LINAS

LINAS is model-agnostic and can be adapted to various baseline methods. We provide our code based on 'Dual Encoding'.

Environments

• Ubuntu 16.04
• CUDA 11.1
• Python 3.8
• Pytorch 1.9.0
• Numpy 1.19.2
• Scipy 1.5.4
• Tensorboard-logger 0.1.0

Dataset

To run our code, please first download the required dataset and a pre-trained word2vec here. Then extract the content in LINAS/dataset/.

Training

Run the script for training.

cd LINAS
./train_all.sh $GPU_DEVICE $support-set-size

Evaluation

The evaluation on the test set will be performed automatically after the training. If you would like to evaluate before the end of training, please add '--test' in the script.

Adaptive Distillation

Change the '--similarity_type' in train_all.sh from 'diag' to 'adapt' for training the mask with our Adaptive Distillation strategy.