PreDiff: Precipitation Nowcasting with Latent Diffusion Models

Authors: Zhihan Gao, Xingjian Shi, Boran Han, Hao Wang, Xiaoyong Jin, Danielle Maddix, Yi Zhu, Mu Li, Yuyang (Bernie) Wang

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct empirical studies on two datasets: N-body MNIST, a synthetic dataset with chaotic behavior, and SEVIR, a real-world precipitation nowcasting dataset. Specifically, we impose the law of conservation of energy in N-body MNIST and anticipated precipitation intensity in SEVIR. Experiments demonstrate the effectiveness of Pre Diff in handling uncertainty, incorporating domain-specific prior knowledge, and generating forecasts that exhibit high operational utility.
Researcher Affiliation Collaboration Zhihan Gao Hong Kong University of Science and Technology zhihan.gao@connect.ust.hk Xingjian Shi Boson AI xshiab@connect.ust.hk Boran Han AWS boranhan@amazon.com Hao Wang AWS AI Labs howngz@amazon.com Xiaoyong Jin Amazon jxiaoyon@amazon.com Danielle Maddix AWS AI Labs dmmaddix@amazon.com Yi Zhu Boson AI yi@boson.ai Mu Li Boson AI mu@boson.ai Yuyang Wang AWS AI Labs yuyawang@amazon.com
Pseudocode Yes Algorithm 1 One training step of the knowledge alignment network Uϕ
Open Source Code No The paper provides a link for generating the N-body MNIST dataset, but not for the Pre Diff model's source code or the experiment code itself. "Code available at https://github.com/amazon-science/earth-forecasting-transformer/tree/main/src/earthformer/datasets/nbody".
Open Datasets Yes We conduct empirical studies on two datasets: N-body MNIST, a synthetic dataset with chaotic behavior, and SEVIR, a real-world precipitation nowcasting dataset. Dataset is available at https://sevir.mit.edu/ (for SEVIR) and Code available at https://github.com/amazon-science/earth-forecasting-transformer/tree/main/src/earthformer/datasets/nbody (for N-body MNIST).
Dataset Splits Yes We generate 20,000 sequences for training and 1,000 sequences for testing. (N-body MNIST) ... a real-world precipitation nowcasting benchmark SEVIR2 [55] (SEVIR test set implicitly used as a benchmark with standard splits).
Hardware Specification Yes All experiments are conducted on machines with NVIDIA A10G GPUs (24GB memoery).
Software Dependencies No The paper mentions optimizers (Adam, AdamW) and activation functions (SiLU, GELU) but does not provide specific version numbers for software dependencies like Python, PyTorch, or CUDA.
Experiment Setup Yes Detailed configurations are shown in Table 10, Table 11 and Table 12 for the frame-wise VAE, the latent Earthformer-UNet and the knowledge alignment network, respectively. These tables list hyperparameters such as Learning rate, β1, β2, Weight decay, Batch size, Training epochs, Warm up percentage, and Learning rate decay.