unable to reproduce results

[cnut1648]

unable to reproduce results

[cnut1648]

After several days I have been successfully reproduce the result! Here is my attempt. The code is using python3 and contains FCN training code (use torchvision) and also object detection module (using detectron2). Thank you again for the authors' great work!

[cnut1648]

An update, so I have trained Detectron2 faster rcnn (pretrained on COCO), and the result are quite low

I think I kept the original parameters, and my config is below

class Trainer(DefaultTrainer):
    @classmethod
    def build_evaluator(cls, cfg, dataset_name):
        return COCOEvaluator("synthetic_train", output_dir=os.path.join(cfg.OUTPUT_DIR, "eval", dataset_name))

with open(cutpasted_dir / "label2id.json") as f:
        label2id = json.load(f)

from PIL import ImageFile, Image
ImageFile.LOAD_TRUNCATED_IMAGES = True

register_coco_instances("synthetic_train", metadata=label2id, json_file=str(cutpasted_dir / "COCO.json"), image_root=str(cutpasted_dir / "images"))
register_coco_instances("kitchen_test", metadata={}, json_file=str(data_dir / "Kitchen_fold1_COCO.json"), image_root="")
cfg = get_cfg()

cfg.merge_from_file(model_zoo.get_config_file("COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml"))
cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5  # set threshold for this model
cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url("COCO-Detection/faster_rcnn_R_50_FPN_3x.yaml")
cfg.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE = 256
cfg.MODEL.ROI_HEADS.NUM_CLASSES = len(label2id)

cfg.DATASETS.TRAIN = ("synthetic_train",)
cfg.DATASETS.TEST = ("synthetic_train", )
cfg.DATALOADER.NUM_WORKERS = 10
cfg.TEST.EVAL_PERIOD = 5000

cfg.SEED = 42

cfg.SOLVER.IMS_PER_BATCH = 4
cfg.SOLVER.BASE_LR = 0.001
cfg.SOLVER.MAX_ITER = 25000
cfg.SOLVER.STEPS = (15000, )
cfg.SOLVER.WEIGHT_DECAY = 0.0005

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

what did I miss here? Thank you!

[zeldrinn]

@cnut1648 any chance you still have that code lying around? Getting a 404 :)