A previous fix to let validation run across more
than one batch caused an issue with multi-GPU
training. The issue seems to be in how Keras
averages loss and metric values, where it expects
them to be scalars rather than arrays. This fix
causes scalar outputs from a model to remain
scalar in multi-GPU training.

Although a bug, it has no practical side-effect
because it works if the masks are square, which is
the case always during training and this code is
invoked during training.

And cleaner path split in nucleus sample.

And a couple other minor fixes.

The detect() method resizes and molds the input
images before detection. In case we already have
the images molded, which is common in debugging or
visualizations, when loading with load_image_gt()
then use this method.

display_instances() calls plt.show() at the end,
which is good in most cases. But if you render
on a figure with multiple instances, the show()
causes only the first to display. This fixes it.
Also fixes a bug in display_differences().

The Dataset class assigns int image IDs to all
images, but keeps track of their original ID from
the source dataset. It’s easy to get the source ID
if you have the image_id. This new mapping makes
it also easy to get the image_id if you have the
source id.

Instead of showing GT and detections on two images
display_differences() shows them on the same image
for comparison.

config.IMAGE_RESIZE_MODE = "crop" can be used in
training only to pick random crops from training
images.

- Override random colors with your own
- Override image captions with your own
- Choose to show or not show masks
- Choose to show or not show bounding boxes

If you stop training at epoch 15 then continue
from where you left off, it resumes with 17 which
skips 16. Does cause any problems other that being
a bit annoying.

Before the fix, the computed AP score is slightly
lower than it should be. It doesn’t affect
training, but just shows slightly lower results AP
in the notebooks.

This is useful to run prediction on images of
variable sizes without resizing them to a fixed
size. While using mode=“none” is another option,
it requires that you pre-process your images to
ensure the width and height are multiples of 64.
This mode handles that automatically.

- Update ROOT_DIR depending on path of file
- Remove unneeded __init__.py files
- Remove setting matplotlib backend

* create package mrcnn
* move notebooks to samples