README.md

Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models (ZSCL)

This is the official implementation of paper "Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models". Our approach Zero-Shot Continual Learning (ZSCL) aims to mitigate forgetting problem existed in the continual learning of large pretrained vision-language models. This repo includes experiments for Multi-domain Task Increamental Learning (MTIL) in mtil and Class Incremental Learning in cil.

Multi-domain Task Increamental Learning

Go into the mtil directory.

Dataset Preparation

• Required Datasets: ImageNet, Conceptual_Captions
• Target Datasets: Aircraft, Caltech101,CIFAR10, CIFAR100, DTD, EuroSAT, Flowers, Food, MNIST, OxfordPet,StanfordCars, SUN397

Download required datasets (ImageNet and Conceptual_Captions) and datasets to be learned (e.g., DTD, MNIST, ect.) to your experiment directory. You can refer to datasets.md for more details.

Replicate our results

If you want to replicate our results in our paper, you can directly run the command line for that table. For example, if you want to replicate the results of ZSCL, then run bash scripts/ZSCL.sh.

More commands are listed in scripts directory.

To calculate Transfer, Avg., and Last scores, you can use results.ipynb. The results in the paper are also listed in results.ipynb.

Training and Evaluation

Following cammand lines are examples of training and evaluating the model.

# train from clip model
python -m src.main \
    --train-mode=whole \
    --train-dataset=DTD \
    --lr=1e-5\
    --ls 0.2 \
    --iterations 1000 \
    --method ZSCL \
    --image_loss \
    --text_loss \
    --we \
    --avg_freq 100 \
    --l2 1 \
    --ref-dataset ImageNet \
    --ref-sentences conceptual_captions \
    --save ckpt/exp_000

# evaluation on all datasets
python -m src.main --eval-only \
    --train-mode=whole \
    --eval-datasets=Aircraft,Caltech101,CIFAR100,DTD,EuroSAT,Flowers,Food,MNIST,OxfordPet,StanfordCars,SUN397 \
    --load ckpt/exp_000/Flowers.pth

# continual training
python -m src.main \
    --train-mode=whole \
    --train-dataset=MNIST \
    --lr=5e-5 \
    --ls 0.2 \
    --method ZSCL \
    --image_loss \
    --text_loss \
    --we \
    --avg_freq 100 \
    --l2 1 \
    --ref-dataset ImageNet \
    --ref-sentences conceptual_captions \
    --iterations 1000 \
    --save ckpt/exp_000 \
    --load ckpt/exp_000/DTD.pth

Hyperparameters

The meaning of hyperparameters in the command line is as follows:

params	name
--batch-size	batch size
--iterations	iterations during training
--lr	learning rate
--method	approaches for continual learning
--train-model	components of the model to be trained
--ls	label smoothing
--image_loss	initialize loss for image encoder
--text_loss	initialize loss for text encoder
--we	initialize weight ensemble
--avg_freq	frequency to conduct weight ensemble
--l2	initialize weight constraint
--ref-dataset	dataset for image encoder distillation
--ref-sentences	texts for text encoder distillation
--dataset_order	the order of datasets for iCaRL

Class Incremental Learning

Dataest Preparation

• Required Datasets: ImageNet, Conceptual_Captions
• Target Datasets: ImageNet, tinyImageNet, CIFAR100

Download required datasets (ImageNet and Conceptual_Captions) and datasets to be learned (e.g., tinyImageNet, CIFAR100 ect.) to your experiment directory.

Training

First, change your configs in .yaml files in configs/class

Then, use below cammand lines to train and evaluate the model. if you want to replicate our results in our paper, you can directly run the command line for that table and run the command line below.

python main.py \
    --config-path configs/class \
    --config-name imagenet100_10-10-ZSCL.yaml \
    dataset_root="../datasets/" \
    class_order="class_orders/imagenet100.yaml"

Hyperparameters

The meaning of hyperparameters in the command line is as follows:

params	name
--batch-size	batch size
--initial_increment	initial increment
--lr	learning rate
--dataset	the dataset to be learned
--method	appraoch for Continual Learning
--ls	label smoothing
--weight_decay	weight decay
--text_loss	initialize loss for text encoder
--we	initialize weight ensemble
--avg_freq	frequency to conduct weight ensemble
--l2	initialize weight constraint
--ref-dataset	dataset for image encoder distillation
--ref-sentences	texts for text encoder distillation
--ce_method	an optional method to add external texts

Citation

@article{zheng2022preventing,
  title={Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models},
  author={Zheng, Zangwei and Ma, Mingyuan and Wang, Kai and Qin, Ziheng and Yue, Xiangyu and You, Yang},
  journal={arXiv preprint arXiv:2303.06628},
  year={2023}
}

Acknowledgement

mil is built on wise-ft, and cil is built on Continual-CLIP. We thank the authors for sharing their codes.