support customized dataset about adelaidet HOT 4 CLOSED

finally I found the way.

1. Convert your own dataset to coco-format
2. put the dataset folder in, or link it to "./datasets/ ". "." is the working directory where you execute train_net.py
   the structure should be:
   --datasets
   | |--coco
   | | |--annotations
   | | | |--instances_train2017.json
   | | | |--instances_val2017.json
   | | |--train2017
   | | | |--[put your train images here]
   | | |--val2017
   | | | |--[put your val images here]
   (pay attention to folder name and annotation json file name)
3. find the path of Detecton2 (search your python site-packages)
   and then find detectron2/data/datasets/builtin_meta.py
   locate the method named " _get_coco_instances_meta()"
   comment all the code before return ret and insert your own code

def _get_coco_instances_meta():
    #thing_ids = [k["id"] for k in COCO_CATEGORIES if k["isthing"] == 1]
    #thing_colors = [k["color"] for k in COCO_CATEGORIES if k["isthing"] == 1]
    #assert len(thing_ids) == 80, len(thing_ids)
    ## Mapping from the incontiguous COCO category id to an id in [0, 79]
    #thing_dataset_id_to_contiguous_id = {k: i for i, k in enumerate(thing_ids)}
    #thing_classes = [k["name"] for k in COCO_CATEGORIES if k["isthing"] == 1]
    #ret = {
    #    "thing_dataset_id_to_contiguous_id": thing_dataset_id_to_contiguous_id,
    #    "thing_classes": thing_classes,
    #    "thing_colors": thing_colors,
    #}
    ret = {
        "thing_dataset_id_to_contiguous_id": {your class id mapping}
        # for example : {0:0, 1:1} 
        "tining_classes": [your class names]
        # for example : ["person", "bird"]
        "thing_colors": [color for different class]
        # for example : [[0, 255, 0], [255, 0, 0]]
    return ret

4. run train_net.py in your working directory.

from adelaidet.

@toxic-0518 Thank you for your question. We directly make use of the data loader in detectron2, so you can refer to the document of detectron2 to know how to use your own data.

from adelaidet.

ok, I'll look into detectron2 dataloader.
Thank you!

from adelaidet.

@toxic-0518 thank you very much for sharing it here:-).

from adelaidet.