Semantic segmentation with detectron2

Question

I used Detectron2 to train a custom model with Instance Segmentation and worked well. There are several Tutorials on google colab with Detectron2 using Instance Segmentation, but nothing about Semantic Segmentation. So, to train the Custom Instance Segmentation the code based on colab (https://colab.research.google.com/drive/16jcaJoc6bCFAQ96jDe2HwtXj7BMD_-m5#scrollTo=7unkuuiqLdqd) is this:

from detectron2.engine import DefaultTrainer

cfg = get_cfg()
cfg.merge_from_file(model_zoo.get_config_file("COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml"))
cfg.DATASETS.TRAIN = ("balloon_train",)
cfg.DATASETS.TEST = ()
cfg.DATALOADER.NUM_WORKERS = 2
cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url("COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml")  # Let training initialize from model zoo
cfg.SOLVER.IMS_PER_BATCH = 2
cfg.SOLVER.BASE_LR = 0.00025  # pick a good LR
cfg.SOLVER.MAX_ITER = 300    # 300 iterations seems good enough for this toy dataset; you will need to train longer for a practical dataset
cfg.SOLVER.STEPS = []        # do not decay learning rate
cfg.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE = 128   # faster, and good enough for this toy dataset (default: 512)
cfg.MODEL.ROI_HEADS.NUM_CLASSES = 1  # only has one class (ballon). (see https://detectron2.readthedocs.io/tutorials/datasets.html#update-the-config-for-new-datasets)
# NOTE: this config means the number of classes, but a few popular unofficial tutorials incorrect uses num_classes+1 here.

os.makedirs(cfg.OUTPUT_DIR, exist_ok=True)
trainer = DefaultTrainer(cfg) 
trainer.resume_or_load(resume=False)
trainer.train()

And to run Semantic Segmentation train I replaced "COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml" to "/Misc/semantic_R_50_FPN_1x.yaml", basicly I changed the pre-trainded model, just this. And I got this error:

TypeError: cross_entropy_loss(): argument 'target' (position 2) must be Tensor, not NoneType

How I set up to Semantic Segmentation on Google Colab?

score 3 · Accepted Answer · answered May 01 '22 at 13:23

To train for semantic segmentation you can use the same COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml model. You don't have to change this line.

The training code you showed in your question is correct and can be used for semantic segmentation as well. All that changes are the label files.

Once the model is trained, you can use it for inference by loading the model weights from the trained model

cfg = get_cfg()
cfg.merge_from_file(model_zoo.get_config_file("COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml"))
cfg.MODEL.WEIGHTS = os.path.join(cfg.OUTPUT_DIR, "model_final.pth")
cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5   # set the testing threshold for this model
cfg.DATASETS.TEST = ("Detectron_terfspot_" + "test", )                      # the name given to your dataset when loading/registering it
cfg.DATALOADER.NUM_WORKERS = 2
cfg.SOLVER.IMS_PER_BATCH = 2
cfg.MODEL.ROI_HEADS.NUM_CLASSES = 1
predictor = DefaultPredictor(cfg)

The COCOEvaluator has a tasks param. If tasks includes “segm”, will this handle semantic and instance segmentation cases? If yes, does the data (eg ground truth) need to be setup differently for semantic segmentation and instance segmentation? — user3731622, Aug 26 '22 at 20:23

Semantic segmentation with detectron2

1 Answers1