PACE: Marrying generalization in PArameter-efficient fine-tuning with Consistency rEgularization

Authors: Yao Ni, Shan Zhang, Piotr Koniusz

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental evidence supports our theories. PACE surpasses existing PEFT methods in visual adaptation tasks (VTAB-1k, FGVC, few-shot learning, domain adaptation) showcasing its potential for resource-efficient finetuning. It also improves Lo RA in text classification (GLUE) and mathematical reasoning (GSM-8K).
Researcher Affiliation Collaboration Yao Ni Shan Zhang , Piotr Koniusz , , The Australian National University Data61 CSIRO Australian Institute for Machine Learning, The University of Adelaide yao.ni@anu.edu.au shan.zhang@adelaide.edu.au piotr.koniusz@data61.csiro.au
Pseudocode No The paper describes the pipeline and steps involved in PACE, particularly in Section 3.5 and Figure 2, but it does not present a clearly labeled pseudocode block or algorithm box.
Open Source Code Yes The code is available at github.com/Maxwell Yao Ni/PACE.
Open Datasets Yes Datasets and evluations. VTAB-1K comprises 19 datasets organized into (i) Natural images, (ii) Specialized datasets (remote sensing, medical) and (iii) Structured datasets (scene structure) domains. Each dataset has 1K training examples... Few-shot learning involves 5 fine-grained datasets: FGVC-Aircraft [46], Food101 [4], Oxford Flowers102 [51], Oxford Pets [53] and Stanford Cars [35]... FGVC includes 5 fine-grained datasets: CUB200-2011 [69], NABirds [66], Oxford Flowers [51], Stanford Dogs [10] and Stanford Cars [35]... For domain adaptation, following [82, 7], we train on Image Net [11]... We evaluate on GLUE [70] for text classification and GSM-8K [9] for mathematical reasoning.
Dataset Splits Yes VTAB-1K... Following [78, 28], we use the provided 800-200 train split for hyperparameter selection, evaluate using the full training set and report average accuracy across three trails... Few-shot learning... tune hyperparameters using validation and report average test accuracy over three random seeds... For domain adaptation... use the validation split by [82] for hyperparameter selection and report the results on the official validation set and 4 out-of-domain datasets...
Hardware Specification Yes All experiments used an NVIDIA H100 GPU.
Software Dependencies No The paper mentions using "Adam optimizer [32]" but does not specify its version number or versions for other key software components like programming languages (e.g., Python) or deep learning frameworks (e.g., PyTorch, TensorFlow, CUDA).
Experiment Setup Yes Models are fine-tuned for 300 epochs on VTAB-1K and 100 epochs on other vision adaptation tasks, with batch size 64. For text classification we follow [26]... We use the Adam optimizer [32] with cosine learning rate decay and linear warm-up (first 10 epochs)... Building on the strong Lo RAmul+VPTadd, we use the grid search for our λ and σ, following strategies from previous studies [28, 41, 26]. ... Table 16, 17, 18 and 19 present the hyperparameters and number of trainable parameters used in our strong baseline for VTAB-1K, few-shot learning, FGVC and domain adaptation tasks.