weirayao / leap

LEAP is a tool for discovering latent temporal causal relations with gradient-based neural network.

I'm Weiran! I'm a Research Scientist working on Large Language Agents at Salesforce AI. I got my PhD degree from CMU. I'm a co-founder of TraffiQure, a CMU AI/ML spinoff for curing traffic pains with machine learning solutions.

leap's Issues

ValueError: The parameter loc has invalid values

Hello, I tried to run your code. I rewrote a class similar to SRNNSynthetic to handle my multidimensional time series data, keeping the network structure and training hyperparameters basically unchanged. During this process, I encountered an error. Around line 271 in srnn.py, q_dist = D.Normal(mus, torch.exp(logvars / 2)), it throws an error "ValueError: The parameter loc has invalid values". Upon investigation, I found that during the multiplication of latent variables by the transition, the values explode as time increases, causing the mean and variance to become inf or nan. Do you have any insights into this issue? The data has been preprocessed to ensure that the mean of each observation variable is 0 and the variance is 1. Thank you.