Skip to content

README.md

Latest commit

 

History

History
39 lines (33 loc) · 2.05 KB

README.md

File metadata and controls

39 lines (33 loc) · 2.05 KB

Interpolation-based LoRA (I-LORA)

[ English | 中文 ]

Main Idea

I-LORA is a continual learning method that improves upon LoRA by constructing a dual-memory experience replay framework based on LoRA parameter interpolations.

Method-Specific Parameters

  • use_ilora (bool): Whether to enable I-LORA method. Default: False
  • ema_alpha (float): EMA smoothing coefficient for updating stable adapter. Default: 0.999
  • consistency_weight (float): Weight for the consistency loss between plastic and stable adapters. Default: 1.0
  • ilora_buffer_size (int): Maximum number of samples to store in memory buffer. Default: 500
  • selective_update (bool): Whether to selectively update only when plastic model performs better. Default: False
  • min_update_threshold (float): Minimum threshold for selective updates. Default: 0.1
  • hidden_state_layers (list): List of hidden state layers to compute consistency loss on. Default: [-1]
  • save_ema_adapter (bool): Whether to save the EMA adapter separately. Default: True
  • ema_adapter_path (str): Path to save/load the EMA adapter. Default: None
  • previous_task_model (str): Path to the model from previous task. Default: None
  • current_task_id (str): ID of the current task. Default: None
  • prev_task_id (str): ID of the previous task. Default: None

Files in Directory

  • ilora.py: Core implementation of I-LORA including buffer management and consistency loss computation
  • ilora_adapter.py: Adapter initialization and management for I-LORA
  • ilora_trainer.py: Custom trainer with I-LORA specific training logic
  • ilora_workflow.py: Training workflow management for I-LORA
  • ilora_loader.py: Model loading utilities for I-LORA
  • README.md: English documentation
  • README_zh.md: Chinese documentation

Citation

@article{ren2024analyzing, title={Analyzing and reducing catastrophic forgetting in parameter efficient tuning}, author={Ren, Weijieying and Li, Xinlong and Wang, Lei and Zhao, Tianxiang and Qin, Wei}, journal={arXiv preprint arXiv:2402.18865}, year={2024} }