Visual Prompt Tuning in Null Space for Continual Learning

Authors: Yue Lu, Shizhou Zhang, De Cheng, Yinghui Xing, Nannan Wang, PENG WANG, Yanning Zhang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental results demonstrate the effectiveness of anti-forgetting on four classincremental benchmarks with diverse pre-trained baseline models, and our approach achieves superior performances to state-of-the-art methods.
Researcher Affiliation Academia 1 School of Computer Science, Northwestern Polytechnical University, China 2 School of Telecommunications Engineering, Xidian University, China
Pseudocode Yes An overview and algorithm of our approach are provided in Figure 6 and Algorithm 1 , respectively.
Open Source Code Yes Our code is available at https://github.com/zugexiaodui/VPTin NSfor CL .
Open Datasets Yes Our experiments are conducted across 4 classincremental benchmarks: 10and 20-split CIFAR-100, 10-split Image Net-R [39] and 10-split Domain Net [38].
Dataset Splits Yes 25% samples of the training data in each dataset are picked as a validation set for searching optimal hyper-parameters.
Hardware Specification Yes The experiments are performed on a server with 128 GB RAM and four NVIDIA RTX 4090 GPUs.
Software Dependencies No We implement our approach in Py Torch [24] with the timm library [41]. Specific version numbers for PyTorch and timm are not provided.
Experiment Setup Yes For the VPT-based models, we use the Adam optimizer [16] with β1 = 0.9, β2 = 0.999 and a weight decay of 5 10 5 to train 100 epochs with an initial learning rate of 0.01 and a batch size of 256 on all benchmarks. The learning rate is scaled by a factor of 0.1 at the 50-th and 80-th epoch.