* create package mrcnn
* move notebooks to samples

This also changes utils.resize_image() such that
it doesn’t convert images to 0-255 range, but
rather keep the input range. This would only
affect users who use input image of range 0-1
rather than 0-255.

This doesn’t change the functionality, other than
cleaning up the code to prepare for future changes
to support detection on different image sizes.

The image_meta tensor stores meta data about the
image. This update adds the scale by which the
image was resized and it’s final shape.

Make DetectionLayer return normalized coordinates
to avoid unnecessary conversion to pixels and back
to normalized.

Move code to separate functions and fix off-by-1
error.

Allows using a config variable, TRAIN_NB, rather
that updating the BatchNorm class.

skimage.transform.resize() crashes with a floating
point exception if the image has 1-pixel width or
height and the mode=“reflection”. Changing mode to
“constant” (the default) fixes the issue. And, 
it’s the correct mode to use anyway.

If the mask is a rectangle, its mini-mask pixels
become all 1s, and it has no zeros. Scipy imresize
fails on that and returns 0s instead. This update
fixes the problem.

Requires installing imgaug (pip3 install imgaug)
https://github.com/aleju/imgaug

List of augmentations:
http://imgaug.readthedocs.io/en/latest/source/augmenters.html

This is more reliable than using the batch size to
determine worker count because load per batch of 
images varies depending on image sizes and where
they’re loaded from.

shapes more-or-less match now. weights load. current error=InvalidArgumentError (see above for traceback): WhereOp: Unhandled input dimensions: 0

Remove py_func for cross-environment serializability. Convert DetectionLayer/refine_detections to tf

* Remove dropout layer. Unnecessary.
* Reduce anchor stride from 2 to 1
* Increase training proposals to 200/image
* Reduce pre-NMS limit to 6000
* Improve computing proposal post:neg ratio
* Updated COCO training schedule