Structure-Aware Feature Fusion for Unsupervised Domain Adaptation
Authors: Qingchao Chen, Yang Liu10567-10574
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments and Results We evaluate the proposed STAFF network with state-of-the-art deep learning based unsupervised domain adaptation methods. In this section, we first illustrate the datasets and implementation details. Then we show extensive experimental results and analysis. Our STAFF works reasonably well on all benchmarks, including Digit, Office-31 and Office-Home dataset. |
| Researcher Affiliation | Academia | Department of Engineering Science, University of Oxford, UK qingchao.chen@eng.ox.ac.uk, yangl@robots.ox.ac.uk |
| Pseudocode | No | The paper describes its methodology in text and through network architecture diagrams but does not provide structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any explicit statements about releasing source code for the described methodology, nor does it include links to a code repository. |
| Open Datasets | Yes | Digitis: We investigate three digits datasets of varying difficulties, including MNIST, USPS and the SVHN. We adopt the train-test protocol of (Russo et al. 2017) for a fair comparison with four transfer tasks: MNIST USPS (M U), USPS MNIST (U M), SVHN MNIST, (S M) and MNIST SVHN (M S). Office-31 is the most widely used dataset for unsupervised domain adaptation. Office-Home is a more difficult dataset than Office-31. |
| Dataset Splits | No | The paper states it adopts a 'train-test protocol' from a cited work but does not provide specific details on training, validation, or test dataset splits (percentages, counts, or explicit validation set usage) within the paper itself. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments, only mentioning the base network architectures used. |
| Software Dependencies | No | The paper mentions optimization details like SGD but does not specify any software libraries (e.g., PyTorch, TensorFlow) or their version numbers, which are necessary for reproducible ancillary software details. |
| Experiment Setup | Yes | We fix α = 1, β = 0.01, γ = 0.01 for all experiments. We train our network from scratch use SGD with momentum of 0.9, learning rate of 0.002 and batch size of 128. We used the following parameters α = 1, β = 0.1, γ = 0.05 for all experiments. The SGD with 0.9 momentum is used and the learning rate is annealed by up = u0(1 + ηp) φ, where p is the training progress changing from 0 to 1, and u0 = 0.01, η = 10, φ = 0.75 (Ganin et al. 2016). Whatever module trained from scratch, its learning rate was set to be 10 times that of the fine-tuning layers. |