The current version of torchdyn contains various quickstart examples / tutorials which explore different aspects of continuous / implicit learning and related numerical methods. Here, we plan to introduce standardized benchmarks for a variety of domains most relevant to the field, focusing on: time series tasks, generative modeling, optimal control (and some image classification).

These benchmarks are designed to inform practitioners of particular SOTA ML + numerics techniques for neural differential equations and implicit models, given a particular target application. We are also planning to release loadable pretrained weights for top scoring models.