Accepted at ICCV 2021

1. ILSVRC2012: Register at ImageNet and request for a username and an access key to download ILSRVC-2012 data set.

2. CUB-200-2011 Birds: Birds

3. FGVC-Aircraft: Aircraft

4. FC100: FC100

5. Omniglot: Omniglot

6. Texture: Texture

7. Traffic Sign: Traffic Sign

8. FGCVx Fungi: Fungi Annotations

9. Quick Draw: Quick Draw

• Use gsutil to download the data:

gsutil -m cp gs://quickdraw_dataset/full/numpy_bitmap/*.npy data/quickdraw

• Convert numpy files to jpeg images using this conversion code

10. VGG Flower: VGG Flower, Labels

python extract_pretrained_features.py <path_to_data> -f <result_folder> -b <batch_size> --gpu <gpu_ID>

• Single library classifier example:

python classifier_single.py data/aircraft --model resnet18 --nway 5 --kshot 1 --kquery 15 --num_epochs 200 --n_problems 600 --hidden_size 512 --lr 0.001 --gamma 0.2

• Full library classifier example:

python classifier_full_library.py data/aircraft --nway 20 --kshot 5 --kquery 15 --num_epochs 100 --n_problems 600 --hidden_size 512 --lr 0.0005 --gamma 0.1

• data_path: path to the folder containing all images: data/<dataset>
• --model: model name for single classifier (Resnet18, Densenet121, etc.)
• --gpu: GPU ID to be used
• Hyperparameters
  • --lr: learning rate for the classifier: 0.001
  • --kshot: number of images from each class in training set: 1
  • --kquery: number of images from each class in test set: 15
  • --nway: number of classes per task: 5
  • --hidden_size: hidden state size for the classifier: 1024
  • --num_epochs: number of training epochs: 100
  • --n_problems: number of tasks used for testing: 600
  • --gamma: L2 regularization constant: 0.5
  • --linear: Use for a linear NN architecture (no hidden layer)
  • --nol2: Use to get rid of L2 regularization
  • --soft: Use for soft bagging when applying ensemble method, otherwise hard bagging