Towards Identifiable Unsupervised Domain Translation: A Diversified Distribution Matching Approach
Authors: Sagar Shrestha, Xiao Fu
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments corroborate with our theoretical claims. NUMERICAL VALIDATION |
| Researcher Affiliation | Academia | Sagar Shrestha & Xiao Fu School of Electrical Engineering and Computer Science Oregon State University Corvallis, OR 97331, USA {shressag,xiao.fu}@oregonstate.edu |
| Pseudocode | No | The paper describes its methods and algorithms in text and mathematical formulas but does not include any clearly labeled 'Pseudocode' or 'Algorithm' blocks or figures. |
| Open Source Code | Yes | Source code is available at https://github.com/Xiao Fu Lab/Identifiable-UDT.git |
| Open Datasets | Yes | MNIST to Rotated MNIST (Mr M). We use 60, 000 training samples of the MNIST digits (Le Cun et al., 2010)... Edges to Rotated Shoes (Er S). Edges2Shoes dataset (Isola et al., 2017)... Celeb A-HQ to Bitmoji Faces (CB) We use 29, 900 training samples from Celeb A-HQ (Karras et al., 2017) as the x-domain, and 3, 984 training samples from Bitmoji faces (Mozafari, 2020) as the y-domain. |
| Dataset Splits | No | The paper mentions 'training samples' and 'testing samples' for both synthetic and real-world experiments. For example, 'In addition, we have 1,000 testing samples' for synthetic data, and for Bitmoji faces, '100 samples are held out as the test samples'. However, it does not explicitly mention a separate 'validation' split or its size/percentage for any of the datasets used. |
| Hardware Specification | Yes | For the translation tasks with 256 256 images (Celeb A-HQ to Bitmoji Faces), the runtime using a single Tesla V100 GPU is approximately 55 hours. For the translation tasks with 128 128 images (Edges to Rotated Shoes), the runtime using a single Tesla V100 GPU is approximately 35 hours. |
| Software Dependencies | No | The paper mentions using the Adam optimizer and setting hyperparameters (β1, β2), and refers to (Kingma & Ba, 2015) for Adam. However, it does not specify software dependencies like programming languages (e.g., Python), deep learning frameworks (e.g., PyTorch, TensorFlow), or other libraries with their specific version numbers. |
| Experiment Setup | Yes | Hyperparameter Settings. We use the Adam optimizer with an initial learning rate of 0.0001 with hyperparameters β1 = 0.5 and β2 = 0.999 (Kingma & Ba, 2015). We use a batch size of 1000 and train the models for 2000 iterations, where one iteration refers to one step of gradient descent of the translation and discriminator neural networks. We use λ = 10 for (7). ... For the real-data experiments: We use the Adam optimizer with an initial learning rate of 0.0001 with hyperparameters β1 = 0.0 and β2 = 0.999... We set our regularization parameter λ = 10. We use a batch size of 16. We train the networks for 100,000 iterations. |