Anticipatory Recovery in Sequential Cyclic Fine-tuning

Code for the paper "Reawakening knowledge: Anticipatory recovery from catastrophic interference via structured training" (NeurIPS 2024).

The llm-experiments folder includes the Language model (Pythia) experiments; the igpt-experiments folder includes the Image GPT experiments; the imagenet-experiments folder includes the image classification experiments.

Installation

To install requirements:

pip install -r requirements.txt

LLM Experiments

Example commands for cyclic fine-tuning experiments can be found in the llm-experiments/scripts folder.

Code for visualizing the pairwise recovery matrix (Figures 8b and 18), PCA in the last layer weights (Figure 9), and representations (Figure 8d) can be found in the llm-experiments/visualization folder.

Image GPT Experiments

Example command for cyclic fine-tuning with Image GPT:

python train_interleave_igpt.py \
        --learning_rate 0.001 \
        --model_size medium \
        --output_dir ./medium-20steps \
        --save_prefix batch1_gpu1 \
        --num_train_epochs 5 \
        --num-grad-steps 20 \
        --num-data-samples 25

where num-grad-steps is the number of consecutive gradient update steps on each image, and num-data-samples is the number of images in the sequence.

Acknowledgements

The code is adapted from Huggingface Transformers and Emin Orhan's LLM Memory Experiments.