open-mmlab
[Bug] CenterHead with multi task.
Prerequisite
- [X] I have searched Issues and Discussions but cannot get the expected help.
- [X] I have read the FAQ documentation but cannot get the expected help.
- [X] The bug has not been fixed in the latest version (dev-1.x) or latest version (dev-1.0).
Task
I'm using the official example scripts/configs for the officially supported tasks/models/datasets.
Branch
main branch https://github.com/open-mmlab/mmdetection3d
Environment
python3.7
Reproduces the problem - code sample
xxx
Reproduces the problem - command or script
xxx
Reproduces the problem - error message
https://github.com/open-mmlab/mmdetection3d/blob/0f9dfa97a35ef87e16b700742d3c358d0ad15452/mmdet3d/models/dense_heads/centerpoint_head.py#L484
Here, I found in CenterHead.py. in centerpoint models on nuscenes dataset, common_heads are multi-task specific. But I found the above code, while selecting gt boxes according to gt label for one specific task, directly use class name order in current task to index the original class name order in nus-3d.py, I think it should be under the premise of the class name order in common_heads the same as class name order in nus-3d.py. But I found they are not the same. I am confused about it, could you please show me the reason?
Additional information
class names in nus-3d.py
class_names = [
'car', 'truck', 'trailer', 'bus', 'construction_vehicle', 'bicycle',
'motorcycle', 'pedestrian', 'traffic_cone', 'barrier'
]
class names in common_head in file https://github.com/open-mmlab/mmdetection3d/blob/0f9dfa97a35ef87e16b700742d3c358d0ad15452/configs/base/models/centerpoint_voxel01_second_secfpn_nus.py#L42C3-L42C3
tasks=[
dict(num_class=1, class_names=['car']),
dict(num_class=2, class_names=['truck', 'construction_vehicle']),
dict(num_class=2, class_names=['bus', 'trailer']),
dict(num_class=1, class_names=['barrier']),
dict(num_class=2, class_names=['motorcycle', 'bicycle']),
dict(num_class=2, class_names=['pedestrian', 'traffic_cone']),
]
What I mean is that the class_names order of tasks while flattening in one list should be the same as class_names in nus-3d.py. Or we will get a wrong class label using class_name.index(i) + flag which is not aligned with the original class_names index order in gt_instances_3d.labels_3d.
In the final config, we refactor the class_names to adjust the tasks. Please see here (https://github.com/open-mmlab/mmdetection3d/blob/main/configs/centerpoint/centerpoint_pillar02_second_secfpn_8xb4-cyclic-20e_nus-3d.py) for reference.
Another question: I met a problem when using analysis_tools/get_flops.py
/opt/conda/lib/python3.7/site-packages/traits/etsconfig/etsconfig.py:428: UserWarning: Environment variable "HOME" not set, setting home directory to /tmp
% (environment_variable, parent_directory)
Traceback (most recent call last):
File "tools/analysis_tools/get_flops.py", line 83, in <module>
main()
File "tools/analysis_tools/get_flops.py", line 73, in main
flops, params = get_model_complexity_info(model, input_shape)
File "/opt/conda/lib/python3.7/site-packages/mmcv/cnn/utils/flops_counter.py", line 107, in get_model_complexity_info
_ = flops_model(batch)
File "/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1120, in _call_impl
result = forward_call(*input, **kwargs)
File "/root/code/mmdetection3d/mmdet3d/models/detectors/base.py", line 88, in forward
return self._forward(inputs, data_samples, **kwargs)
TypeError: _forward() takes 1 positional argument but 3 were given
how to solve it?
python tools/analysis_tools/get_flops.py \
/root/code/mmdetection3d/configs/centerpoint/centerpoint_voxel0075_second_secfpn_8xb4-cyclic-20e_nus-3d.py \
Run this command
It seems that model input is not only the arg shape, how can I compute the 3D detector flops using the script? Looking forward to your reply and solution.
I get a similar error for Pointpillars #3028. Did you find a solution to fix this problem? @Xiangxu-0103 @DezeZhao
