torment123 / chef

CHEF

Cross-domain few-shot learning by representation fusion

paper, blog

Download and prepare data

miniImagenet

Download the miniImagenet data from here. We use the data split proposed by Ravi & Larochelle. Create the folders images_train, images_val, images_test and place the respective files in them, as well as a folder images_trainval that must contain all images from images_train and images_val.

tieredImagenet

Download the tiereImagenet data from here. Extract it and create a folder trainval containing all images from the folders train and val.

Set up horizontal data splits

Run python3 make_miniImagenet_hsplit.py and python3 make_tieredImagenet_hsplit.py to set up the horizontal data splits for pretraining.

Cross-domain data

Follow the instructions in this repo to acquire and set up the cross-domain data sets.

Pre-training

Run python3 pretrain.py config/pretrain_{res10,res12,conv64}_{tier,mini}.json.

Testing

To test the pre-trained ResNet-10 on the four cross-domain data sets run python3 xdom_res10.py config/xdom_res10.json --dataset {isic,cropdisease,eurosat,chest}. To test the Imagenet-pre-trained ResNet-18 from PyTorch on the four cross-domain data sets run python3 xdom_res18.py config/xdom_res18.json --dataset {isic,cropdisease,eurosat,chest}. Without the --dataset option, the model will be run on the miniImagenet test set by default.

To test the pre-trained ResNet-12 and Conv-4 on the miniImagenet and tieredImagenet test sets run python3 test.py config/{res12,conv64}_{mini,tier}_{1,5}shot.json.