Gradual Domain Adaptation via Gradient Flow

Authors: Zhan Zhuang, Yu Zhang, Ying Wei

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on several benchmark datasets demonstrate the superiority of the proposed GGF method compared to state-of-the-art baselines. We conduct comprehensive experiments to evaluate the proposed GGF on various domain adaptation scenarios, where GGF outperforms state-of-the-art methods.
Researcher Affiliation Academia Zhan Zhuang1,2, Yu Zhang1, , Ying Wei3, 1Southern University of Science and Technology, 2City University of Hong Kong 3Nanyang Technological University
Pseudocode Yes Algorithm 1: Construct Next Intermediate Domain (xt, yt, sπ,ϕ, ht, vθ, α, η) Algorithm 2: Complete Algorithm with Fixed Hyperparameters Algorithm 3: Complete Algorithm with Bilevel Optimization
Open Source Code Yes Our code is publicly available at https://github.com/zwebzone/ggf.
Open Datasets Yes Portraits is a gender (binary) classification dataset with 37,921 facing portraits from 1905 to 2013. We follow the chronological split from Kumar et al. (2020)... Rotated MNIST is a variant of the MNIST dataset (Le Cun, 1998)... Office-Home (Venkateswara et al., 2017) is a well-known UDA dataset... Vis DA-2017 (Peng et al., 2017) is a large-scale dataset...
Dataset Splits Yes We follow the chronological split from Kumar et al. (2020), creating a source domain (first 2000 images), intermediate domains (14000 images not used here), and a target domain (next 2000 images).
Hardware Specification Yes We conduct experiments on a single NVIDIA 2080Ti GPU. ...on a single NVIDIA V100 GPU...
Software Dependencies No The paper mentions several software tools and libraries used, such as UMAP, Rectified Flow, and the Sinkhorn solver from the POT toolbox, and refers to official implementations. However, it does not provide specific version numbers for any of these software components (e.g., 'UMAP 0.5.1' or 'PyTorch 1.10'), which is necessary for reproducible software dependencies.
Experiment Setup Yes We apply SGD optimizer with a learning rate of 10^-4 for training all modules and updating the classifier. The batch size for each domain is set to 1024. ...fine-tuning with five epochs. We provide the hyperparameters in Table 9, where λ is the balancing weight between the cross-entropy and entropy in the class-based energy function, and α, T, and η are used in the sampling process.