When profiling my program, I find that importing cheap_repr is very expensive. Here's one profiling result using py-spy:

https://laike9m.github.io/images/af9b0d4.svg

Importing cheap_repr took ~8% of the total time. In __init__.py, it seems that it should be possible to make @try_register_repr calls lazy. Especially for registering Pandas handlers, they took most of the time. This becomes worse when users have Pandas installed but didn't really call cheap_repr on Pandas objects.