Pre-train on diverse mix of video datasets

[JackKelly]

The paper described in this video https://youtu.be/FbRcbM4T-50 suggests that pre-training on a mix of diverse datasets is helpful. For us, we want our models to learn video prediction. So we could train on a mix of satellite imagery, and video like YouTube, moving mnist, simple cartoons of clouds (to teach object permanence etc)