[Bug] mmdeploy Unsupported nms type: soft_nms.

[liyunlaiii]

Checklist

• I have searched related issues but cannot get the expected help.
• 2. I have read the FAQ documentation but cannot get the expected help.
• 3. The bug has not been fixed in the latest version.

Describe the bug

我在将 ConvNeXt-V2 转为 onnx 时遇到报错：Unsupported nms type: soft_nms.
最新的mmcv nms type 包含了 soft_nms，而当前的mmdeploy并未支持此nms type

Reproduction

from mmdeploy.apis import torch2onnx

img = '/mmdetection/demo/demo.jpg'
work_dir = '/work/'
deploy_cfg = '/mmdeploy/configs/mmdet/detection/detection_onnxruntime_dynamic.py'
device = 'cpu'

save_file = 'ConvNeXt-V2.onnx'
model_cfg = '/mmdetection/projects/ConvNeXt-V2/configs/mask-rcnn_convnext-v2-b_fpn_lsj-3x-fcmae_coco.py'
model_checkpoint = '/checkpoints/mask-rcnn_convnext-v2-b_fpn_lsj-3x-fcmae_coco_20230113_110947-757ee2dd.pth'

torch2onnx(img, work_dir, save_file, deploy_cfg, model_cfg,
        model_checkpoint, device)

Environment

NA

Error traceback

_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
mmdeploy/apis/core/pipeline_manager.py:356: in _wrap
    return self.call_function(func_name_, *args, **kwargs)
mmdeploy/apis/core/pipeline_manager.py:326: in call_function
    return self.call_function_local(func_name, *args, **kwargs)
mmdeploy/apis/core/pipeline_manager.py:275: in call_function_local
    return pipe_caller(*args, **kwargs)
mmdeploy/apis/core/pipeline_manager.py:107: in __call__
    ret = func(*args, **kwargs)
mmdeploy/apis/pytorch2onnx.py:111: in torch2onnx
    optimize=optimize)
mmdeploy/apis/core/pipeline_manager.py:356: in _wrap
    return self.call_function(func_name_, *args, **kwargs)
mmdeploy/apis/core/pipeline_manager.py:326: in call_function
    return self.call_function_local(func_name, *args, **kwargs)
mmdeploy/apis/core/pipeline_manager.py:275: in call_function_local
    return pipe_caller(*args, **kwargs)
mmdeploy/apis/core/pipeline_manager.py:107: in __call__
    ret = func(*args, **kwargs)
mmdeploy/apis/onnx/export.py:148: in export
    verbose=verbose)
/opt/conda/lib/python3.7/site-packages/torch/onnx/__init__.py:280: in export
    custom_opsets, enable_onnx_checker, use_external_data_format)
/opt/conda/lib/python3.7/site-packages/torch/onnx/utils.py:94: in export
    use_external_data_format=use_external_data_format)
/opt/conda/lib/python3.7/site-packages/torch/onnx/utils.py:695: in _export
    dynamic_axes=dynamic_axes)
mmdeploy/apis/onnx/optimizer.py:27: in model_to_graph__custom_optimizer
    graph, params_dict, torch_out = ctx.origin_func(*args, **kwargs)
/opt/conda/lib/python3.7/site-packages/torch/onnx/utils.py:459: in _model_to_graph
    _retain_param_name)
/opt/conda/lib/python3.7/site-packages/torch/onnx/utils.py:422: in _create_jit_graph
    graph, torch_out = _trace_and_get_graph_from_model(model, args)
/opt/conda/lib/python3.7/site-packages/torch/onnx/utils.py:373: in _trace_and_get_graph_from_model
    torch.jit._get_trace_graph(model, args, strict=False, _force_outplace=False, _return_inputs_states=True)
/opt/conda/lib/python3.7/site-packages/torch/jit/_trace.py:1160: in _get_trace_graph
    outs = ONNXTracedModule(f, strict, _force_outplace, return_inputs, _return_inputs_states)(*args, **kwargs)
/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py:1051: in _call_impl
    return forward_call(*input, **kwargs)
/opt/conda/lib/python3.7/site-packages/torch/jit/_trace.py:132: in forward
    self._force_outplace,
/opt/conda/lib/python3.7/site-packages/torch/jit/_trace.py:118: in wrapper
    outs.append(self.inner(*trace_inputs))
/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py:1051: in _call_impl
    return forward_call(*input, **kwargs)
/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py:1039: in _slow_forward
    result = self.forward(*input, **kwargs)
mmdeploy/apis/onnx/export.py:123: in wrapper
    return forward(*arg, **kwargs)
mmdeploy/codebase/mmdet/models/detectors/two_stage.py:91: in two_stage_detector__forward
    x, rpn_results_list, data_samples, rescale=False)
/mmdetection/mmdet/models/roi_heads/base_roi_head.py:123: in predict
    rescale=bbox_rescale)
mmdeploy/codebase/mmdet/models/roi_heads/standard_roi_head.py:68: in standard_roi_head__predict_bbox
    rescale=rescale)
mmdeploy/codebase/mmdet/models/roi_heads/bbox_head.py:145: in bbox_head__predict_by_feat
    keep_top_k=keep_top_k)
mmdeploy/core/optimizers/function_marker.py:266: in g
    rets = f(*args, **kwargs)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

