To train our model on ImageNet, run following command:
Setting 1: Train our model without prototype loss:
python main_distributed.py --dataset imagenet --arch resnet50 --epochs 200 --data /path/to/imagenet
Setting 2: Train model with another augment_weight config:
python main_distributed.py --dataset imagenet --arch resnet50 --epoch 200 --use-class-temperature --data /data/ImageNet/ILSVRC/Data/CLS-LOC/ --batch-size 128 --augment-weight 0.25
Setting 3: Enable prototype loss with constant temperature :
python main_distributed.py --dataset imagenet --arch resnet50 --epoch 200 --use-center --data /data/ImageNet/ILSVRC/Data/CLS-LOC/ --batch-size 128 --augment-weight 0.5
Setting 4: Enable prototype loss with adaptive temperature :
python main_distributed.py --dataset imagenet --arch resnet50 --epochs 200 --data /path/to/imagenet --use-center --use-class-temperature
Setting 5 : Fine-tuning top classifier layer for pretrained model on ImageNet:
python train_linear_classifier.py --loss_type LDAM --train_rule Reweight --arch resnet50 --lr 0.1 --dataset imagenet --pretrained_model path/to/model/model_best.pth --data_path /data/ImageNet/ILSVRC/Data/CLS-LOC/
Setting 6 : Using TSC setting to fine-tuning ImageNet:
python TSC_fine_tuning.py --dataset imagenet --pretrained ../path_to/model_best.pth --epochs 40 --schedule 20 30 --seed 0 -b 2048 --data /data/ImageNet/ILSVRC/Data/CLS-LOC/
For testing accuracy of pretrained model with knn-evaluation:
Download the pretrained checkpoint here:
https://drive.google.com/u/0/uc?id=1XritMl3dYa9iW-TomaKU1XLQJVqgopMz&export=download
and modify path of pretrained model in metric_evaluate.py file, then run:
python metric_evaluate.py