VirDA: Reusing Backbone for Unsupervised Domain Adaptation with Visual Reprogramming

Authors: Duc-Duy Nguyen, Dat Nguyen

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate Vir DA on Office-31 and obtain 92.8% mean accuracy with only 1.5M trainable parameters. We conduct experiments to evaluate Vir DA s capability in classification effectiveness, training parameter size, and storage requirement for each source and target domain pair. Our experiments demonstrate that the proposed Vir DA, requiring only a maximum of 1.5 million of training parameters (less than 2% of PMTrans (Zhu et al., 2023)) and only 6 MB for storage per domain (compared to over 340 MB of both PMTrans and CDTrans (Xu et al., 2022)), and fully reusing the backbone s parameters, achieves comparable performance to state-of-the-art (SOTA) methods across standard domain adaptation benchmarks, including Office-31 (Saenko et al., 2010), Office-Home (Venkateswara et al., 2017), and Digits (MNIST (Le Cun et al., 1998), USPS (Netzer et al., 2011), SVHN (Hull, 1994)).
Researcher Affiliation Academia Duy Nguyen Hanoi University of Science and Technology EMAIL Dat Nguyen Harvard University EMAIL Basis.ai EMAIL
Pseudocode No The paper describes the methodology using mathematical equations and textual descriptions (e.g., Section 3, Equations 1-14), but does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes We release our implementation and reproduction package at https://github.com/Duy-Nguyen-Duc/Vir DA
Open Datasets Yes We evaluate our proposed method on widely used three domain adaptation benchmarks, namely Office-31, Office-Home, and Digits... Digits is a dataset composed from three other digit datasets, which are MNIST (Le Cun et al., 1998), USPS (Hull, 1994), and Street View House Numbers (SVHN) (Hull, 1994). Office-31 (Saenko et al., 2010) is the most popular dataset for real-world domain adaptation. Office-Home (Venkateswara et al., 2017) is a more challenging benchmark than Office-31.
Dataset Splits Yes For the UDA test, we adopted three commonly used cross-dataset transfer settings with the standard data split: S M, U M, M U. We evaluated all methods on six domain adaptation tasks.
Hardware Specification Yes The experiment results reported here were obtained on a machine equipped with Intel Xeon Gold 6130 CPU at 2.10GHz clock speed with 16 cores and 64 GB of RAM running Linux and using a single NVIDIA GTX 3090 device.
Software Dependencies No The paper mentions using Adam W optimizer with specific hyperparameters but does not provide specific version numbers for software dependencies like programming languages, deep learning frameworks, or libraries.
Experiment Setup Yes In all experiments, we use both Resnet (He et al., 2016) and Vi T (Dosovitskiy et al., 2020) models pre-trained on Image Net (Deng et al., 2009) as the fixed backbone for Vir DA. For the Digits tasks, we use Res Net-18 with a learning rate of 3e 4 for the classifier heads and 5e 4 for the visual reprogramming modules, using a batch size of 128. The dropout rate for the classifier and the mask generator is set as pmask = 0.5 and p C = 0.3, respectively. On the Office-Home and Office-31 datasets, we adopt Vi T-B/32 as the backbone for all transfer tasks. We set the same learning rate as above, using a batch size of 32 and pmask = 0.3 with p C = 0.1. For all experiments, we adopt Adam W (Loshchilov & Hutter, 2017) with the default configuration of (β1, β2) is (0.9, 0.999), and a weight decay of 1e 5. On the Office-31 and Office-Home datasets, we set Lvr and Nvr corresponding to 6 and 5, for coarse object-level mask, while on Digits we set Lvr = 5 and Nvr = 4, as the dataset s characteristics demonstrate mild transformation. On all tasks, we set the number of forward passes to estimate uncertainty M = 4. We train our method in two phases, namely burn-in in 20 epochs and domain adaptation in 30 epochs for all data settings following Li et al. (2022). The scaling factor for each loss is set as αunc = 0.3, αadv = 0.1 and αintra = 0.15, in which these hyperparameters are obtained via grid-search.