boxes = tensor([[[415.9427, 337.5312, 849.6290, 721.6528],
         [900.3839, 207.7246, 978.8644, 245.0121],
         [357.08...960, 282.0446],
         [403.1835, 317.6036, 859.9055, 727.7901],
         [  0.0000,   0.0000,   0.0000,   0.0000]]])
scores = tensor([[[6.0523e-06, 1.5483e-05, 1.0588e-05,  ..., 1.1332e-09,
          5.3504e-14, 9.9225e-11],
         [7.6693e-0...e-12, 9.5376e-10],
         [1.3522e-04, 1.6717e-06, 3.5402e-05,  ..., 6.7683e-07,
          1.2107e-08, 5.0770e-07]]])
max_output_boxes_per_class = 200, iou_threshold = 0.5, score_threshold = 0.05, pre_top_k = 5000, keep_top_k = 100
output_index = False, nms_type = 'soft_nms'

    @mark(
        'multiclass_nms',
        inputs=['boxes', 'scores'],
        outputs=['dets', 'labels', 'index'])
    def multiclass_nms(boxes: Tensor,
                       scores: Tensor,
                       max_output_boxes_per_class: int = 1000,
                       iou_threshold: float = 0.5,
                       score_threshold: float = 0.05,
                       pre_top_k: int = -1,
                       keep_top_k: int = -1,
                       output_index: bool = False,
                       nms_type='nms'):
        """Apis for multiclass nms."""
        if nms_type == 'nms':
            return _multiclass_nms(
                boxes,
                scores,
                max_output_boxes_per_class=max_output_boxes_per_class,
                iou_threshold=iou_threshold,
                score_threshold=score_threshold,
                pre_top_k=pre_top_k,
                keep_top_k=keep_top_k,
                output_index=output_index)
        elif nms_type == 'nms_rotated':
            return multiclass_nms_rotated(
                boxes,
                scores,
                max_output_boxes_per_class=max_output_boxes_per_class,
                iou_threshold=iou_threshold,
                score_threshold=score_threshold,
                pre_top_k=pre_top_k,
                keep_top_k=keep_top_k)
        elif nms_type == 'nms_match':
            return multiclass_nms_match(
                boxes,
                scores,
                max_output_boxes_per_class=max_output_boxes_per_class,
                iou_threshold=iou_threshold,
                score_threshold=score_threshold,
                pre_top_k=pre_top_k,
                keep_top_k=keep_top_k)
        else:
>           raise NotImplementedError(f'Unsupported nms type: {nms_type}.')
E           NotImplementedError: Unsupported nms type: soft_nms.

mmdeploy/mmcv/ops/nms.py:545: NotImplementedError