Knowledge Amalgamation for Multi-Label Classification (KA-MLC)

This is the implementation repository of Adaptive kNowledge Transfer (ANT), the solution for KA-MLC, AAAI 2023.

File listing

• main.py : Code for training ANT
• model.py : Supporting models
• utils.py : Supporting utility functions
• requirements.txt : Library requirements

Instructions

Prepared folders:

• data : directory to place training data
• teachers : directory to place teacher models
• output : directory for training logs and outputs

The datasets and the teachers we use in our paper are available here: https://drive.google.com/drive/folders/1QlFDmR4D5fA6MQMnkqy1-fG-6leQI-Od?usp=sharing

Run script as:

python main.py -t_numlabel 4 4 -t_labels 1 2 3 4 3 4 5 6 -stu_labels 1 2 3 4 5 6 \
-t_path './teachers/t0.sav' './teachers/t1.sav' -dataname 'sample_data'   

Parameters:

• Required:

  • -t_path : path of each teacher model, e.g., './t0.sav' './t1.sav'
  • -t_numlabel : #labels corresponding to each teacher in t_path, e.g., 4 4'
  • -t_labels : concatenated specialized labels of each teacher corresponding to t_path, e.g., t0_label: 1 2 3 4 and t1_label: 3 4 5 6, then t_labels: 1 2 3 4 3 4 5 6
  • -stu_labels : student labels, e.g., 1 2 3 4 5 6
  • -dataname : unlabelled data for training the student

• Hyperparameters:

  • -lr : learning rate, default 0.001
  • -ep : epochs, default 500
  • -bs : batch size, default 8
  • -layer : #layers, default 1
  • -hidden : #hidden size, default 32