Likelihood Training of Schrödinger Bridge using Forward-Backward SDEs Theory
Authors: Tianrong Chen, Guan-Horng Liu, Evangelos Theodorou
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | we show that the resulting training algorithm achieves comparable results on generating realistic images on MNIST, Celeb A, and CIFAR10. Our code is available at https://github.com/ghliu/SB-FBSDE. |
| Researcher Affiliation | Academia | Tianrong Chen , Guan-Horng Liu , Evangelos A. Theodorou Georgia Institute of Technology, USA {tianrong.chen, ghliu, evangelos.theodorou}@gatech.edu |
| Pseudocode | Yes | Algorithm 1 Likelihood training of SB-FBSDE, Algorithm 2 Joint (diffusion flow-based) training, Algorithm 3 Alternate (IPF-based) training, Algorithm 4 Generative Process of SB-FBSDE |
| Open Source Code | Yes | Our code is available at https://github.com/ghliu/SB-FBSDE. |
| Open Datasets | Yes | We testify SB-FBSDE on two toy datasets and three image datasets, i.e. MNIST, Celeb A, and CIFAR10. |
| Dataset Splits | No | No explicit statements on training, validation, or test dataset splits (e.g., percentages or counts) are provided for MNIST, Celeb A, or CIFAR10. While these are standard datasets with common splits, the paper does not specify them. |
| Hardware Specification | Yes | All networks adopt position encoding and are trained with Adam W (Loshchilov & Hutter, 2017) on a TITAN RTX. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, or TensorFlow versions) that are required to replicate the experiment. |
| Experiment Setup | Yes | We testify SB-FBSDE on two toy datasets and three image datasets, i.e. MNIST, Celeb A, and CIFAR10. pprior is set to a zero-mean Gaussian whose variance varies for each task and can be computed according to Song & Ermon (2020). We parameterize Z( , ; θ) and b Z( , ; φ) with residual-based networks for toy datasets and consider Unet (Ronneberger et al., 2015) and NCSN++ (Song et al., 2020) respectively for MNIST/Celeb A and CIFAR10. All networks adopt position encoding and are trained with Adam W (Loshchilov & Hutter, 2017) on a TITAN RTX. We adopt VE-SDE (i.e. f := 0; see Song et al. (2020)) as our SDE backbone... On all datasets, we set the horizon T =1.0 and solve the SDEs via the Euler-Maruyama method. The interval [0, T] is discretized into 200 steps for CIFAR10 and 100 steps for all other datasets... Other details are left in Appendix D. |