Pytorch-segmentation-models-trainer

- New Naive Mod PolyMapper model (Object Detection + PolygonRNN);
- New Naive Mod Polymapper dataset;
- New callback: Frame Field Only Crossfield Warmup Callback;
- New inference processors for Object Detection and PolygonRNN;
- Bug fix on object detection model;
- Bug fix on bounding box mask building;
- Bug fix on polygon iou with invalid geometries;
- Minor code refactor;

- Bug fix on PolygonRNN polygon tokenizer.

- Bug fix on convert dataset;
- Bug fix on PolygonRNNDataset;
- Bug fix on PolygonRNNResultCallback when using gpu;
- Bug fix on PolygonRNNPLModel;

- Vector IOU;
- Polis metric added;
- IoU added to PolygonRNN training loop;
- Object detection visualization callback added;
- PolygonRNN visualization callback added;
- Bug fix on polygon building on build mask geometry handling;

- Bug fixes on SegLoss parameters;

- Dataset conversion added. It is possible to convert between some formats of dataset;
- Tversky Loss and Focal Tversky Loss added;
- LabelSmoothingLoss added;
- MixUpAugmentationLoss added;
- KnowledgeDistillationLoss added;
- Mixup augmentation added to Frame Field Model;