Learning to Transfer: Unsupervised Domain Translation via Meta-Learning
Authors: Jianxin Lin, Yijun Wang, Zhibo Chen, Tianyu He11507-11514
AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate effectiveness of our model on ten diverse two-domain translation tasks and multiple face identity translation tasks. We show that our proposed approach significantly outperforms the existing domain translation methods when each domain contains no more than ten training samples. |
| Researcher Affiliation | Collaboration | 1CAS Key Laboratory of Technology in Geo-spatial Information Processing and Application System, University of Science and Technology of China 2University of Science and Technology of China, 3DAMO Academy, Alibaba Group {linjx, wyjun}@mail.ustc.edu.cn, chenzhibo@ustc.edu.cn, timhe.hty@alibaba-inc.com |
| Pseudocode | Yes | Algorithm 1 MT-GAN training process |
| Open Source Code | No | The paper does not provide an explicit statement or link indicating that the source code for the described methodology is publicly available. |
| Open Datasets | Yes | The first one (denoted as P1(T)) contains 10 diverse translation tasks collected by (Zhu et al. 2017): labels photos, horses zebras, summer winter, apple orange, monet photo, cezanne photo, ukiyoe photo, vangogh photo, photos maps and labels facades. In addition, the Facescrub dataset (Ng and Winkler 2014), which comprises 531 different celebrities, is utilized as another collection of domain translation tasks (denoted as P2(T)). |
| Dataset Splits | Yes | For a translation task T, we take the other 9 different tasks as the training dataset and test the meta-learned parameter initialization on task T. This procedure could be seen as task-level 10 fold cross-validation. ... We mainly focus on 5-shot domain translation and 10-shot domain translation in all experiments. Moreover, we set the query set s size L to 10. |
| Hardware Specification | No | The paper mentions "to fit the memory limit of the GPU" but does not specify any particular GPU model (e.g., NVIDIA A100, RTX 2080 Ti), CPU, or other hardware details used for the experiments. |
| Software Dependencies | No | The paper mentions using "stochastic gradient descent (SGD)" and "Adam optimizer (Kingma and Ba 2014)" but does not specify versions for these or any other software libraries (e.g., Python, PyTorch, TensorFlow, CUDA versions). |
| Experiment Setup | Yes | For each meta-training period, we use stochastic gradient descent (SGD) with learning rate α = 0.0001 to fine-tune the generators and discriminators. At meta-optimization time, we use the Adam optimizer (Kingma and Ba 2014) with learning rate β = 0.0002 to update both meta-generator and metadiscriminator. ... The overall iteration number of meta-training T is set to 100. We set the loss function balance parameters λcyc and λidt to be 10 and 5. |