cloud / Google Drive |
