Self-ensembling for visual domain adaptation
Authors: Geoff French, Michal Mackiewicz, Mark Fisher
ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our approach achieves state of the art results in a variety of benchmarks, including our winning entry in the VISDA2017 visual domain adaptation challenge. In small image benchmarks, our algorithm not only outperforms prior art, but can also achieve accuracy that is close to that of a classifier trained in a supervised fashion. |
| Researcher Affiliation | Academia | Geoff French, Michal Mackiewicz & Mark Fisher School of Computing Sciences University of East Anglia Norwich UK {g.french,m.mackiewicz,m.fisher}@uea.ac.uk |
| Pseudocode | No | The paper describes network architectures in detail with tables (Tables 6-8) but does not provide pseudocode or algorithmic blocks for the overall self-ensembling or domain adaptation algorithm. |
| Open Source Code | Yes | Our implementation was developed using Py Torch (Chintala et al.) and is publically available at http://github.com/Britefury/self-ensemble-visual-domain-adapt. |
| Open Datasets | Yes | The datasets used in this paper are described in Table 3. ... USPSa Available from http://statweb.stanford.edu/~tibs/ElemStatLearn/datasets/zip.train.gz and http://statweb.stanford.edu/~tibs/ElemStatLearn/datasets/zip.test.gz ... STLb Available from http://ai.stanford.edu/~acoates/stl10/ ... Syn-Digitsc Available from Ganin’s website at http://yaroslav.ganin.net/ |
| Dataset Splits | Yes | The Vis DA-2017 image classification challenge is a 12-class domain adaptation problem consisting of three datasets: a training set consisting of 3D renderings of sketchup models, and validation and test sets consisting of real images (see Figure 1) drawn from the COCO Lin et al. (2014) and You Tube Bounding Boxes Real et al. (2017) datasets respectively. |
| Hardware Specification | Yes | batch size of 56 (reduced from 64 to fit within the memory of an n Vidia 1080-Ti) |
| Software Dependencies | No | Our implementation was developed using Py Torch (Chintala et al.). While PyTorch is mentioned, no specific version number is provided to ensure full reproducibility of software dependencies. |
| Experiment Setup | Yes | Our networks were trained for 300 epochs. We used the Adam Kingma & Ba (2015) gradient descent algorithm with a learning rate of 0.001. We trained using mini-batches composed of 256 samples... The self-ensembling loss was weighted by a factor of 3 and the class balancing loss was weighted by 0.005. Our teacher network weights ti were updated... with a value of 0.99 for α. |