The implementaion of paper Multi-directional Knowledge Transfer for Few-Shot Learning, which has been accepted by ACM Multimedia 2022.

Requirements

• Python3
• PyTorch1.8, torchvision

Running the code

Our extracted features from ImNet dataset are available at onedrive. Download the extracted features and place them into an empty directory Features/.

The benchmark of ImNet dataset with 5 different settings for the number of novel category examples K = {1,2,3,4,5}. The main entry point for running the low shot benchmark islow_shot.py, which will run a single experiment for a single value of K. (Refer to Low-shot Visual Recognition by Shrinking and Hallucinating Features)

For example, running the first experiment with K=1 will look like:

python ./low_shot.py --nshot 1 \
  --lr 0.001 --wd 0.001 \
  --batchsize 1000

Performance

Method	K=1	K=2	K=3	K=4	K=5
baseline	36.1	47.9	54.0	58.1	60.8
MDKT	44.4	53.3	58.1	61.7	63.8

*We use a visual classifier composed of a full connection layer as the baseline model.