Latent Convolutional Models
Authors: ShahRukh Athar, Evgeny Burnaev, Victor Lempitsky
ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments are performed on Celeb A (Liu et al., 2015) (128x128 resolution), SUN Bedrooms (Yu et al., 2015) (256x256 resolution), Celeb A-HQ (Karras et al., 2018) (1024x1024 resolution) datasets, and we demonstrate that the latent models, once trained, can be applied to large hole inpainting, superresolution of very small images, and colorization tasks, outperforming other latent models in our comparisons. |
| Researcher Affiliation | Collaboration | Shah Rukh Athar Evgeny Burnaev Victor Lempitsky Skolkovo Institute of Science and Technology (Skoltech), Russia Currently at Stony Brook University. Currently also with Samsung AI Center, Moscow. |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code of our implementation is available at the project website. |
| Open Datasets | Yes | The Celeb A dataset was obtained by taking the 150K images from (Liu et al., 2015)... The Bedrooms dataset from the LSUN (Yu et al., 2015)... Finally, the Celeb A-HQ dataset from (Karras et al., 2018)... |
| Dataset Splits | No | The paper mentions comparing on 'hold-out sets not used for training' and refers to a 'validation set' in the context of a baseline model (GAN) but does not provide explicit training, validation, and test splits (e.g., percentages or counts) for its own experiments. |
| Hardware Specification | No | The paper mentions '14 GPU-days' for training time but does not specify any particular GPU models, CPU models, or other hardware components used for the experiments. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies such as libraries or frameworks used in the implementation. |
| Experiment Setup | Yes | Following (Bojanowski et al., 2018), we use plain SGD with very high learning rate of 1.0 to train LCM and of 10.0 to train the GLO models. The exact architectures are given in Appendix D. ... The Generator Network gθ: The generator network gθ has an hourglass architecture in all three datasets. In Celeb A the map size varies as follows: 32 32 4 4 128 128 and the generator has a total of 38M parameters. ... Latent Network fφi: The latent network used in Celeb A128 consists of 4 convolutional layers with no padding. The latent network used in Bedrooms and Celeb A-HQ consists of 5 and 7 convolutional layers respectively with no padding. |