Question about .register_module() function

[opendrivelab]

Question about .register_module() function

[xanhug]

When I tried to mimic ThinkTwice and create a simple model following the mmdet style, I encountered issues related to registration.

#test_config.py

model = dict(
    type='EncoderDecoder'
)

#encoder_decoder_framework.py

from mmcv.runner import BaseModule
from mmdet.models import DETECTORS


@DETECTORS.register_module()
class EncoderDecoder(BaseModule):
    def __init__(self, train_cfg=None,
                test_cfg=None,):
        super().__init__()

    def forward(self):
        print('SimpleModel forward')

#train.py

from mmdet3d.models import build_model
from mmcv import Config, DictAction
import argparse

def parse_args():
    parser = argparse.ArgumentParser(description='Train a decision-maker')
    parser.add_argument('--config', default='open_loop_training/configs/test_config.py', required=False, help='train config file path')
    args = parser.parse_args()
    return args
def main():
    args = parse_args()
    cfg = Config.fromfile(args.config)

    model = build_model(
        cfg.model,
        train_cfg=None,
        test_cfg=None
    )
    print(model)

if __name__ == "__main__":
    main()

Error occurs:

KeyError: 'EncoderDecoder is not in the models registry'

The program runs successfully and the module is registered only when I place encoder_decoder_framework.py in the python3.7/site-packages/mmdet/models/detectors/ folder and make modifications to the __init__.py file in that folder.

# Copyright (c) OpenMMLab. All rights reserved.
from .atss import ATSS
from .autoassign import AutoAssign
from .base import BaseDetector
from .cascade_rcnn import CascadeRCNN
from .centernet import CenterNet
from .cornernet import CornerNet
from .ddod import DDOD
from .deformable_detr import DeformableDETR
from .detr import DETR
from .fast_rcnn import FastRCNN
from .faster_rcnn import FasterRCNN
from .fcos import FCOS
from .fovea import FOVEA
from .fsaf import FSAF
from .gfl import GFL
from .grid_rcnn import GridRCNN
from .htc import HybridTaskCascade
from .kd_one_stage import KnowledgeDistillationSingleStageDetector
from .lad import LAD
from .mask2former import Mask2Former
from .mask_rcnn import MaskRCNN
from .mask_scoring_rcnn import MaskScoringRCNN
from .maskformer import MaskFormer
from .nasfcos import NASFCOS
from .paa import PAA
from .panoptic_fpn import PanopticFPN
from .panoptic_two_stage_segmentor import TwoStagePanopticSegmentor
from .point_rend import PointRend
from .queryinst import QueryInst
from .reppoints_detector import RepPointsDetector
from .retinanet import RetinaNet
from .rpn import RPN
from .scnet import SCNet
from .single_stage import SingleStageDetector
from .solo import SOLO
from .solov2 import SOLOv2
from .sparse_rcnn import SparseRCNN
from .tood import TOOD
from .trident_faster_rcnn import TridentFasterRCNN
from .two_stage import TwoStageDetector
from .vfnet import VFNet
from .yolact import YOLACT
from .yolo import YOLOV3
from .yolof import YOLOF
from .yolox import YOLOX
from .encoder_decoder_framework import EncoderDecoder

__all__ = [
    'ATSS', 'BaseDetector', 'SingleStageDetector', 'TwoStageDetector', 'RPN',
    'KnowledgeDistillationSingleStageDetector', 'FastRCNN', 'FasterRCNN',
    'MaskRCNN', 'CascadeRCNN', 'HybridTaskCascade', 'RetinaNet', 'FCOS',
    'GridRCNN', 'MaskScoringRCNN', 'RepPointsDetector', 'FOVEA', 'FSAF',
    'NASFCOS', 'PointRend', 'GFL', 'CornerNet', 'PAA', 'YOLOV3', 'YOLACT',
    'VFNet', 'DETR', 'TridentFasterRCNN', 'SparseRCNN', 'SCNet', 'SOLO',
    'SOLOv2', 'DeformableDETR', 'AutoAssign', 'YOLOF', 'CenterNet', 'YOLOX',
    'TwoStagePanopticSegmentor', 'PanopticFPN', 'QueryInst', 'LAD', 'TOOD',
    'MaskFormer', 'DDOD', 'Mask2Former', 'EncoderDecoder'
]

However, it seems that in ThinkTwice, there is no modification to the __init__.py file or placement of encoder_decoder_framework.py in a specific folder. Yet, the module is successfully registered and can be called to create models in other files. How is this achieved?

[xanhug]

Sorry for taking up your valuable time, and thank you for your potential response.

[jiaxiaosong1002]

@xanhug Thanks for your interests. We note that some of the modules are not built by mmcv mechanism and we choose the brute-force building directly by Pytorch for fast development.

Besides, make sure you are reading the document of mmcv 1.x version since the latest/default mmcv 2.0/mmengine has made significant API changes.

[xanhug]

@xanhug Thanks for your interests. We note that some of the modules are not built by mmcv mechanism and we choose the brute-force building directly by Pytorch for fast development.

Besides, make sure you are reading the document of mmcv 1.x version since the latest/default mmcv 2.0/mmengine has made significant API changes.

thanks for your reply.