Kai Xu

This looks like a great algorithm to have. You're welcome to fire a PR to add this. Regarding the extra depedencies, what I would suggest is we host codes for...

Feel free to fire a PR to allow optional initial step size in `find_good_step_size`.

I'm not sure if this is the best place to mention this, but the current work-around for `reattach-to-user-namespace`, i.e. below ``` if is_osx && command_exists "reattach-to-user-namespace" && option_value_not_changed "default-command" "";...

I think it's relatively straightforward to make `BroadcastedDistribution` GPU compatible as well.

I like 1 and 3. The reason I don't like 2 is because `val` or `value` doesn't indicate the value is transfomed or not.

Similarly, `ans` is a good choice. It's also Julia's default name holder for results. ```bash ➜ ~ julia _ _ _ _(_)_ | Documentation: https://docs.julialang.org (_) | (_) (_) |...

I prefer `logpdf_with_jac`. Also compared with `forward`, `logpdf_with_jac` is more verbose - when people reads the code they at least know it's doing something about computing the log density with...

tensorboardX supports logging a matplotlib object directly. How hard is it to implmenet this feature here?

Cool. I will give a try.

BTW I found the corresponding helper func in tensorboardX: https://github.com/lanpa/tensorboardX/blob/master/tensorboardX/utils.py#L2