Semantic segmentation training model is correct?

[greg007-web]

I dowloaded the semantic model from here:
https://github.com/facebookresearch/detectron2/tree/main/projects/PointRend#semantic-segmentation
And I used this code to train my customized model:

from detectron2.engine import DefaultTrainer
from detectron2.config import get_cfg
from pathlib import Path
import os

cfg = get_cfg()
cfg.merge_from_file("./detectron2_repo/configs/Cityscapes/mask_rcnn_R_50_FPN.yaml")
cfg.DATASETS.TRAIN = ("sem_seg_meat_train",)
cfg.DATASETS.TEST = ()
cfg.DATALOADER.NUM_WORKERS = 2
cfg.MODEL.WEIGHTS =  str(Path("./output/model_final_cf6ac1.pkl").absolute()) ## Downloaded model
cfg.SOLVER.IMS_PER_BATCH = 2
cfg.SOLVER.BASE_LR = 0.00025
cfg.SOLVER.MAX_ITER = 30000
cfg.SOLVER.STEPS = []        # do not decay learning rate
cfg.MODEL.ROI_HEADS.NUM_CLASSES = 3

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

I ran on google colab, and worked and I obtained a model.
But the question is:
I made right? Is it a semantic model? Or is an instance model.

thanks

[innat]

I think you should replace

cfg.merge_from_file("./detectron2_repo/configs/Cityscapes/mask_rcnn_R_50_FPN.yaml")

with something like:

cfg.merge_from_file("..../..../pointrend_semantic_R_101_FPN_1x_cityscapes.yaml")

config.

---

(much better option: use mmsegmentation.)

[greg007-web]

Thanks for your time.
I forgot to cite this, I used this too, but didn`t work, here the code:

from detectron2.engine import DefaultTrainer
from detectron2.config import get_cfg
from detectron2.projects import point_rend
from pathlib import Path
import os

cfg = get_cfg()
point_rend.add_pointrend_config(cfg)
cfg.merge_from_file("./detectron2_repo/projects/PointRend/configs/SemanticSegmentation/pointrend_semantic_R_101_FPN_1x_cityscapes.yaml")#1
cfg.DATASETS.TRAIN = ("sem_seg_meat_train",)
cfg.DATASETS.TEST = ()
cfg.DATALOADER.NUM_WORKERS = 1
cfg.MODEL.WEIGHTS =  str(Path("./output/model_final_cf6ac1.pkl").absolute())
cfg.SOLVER.IMS_PER_BATCH = 1
cfg.SOLVER.BASE_LR = 0.00025
cfg.SOLVER.MAX_ITER = 300000
cfg.SOLVER.STEPS = []        # do not decay learning rate
cfg.MODEL.ROI_HEADS.NUM_CLASSES = 3
#cfg.MODEL.POINT_HEAD.SUBDIVISION_STEPS = 3

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

and I got this error:

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