Conditional Generative Modeling via Learning the Latent Space
Authors: Sameera Ramasinghe, Kanchana Nisal Ranasinghe, Salman Khan, Nick Barnes, Stephen Gould
ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Compared to existing generative solutions, our approach demonstrates faster and more stable convergence, and can learn better representations for downstream tasks. Importantly, it provides a simple generic model that can perform better than highly engineered pipelines tailored using domain expertise on a variety of tasks, while generating diverse outputs. Code available at https://github.com/samgregoost/c GML. 5 EXPERIMENTS AND DISCUSSIONS |
| Researcher Affiliation | Academia | Australian National University, Data61 (CSIRO), Mohamed bin Zayed University of AI sameera.ramasinghe@anu.edu.au |
| Pseudocode | Yes | Eq. 1 can be optimized via Algorithm 1 (proof in App. 2.2). Algorithm 1: Training algorithm |
| Open Source Code | Yes | Code available at https://github.com/samgregoost/c GML. |
| Open Datasets | Yes | The distribution of natural images lies on a high dimensional manifold, making the task of modelling it extremely challenging. Moreover, conditional image generation poses an additional challenge with their constrained multimodal output space (a single input may correspond to multiple outputs while not all of them are available for training). In this section, we experiment on several such tasks. MNIST dataset, Celeb-HQ, Facade (Tyleˇcek & Šára, 2013), CIFAR10 (Krizhevsky et al., 2009), Model Net10 & 40. |
| Dataset Splits | No | While the paper mentions train/test splits, it does not explicitly state the details of any validation splits for its experiments. For example, in Appendix 3.1, it mentions 'the data was split at 80/20 for training/testing', without detailing a validation set. |
| Hardware Specification | No | Appendix 3.1 states 'Our models were trained for up to 200 epochs on a single GPU.' This mention of 'a single GPU' lacks the specificity (e.g., model name, memory) required for hardware reproduction. |
| Software Dependencies | No | The paper does not provide specific version numbers for any software, libraries, or frameworks used in the experiments. Although it mentions deep learning models and provides a code link, it lacks details like 'PyTorch 1.x' or 'Python 3.x'. |
| Experiment Setup | Yes | For hyper-parameters and training details, see App. 3.1. Appendix 3.1: We used Adam optimizer with β1 = 0.9, β2 = 0.999. We used a constant learning rate of 0.0001 (for H and G) and 0.001 (for Z). The latent variable z was randomly initialized for each training batch, and its dimension was set to 10. For ˆE we used an L1 loss, and for the diversity of color, we used KL divergence as described in App. 3.3. |