help with Mask R-CNN

FreeLance Project I ONLY want to work with someone who has successfully trained a Mask R-CNN model in python 3.6 on a Windows computer with a custom image dataset. By successfully trained, I mean that the weights produced by your training produced proper masks on those images in your test folder, those images that the model had not seen before. Especially if your image dataset was comprised of 3 or more classes of objects. Especially if those image classes were small objects. I am trying to train a Mask R-CNN model on images of microscope slides. I have 11 classes of objects, each with their own mask color, all of them soil based: bacteria, fungi, 3 kinds of protozoa, 5 kinds of nematodes, and organic matter. All slide images were taken at 400 total magnification, thus all objects are SMALL. If you have done work with Mask R-CNN, you are familiar with the model program itself. There are a couple of different versions of this program. I am using one that is the original tweaked to allow custom datasets. Most of the versions have a samples folder showing use of the model in a couple of different ways. Usually balloon and nucleus are included in this samples folder. These all have, at most, 2 classes of objects; I have 11. I have been trying to adapt these samples for my particular focus. My train image dataset has 22 images, and the valid folder has 19 images, each manually annotated using VGG/VIA. This means that each soil microorganism and all organic matter in these images has a boundary surrounding it, and a "json" file that describes all of the properties of each annotated object: class, boundary type and points, etc. The training module of my program uses this json file to tell it where the masks should be in the training dataset. I can train only the "heads" or the entire model. With just a CPU, heads takes a couple of hours, and all takes about 9 hours. I am not getting any useful masked images with my test images with either the weights from the "heads" training, nor from an "all" training. I've tried training with coco and imagenet weights. There is enough involved here, that I need to ask for ONLY those who have successfully trained a Mask R-CNN model in python 3.6 on a Windows computer with a custom image dataset to reply. I don't want to spend the time having someone get familiar with the Mask R-CNN program. I want someone familiar enough with it, to be able to look at my program to be able to identify what's wrong.