This is an unofficial pytorch implementation of Low-Shot Learning with Imprinted Weights.

• Python 3.5+
• PyTorch 0.4.1
• torchvision
• pandas (0.23.4)
• progress
• matplotlib
• numpy

Paper: InceptionV1 + RMSProp

This implementation: ResNet-50 + SGD

Download CUB_200_2011 Dataset.

Unzip and locate it in this directory.

The whole directory should be look like this:

imprinted-weights
│   README.md
│   pretrain.py
│   models.py
│   loader.py
│   imprint.py
│   imprint_ft.py
│   alljoint.py
|
└───utils
│   
└───CUB_200_2011
    │   images.txt
    │   image_class_labels.txt
    │   train_test_split.txt
    │
    └───images
        │   001.Black_footed_Albatross
        │   002.Laysan_Albatross
        │   ...

Train the model on the first 100 classes of CUB_200_2011.

python pretrain.py

Trained models will be saved at pretrain_checkpoint.

Use N novel exemplar from the training split to imprint weights.

python imprint.py --model pretrain_checkpoint/model_best.pth.tar --num-sample N

Trained models will be saved at imprint_checkpoint.

For more details and parameters, please refer to --help option.

All w/o FT results of Table 1 and Table 2 in the paper can be reproduced by this script.

Apply fine-tuning to the imprinting model.

python imprint_ft.py --model pretrain_checkpoint/model_best.pth.tar --num-sample N

Trained models will be saved at imprint_ft_checkpoint.

All w/ FT results of Table 1 and Table 2 in the paper can be reproduced by this script.

Train model for base and novel classes jointly without a separate low-shot learning phase.

python alljoint.py --num-sample N

Trained models will be saved at alljoint_checkpoint.

• [1]: H. Qi, M. Brown and D. Lowe. "Low-Shot Learning with Imprinted Weights", in CVPR, 2018.