tiled_pyramid_to_image().

test_box_overlap so they are less confusing and vague.

from objects that are implicitly convertible to a loss details object
just like you can for computational layers.

alias tensors.  Now any kind of tensors are supported.

environments.

expecting an array2d like object.

not destruct and recreate themselves when you switch devices.  Instead, they
keep a table of context objects, for each thread and device, reusing as necessary.

This prevents churn in the context objects when you are switching back and
forth between devices inside a single thread.

order if that's the memory layout of the argument.

with a tensor.

written to with assign_image().  However, you get an error if you try to
assign an image with a different size.

compilation errors when they get pulled into the scope of template
metaprogramming expressions.