Not Just Object, But State: Compositional Incremental Learning without Forgetting

Authors: Yanyi Zhang, Binglin Qiu, Qi Jia, Yu Liu, Ran He

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on two datasets exhibit state-of-the-art performance achieved by Comp ILer.
Researcher Affiliation Academia Yanyi Zhang 1, Binglin Qiu 1, Qi Jia 1, Yu Liu 1, Ran He 2 1 International School of Information Science & Engineering, Dalian University of Technology 2 MAIS&CRIPAC, Institute of Automation, Chinese Academy of Sciences
Pseudocode Yes Algorithm 1: Training Procedure of Comp ILer for composition-IL
Open Source Code Yes Code and datasets are available at: https://github.com/Yanyi-Zhang/Comp ILer.
Open Datasets Yes Code and datasets are available at: https://github.com/Yanyi-Zhang/Comp ILer. As there are no existing datasets suitable for composition-IL, we re-organize the data in Clothing16K [46] and UT-Zappos50K [44], and construct two new datasets tailored for composition-IL, namely Split-Clothing and Split-UT-Zappos.
Dataset Splits No We allocate 80% of the dataset for training and the remaining 20% for testing. Consistent with Split-Clothing, we employ 80% of the images for training and 20% for testing. No explicit mention of a separate validation split percentage is provided.
Hardware Specification No The paper does not specify the hardware used for experiments, such as specific GPU or CPU models, memory, or cloud instance types.
Software Dependencies No The paper mentions using the Adam optimizer but does not specify any software dependencies (e.g., libraries, frameworks) with version numbers.
Experiment Setup Yes For multi-pool prompt learning, the size of each pool is set to 20, and each prompt has 5 tokens. We select top-5 prompts from each pool and generate a fused prompt. During training, we utilize the Adam optimizer [9] with a batch size of 16. The whole Comp ILer undergoes training for 25 epochs on the Split Clothing, for 10 epochs on the 5-task Split-UT-Zappos, and for 3 epochs on the 10-task Split-UT-Zappos. For the Split-Clothing and the 10-task Split-UT-Zappos, we set the learning rate to 0.03, while we use a learning rate of 0.02 for the 5-task Split-UT-Zappos. Note that, for all the methods, their results are averaged over three runs with the corresponding standard deviations reported to mitigate the influence of random factors. As there are a few hyper-parameters in the model, we conduct a rigorous tuning on them. For instance, we set θthre to π/2 for all settings. For Split-Clothing, the loss weights λ1 and λ3 are set to 0.1; λ2 is set to 10^−7; α and β for SCE loss are 0.006 and 0.3, and the parameter µ during inference is 0.5. For 5-task Split-UT-Zappos, λ1, λ2, λ3, α, β and µ are set to 1.0, 3 × 10^−6, 0.7, 0.01, 0.7 and 0.02, respectively. For 10-task Split-UT-Zappos, λ1, λ2, λ3, α, β and µ are set to 0.5, 10^−7, 0.1, 0.05, 0.4 and 0.03.