Potential performance improvement using multiprocess

[wookayin]

GIL sucks.

One alternative way to make it much faster would be to serialize the function body (what if it contains closure?) and execute the function in a multiprocess fashion (e.g. from a child process). This is quite tricky in general.

[wookayin]

Another crazy idea is that we can draw the plot lazily by storing the Tensor arguments and the function body (serialized, cannot contain closures, maybe) into summary eventfiles and then executing the plot code from the tensorboard process. A custom tensorboard extension will be needed to make it work.