Detectron2: Register custom dataset -AssertionError

[andresviana]

❓ How to register custom dataset in detectron2

Hi guys, from Colab, I would like to learn how to register in detectron2 my custom dataset.
https://rosenfelder.ai/Instance_Image_Segmentation_for_Window_and_Building_Detection_with_detectron2/#prepare-the-data

Inputs
I used via.html to make annotations and save them in two json files (train - val)
Each image has street and hole labels.
images are located in Colab in train and val folders, inside of each folder there are images and a json file called

I found a function in beginer's tutorial for converts them into a format that is usable by detectron2.

I expect how outputs about that function converts images into a format that is usable by detectron2

Code implemented
def get_street_dicts(img_dir):
"""This function loads the JSON file created with the annotator and converts it to
the detectron2 metadata specifications.
"""
# load the JSON file
json_file = os.path.join(img_dir, "via_region_data.json")
with open(json_file) as f:
imgs_anns = json.load(f)

dataset_dicts = []
# loop through the entries in the JSON file
for idx, v in enumerate(imgs_anns.values()):
    record = {}
    # add file_name, image_id, height and width information to the records
    filename = os.path.join(img_dir, v["filename"])
    height, width = cv2.imread(filename).shape[:2]

    record["file_name"] = filename
    record["image_id"] = idx
    record["height"] = height
    record["width"] = width

    annos = v["regions"]

    objs = []
    # one image can have multiple annotations, therefore this loop is needed
    for annotation in annos:
        # reformat the polygon information to fit the specifications
        anno = annotation["shape_attributes"]
        px = anno["all_points_x"]
        py = anno["all_points_y"]
        poly = [(x + 0.5, y + 0.5) for x, y in zip(px, py)]
        poly = [p for x in poly for p in x]

        region_attributes = annotation["region_attributes"]["class"]

        # specify the category_id to match with the class.

        if "street" in region_attributes:
            category_id = 1
        elif "hole" in region_attributes:
            category_id = 0

        obj = {
            "bbox": [np.min(px), np.min(py), np.max(px), np.max(py)],
            "bbox_mode": BoxMode.XYXY_ABS,
            "segmentation": [poly],
            "category_id": category_id,
            "iscrowd": 0,
        }
        objs.append(obj)
    record["annotations"] = objs
    dataset_dicts.append(record)
return dataset_dicts

for d in ["train", "val"]:
DatasetCatalog.register("streets_" + d,lambda d=d: get_street_dicts("/content/potholes/"+ d))
street_metadata = MetadataCatalog.get("streets_train")
dataset_dicts = get_street_dicts("/content/potholes/train")

AssertionError
AssertionError Traceback (most recent call last)
in ()
59 from detectron2.data import DatasetCatalog, MetadataCatalog
60 for d in ["train", "val"]:
---> 61 DatasetCatalog.register("streets_" + d,lambda d=d: get_street_dicts("/content/potholes/", d))
62 street_metadata = MetadataCatalog.get("streets_train")
63 dataset_dicts = get_street_dicts("/content/potholes/train")
/content/detectron2_repo/detectron2/data/catalog.py in register(name, func)
38 assert callable(func), "You must register a function with DatasetCatalog.register!"
39 assert name not in DatasetCatalog._REGISTERED, "Dataset '{}' is already registered!".format(
---> 40 name
41 )
42 DatasetCatalog._REGISTERED[name] = func
AssertionError: Dataset 'streets_train' is already registered!

Thanks for check it.

[ppwwyyxx]

As the error says, the dataset is already registered. Registering it again is expected to cause this error

[RishiMalhotra920]

How do you unregister the dataset. I'm creating a huge number of datasets because I can't unregister them.
Even better - how do I attach a new create_dataset_dicts function to a registered dataset

[ghazni123]

@RishiMalhotra920
I am facing same issue and there is no doc talking about it. did you manage to find how to unregister a dataset? thanks.

[rogertrullo]

@ghazni123 @RishiMalhotra920 I was having the same issue; after reading the code I found there is a clear method that you can use to _unregister_:
DatasetCatalog.clear()