I currently try to get EfficientPS (https://github.com/vincrichard/EfficientPS) running.
When I start the training, the initialisation of the net crashes due to a key error belonging to detectron2.

我目前正在尝试运行EfficientPS(https://github.com/vincrichard/EfficientPS)。当我开始训练时，网络的初始化由于属于Detectron2的关键错误而崩溃。

During execution of the init method of InstanceHead, the calls

在执行InstanceHead的init方法期间，调用

self.rpm = build_proposal_generator(cfg, input_shape)
self.roi_heads = build_roi_heads(cfg, input_shape)

fail with output:

失败，输出：

Traceback (most recent call last):
build-agent-1  |   File "/opt/.pycharm_helpers/pydev/pydevd.py", line 1496, in _exec
build-agent-1  |     pydev_imports.execfile(file, globals, locals)  # execute the script
build-agent-1  |   File "/opt/.pycharm_helpers/pydev/_pydev_imps/_pydev_execfile.py", line 18, in execfile
build-agent-1  |     exec(compile(contents+"\n", file, 'exec'), glob, loc)
build-agent-1  |   File "/opt/project/code/perception/src/panoptic_segmentation/src/training/train_net.py", line 168, in <module>
build-agent-1  |     main()
build-agent-1  |   File "/opt/project/code/perception/src/panoptic_segmentation/src/training/train_net.py", line 139, in main
build-agent-1  |     efficientps = EffificientPS(cfg)
build-agent-1  |   File "/opt/project/code/perception/src/panoptic_segmentation/src/efficientps/model.py", line 35, in __init__
build-agent-1  |     self.instance_head = InstanceHead(cfg)
build-agent-1  |   File "/opt/project/code/perception/src/panoptic_segmentation/src/efficientps/instance_head/instance_head.py", line 22, in __init__
build-agent-1  |     self.roi_heads = build_roi_heads(cfg, input_shape)
build-agent-1  |   File "/home/carla/.local/lib/python3.8/site-packages/detectron2/modeling/roi_heads/roi_heads.py", line 43, in build_roi_heads
build-agent-1  |     return ROI_HEADS_REGISTRY.get(name)(cfg, input_shape)
build-agent-1  |   File "/home/carla/.local/lib/python3.8/site-packages/fvcore/common/registry.py", line 71, in get
build-agent-1  |     raise KeyError(
build-agent-1  | KeyError: "No object named 'CustomROIHeads' found in 'ROI_HEADS' registry!"

for roi_heads and:

对于ROI_HEADS和：

Traceback (most recent call last):
build-agent-1  |   File "/opt/project/code/perception/src/panoptic_segmentation/src/training/train_net.py", line 168, in <module>
build-agent-1  |     main()
build-agent-1  |   File "/opt/project/code/perception/src/panoptic_segmentation/src/training/train_net.py", line 139, in main
build-agent-1  |     efficientps = EffificientPS(cfg)
build-agent-1  |   File "/opt/project/code/perception/src/panoptic_segmentation/src/efficientps/model.py", line 35, in __init__
build-agent-1  |     self.instance_head = InstanceHead(cfg)
build-agent-1  |   File "/opt/project/code/perception/src/panoptic_segmentation/src/efficientps/instance_head/instance_head.py", line 22, in __init__
build-agent-1  |     self.rpn = build_proposal_generator(cfg, input_shape=input_shape)
build-agent-1  |   File "/home/carla/.local/lib/python3.8/site-packages/detectron2/modeling/proposal_generator/build.py", line 24, in build_proposal_generator
build-agent-1  |     return PROPOSAL_GENERATOR_REGISTRY.get(name)(cfg, input_shape)
build-agent-1  |   File "/home/carla/.local/lib/python3.8/site-packages/fvcore/common/registry.py", line 71, in get
build-agent-1  |     raise KeyError(
build-agent-1  | KeyError: "No object named 'RPNCustom' found in 'PROPOSAL_GENERATOR' registry!"

for rpn.

用于RPN。

I read about the registry in the documentation and it says that the models have to be registered in the registries.
I searched in the code and both models have the required registration call:

我在文档中读到了注册表，它说模型必须在注册表中注册。我在代码中搜索了一下，两款车都有必要的注册电话：

@PROPOSAL_GENERATOR_REGISTRY.register()
class RPNCustom(RPN):

    def forward(...):
        ...

@ROI_HEADS_REGISTRY.register()
class CustomROIHeads(ROIHeads):
    ...

The corresponding part of the config:

配置的相应部分：

MODEL:
  ANCHOR_GENERATOR:
    SIZES: [[32], [64], [128], [256]]  # One size for each in feature map
    ASPECT_RATIOS: [[0.5, 1.0, 2.0]]  # Three aspect ratios (same for all in feature maps)
  PROPOSAL_GENERATOR:
    NAME: "RPNCustom"
  RPN:
    HEAD_NAME: "DepthwiseSepRPNHead" # Normal RPN Head "StandardRPNHead"
    IN_FEATURES: ["P_4", "P_8", "P_16", "P_32"]
    PRE_NMS_TOPK_TRAIN: 2000  # Per FPN level
    PRE_NMS_TOPK_TEST: 2000  # Per FPN level
    BBOX_REG_LOSS_TYPE: "smooth_l1"
    BBOX_REG_LOSS_WEIGHT: 1.0
    SMOOTH_L1_BETA: 0.11111111 # 1.0 / 9.0

    # Detectron1 uses 2000 proposals per-batch,
    # (See "modeling/rpn/rpn_outputs.py" for details of this legacy issue)
    # which is approximately 1000 proposals per-image since the default batch size for FPN is 2.
    POST_NMS_TOPK_TRAIN: 1000
    POST_NMS_TOPK_TEST: 1000
    SMOOTH_L1_BETA: 0.1111
    IOU_THRESHOLDS: [0.3, 0.7]
  ROI_HEADS:
    NAME: "CustomROIHeads"
    # BATCH_SIZE_PER_IMAGE: 256 # number of proposals to sample for training
    # POSITIVE_FRACTION: 0.25 # fraction of positive (foreground) proposals to sample for training.
    IN_FEATURES: ["P_4", "P_8", "P_16", "P_32"]
    NUM_CLASSES: 8 # There is 8 instance in the city scape dataset
    # PROPOSAL_APPEND_GT:
    IOU_THRESHOLDS: [0.5]
    # IOU_LABELS:
    SCORE_THRESH_TEST: 0.5 # First step of panoptic fusion module
    NMS_THRESH_TEST: 0.5 # Second step of panoptic fusion module
  ROI_BOX_HEAD:
    POOLER_RESOLUTION: 7
    POOLER_SAMPLING_RATIO: 2 # (maybe put to 2) The `sampling_ratio` parameter for the ROIAlign op.
    POOLER_TYPE: "ROIAlign" # "ROIAlignV2"
    SMOOTH_L1_BETA: 1.0
    # SCORE_THRESH_TEST: 0.05
    # NMS_THRESH_TEST: 0.5
    BBOX_REG_LOSS_TYPE: "smooth_l1"
    SMOOTH_L1_BETA: 1.0
    BBOX_REG_LOSS_WEIGHT: 1.0
  ROI_MASK_HEAD:
    POOLER_RESOLUTION: 14
    POOLER_TYPE: "ROIAlign"
TEST:
  DETECTIONS_PER_IMAGE: 100

I never worked with detectron2 and its registries before, so I don't know what the problem here is.

我以前从未使用过Detectron2及其注册表，所以我不知道这里有什么问题。