CLAP4CLIP: Continual Learning with Probabilistic Finetuning for Vision-Language Models

Authors: Saurav Jha, Dong Gong, Lina Yao

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our method on CIFAR100 [3, 30], Image Net100 [41, 43], Image Net-R [59], CUB200 [60], and VTAB [60].
Researcher Affiliation Collaboration Saurav Jha1, Dong Gong1 , Lina Yao1,2 1University of New South Wales (UNSW Sydney), 2CSIRO s Data61 {saurav.jha, dong.gong}@unsw.edu.au; lina.yao@data61.csiro.au
Pseudocode Yes Algorithm 1: A forward CLAP4CLIP pass at test step t
Open Source Code Yes Our code is available at https://github.com/srv Codes/clap4clip.
Open Datasets Yes We evaluate our method on CIFAR100 [3, 30], Image Net100 [41, 43], Image Net-R [59], CUB200 [60], and VTAB [60].
Dataset Splits Yes we tuned our hyperparameters using a validation set comprising 10% of the CIFAR-100 training dataset.
Hardware Specification Yes All our experiments were performed on NVIDIA V100 GPUs hosted on the Gadi supercomputers of the National Computational Infrastructure (NCI Australia).
Software Dependencies No The paper mentions using SGD and refers to models like CLIP, but does not provide specific version numbers for software dependencies such as Python, PyTorch, or other libraries.
Experiment Setup Yes We train CLAP and its variants using SGD, with a batch size of 64, for 5 epochs, including 1 epoch of linear warmup. The initial learning rate (LR) is set to 1e-3 and decays with cosine annealing. At the end of each incremental task (t > 1), we perform memory consolidation training for 2 epochs, with an LR of 1e-4, on the class-balanced memory dataset.