Merge-Friendly Post-Training Quantization for Multi-Target Domain Adaptation
Authors: Juncheol Shin, Minsang Seok, Seonggon Kim, Eunhyeok Park
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | To validate the effectiveness of our proposed HDRQ method, we conduct experiments on multi-target domain adaptation tasks for semantic segmentation and image classification. The harmonic mean of performances across target domains is reported as the primary evaluation metric. Semantic Segmentation For semantic segmentation, we use the GTA synthetic dataset (Richter et al., 2016) as the source domain and two real-world datasets, Cityscapes (Cordts et al., 2016) and Indian Driving datasets (Varma et al., 2019), as the target domains. We adopt HRDA (Hoyer et al., 2022) as the single-target domain adaptation method, using Res Net-101 (He et al., 2016) as the backbone model and simple convolution head. All other settings are kept consistent with the original paper. Image Classification For image classification, we use the Office-Home dataset (Venkateswara et al., 2017), which consists of four domains (Real, Art, Clipart, and Product). We report the harmonic mean of accuracies across target domains, with the best results highlighted in bold, and red-colored results indicating accuracy gains of over > 1% compared to the second-best result. |
| Researcher Affiliation | Academia | 1Graduate School of Artificial Intelligence, Pohang University of Science and Technology (POSTECH), Pohang, Republic of Korea 2Department of Computer Science and Engineering, Pohang University of Science and Technology (POSTECH), Pohang, Republic of Korea. Correspondence to: Eunhyeok Park <EMAIL>. |
| Pseudocode | No | The paper describes methods using mathematical formulas and descriptive text, but does not include any clearly labeled 'Pseudocode' or 'Algorithm' blocks, nor does it present structured steps formatted like code. |
| Open Source Code | No | The paper does not provide any explicit statements about the release of source code for the methodology described, nor does it include links to any code repositories or mention code in supplementary materials. |
| Open Datasets | Yes | Semantic Segmentation For semantic segmentation, we use the GTA synthetic dataset (Richter et al., 2016) as the source domain and two real-world datasets, Cityscapes (Cordts et al., 2016) and Indian Driving datasets (Varma et al., 2019), as the target domains. Image Classification For image classification, we use the Office-Home dataset (Venkateswara et al., 2017), which consists of four domains (Real, Art, Clipart, and Product). |
| Dataset Splits | No | The paper states that the multi-target domain adaptation approach (Li et al., 2024), HRDA (Hoyer et al., 2022), and SHOT (Liang et al., 2020) were adopted, with adaptation settings following original works. However, it does not explicitly provide specific percentages, sample counts, or detailed methodologies for how the datasets were split into training, validation, or test sets within this paper. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware used for running the experiments, such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper mentions using the "Adam optimizer" but does not specify a version number for it or any other software libraries or frameworks. Therefore, it does not provide specific ancillary software details with version numbers. |
| Experiment Setup | Yes | The hyperparameter λ for weight distance regularization is set to 5e-2 for all tasks. HDRQ conducts 20,000 iterations of block-wise reconstruction, including activation quantization with partial dropout. We use the Adam optimizer with an initial learning rate of 0.001 and a cosine annealing with warmup. |