QT-DoG: Quantization-Aware Training for Domain Generalization

Authors: Saqib Javed, Hieu Le, Mathieu Salzmann

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our approach on diverse datasets from Domainbed and WILDS (Koh et al., 2021) Benchmark. All implementation, datasets, metric details and various ablation studies are provided in the Appendix.
Researcher Affiliation Academia 1CVLab, EPFL, Switzerland. 2Swiss Data Science Center, Switzerland.. Correspondence to: Mathieu Salzmann <EMAIL>.
Pseudocode No The paper includes mathematical equations (e.g., Eq. 1, 3, 5, 6) describing the quantization process and flatness calculation, but it does not contain a clearly labeled 'Pseudocode' or 'Algorithm' block.
Open Source Code Yes Code is released at: https: //saqibjaved1.github.io/QT_Do G/.
Open Datasets Yes We evaluate our approach on diverse datasets from Domainbed and WILDS (Koh et al., 2021) Benchmark... PACS (Li et al., 2017) is a 7 object classification challenge encompassing four domains... VLCS (Fang et al., 2013) poses a 5 object classification problem... Office Home (Venkateswara et al., 2017) comprises a total of 15,588 samples... Terra Incognita (Beery et al., 2018) addresses a 10 object classification challenge... Domain Net (Peng et al., 2019) provides a 345 object classification problem...
Dataset Splits Yes During this training phase, 20% of the samples are used for validation and model selection. We validate the model every 300 steps using held-out data from the source domains, and assess the final performance on the excluded domain (target). ... As (Cha et al., 2021), we split the in-domain datasets into training (60%), validation (20%), and test (20%) sets.
Hardware Specification Yes Every experiment in our work was executed on a single NVIDIA A100
Software Dependencies Yes Every experiment in our work was executed on a single NVIDIA A100, Python 3.8.16, Py Torch 1.10.0, Torchvision 0.11.0, and CUDA 12.1.
Experiment Setup Yes We use the same training procedure as Domain Bed (Gulrajani & Lopez-Paz, 2021), incorporating additional components from quantization. Specifically, we adopt the default hyperparameters from Domain Bed (Gulrajani & Lopez-Paz, 2021), including a batch size of 32 (per-domain). We employ a Res Net-50 (He et al., 2016) pre-trained on Image Net (Russakovsky et al., 2015) as initial model and use a learning rate of 5e-5 along with the Adam optimizer, and no weight decay. Following SWAD(Cha et al., 2021), the models are trained for 15,000 steps on Domain Net and 5,000 steps on the other datasets.