Variational Schrödinger Diffusion Models
Authors: Wei Deng, Weijian Luo, Yixin Tan, Marin Biloš, Yu Chen, Yuriy Nevmyvaka, Ricky T. Q. Chen
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, we test the algorithm in simulated examples and observe that VSDM is efficient in generations of anisotropic shapes and yields straighter sample trajectories compared to the single-variate diffusion. We also verify the scalability of the algorithm in real-world data and achieve competitive unconditional generation performance in CIFAR10 and conditional generation in time series modeling. |
| Researcher Affiliation | Collaboration | 1Machine Learning Research, Morgan Stanley, NY 2Peking University 3Duke University 4Meta AI (FAIR). |
| Pseudocode | Yes | Algorithm 1 Variational Schr odinger Diffusion Models (VSDM). |
| Open Source Code | Yes | See code in https://github.com/pkulwj1994/diff_instruct |
| Open Datasets | Yes | We choose the CIFAR10 dataset as representative image data to demonstrate the scalability of the proposed VSDM on generative modeling of high-dimensional distributions. |
| Dataset Splits | Yes | Let {(t1, x1), . . . , (tn, xn)}, x Rd, denote a single multivariate time series. Given a dataset of such time series we want to predict the next P values xn+1, . . . , xn+P . In probabilistic modeling, we want to generate forecasts from learned p(xn+1:n+P |x1:n). |
| Hardware Specification | Yes | We train the VSDM model from scratch on two NVIDIA A100-80G GPUs for two days and generate images from the trained VSDM with the Euler Maruyama numerical solver with 200 discretized steps for generation. |
| Software Dependencies | No | The paper refers to existing code bases (FB-SDE, diffusion distillation code base) for implementation details but does not provide specific version numbers for software dependencies such as Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | We use a batch size of 256 and a learning rate of 1e-4. |