update training code #22
Conversation
| # cv2.destroyAllWindows() | ||
| # exit() | ||
| cfg = get_cfg() | ||
| cfg.merge_from_file(model_zoo.get_config_file("COCO-PanopticSegmentation/panoptic_fpn_R_50_3x.yaml")) |
There was a problem hiding this comment.
@sayatmimar It's not really clear for me why model initialization from COCO-PanopticSegmentation while dataset schema is built on Instance Segmentation, returned in def real_data(args,image_coordinates_val): line 153
There was a problem hiding this comment.
@AhmedNaglah real_data is a function I created to generate the validation data for the model. Custom Dataset Mapper is not working with CocoEvaluator for some reason. As a note, in the DSA version, we won't have validation so I'm removing that part from this PR.
That's right, here we're not doing panoptic segmentation as we're not using training data for stuff classes i.e. cortical interstitium and medulla. But our pretrained model is trained with those classes and predicts for stuff classes as well, so should we keep COCO-PanopticSegmentation or use instance segmentation from model zoo?
| slide.close() | ||
| maskData=xml_to_mask(dataset_dict['xml_loc'], c, [h,w]) | ||
|
|
||
| if random.random()>0.5: |
There was a problem hiding this comment.
@sayatmimar
Is this random augmentation needed in validation? Since the same CustomDatasetMapper is used in overriding both build_detection_train_loader and build_detection_test_loader
There was a problem hiding this comment.
@AhmedNaglah That's a good point yeah we can use that in validation too, it would be nice to use CustomDatasetMapper with CocoEvaluator then we don't need to do duplicated work.
This PR adds the multiC training code in the form of DSA plugin.