feat(codebase): support mmdet3d dev-1.x

configs/mmdet3d/voxel-detection/voxel-detection_static.py

@@ -2,5 +2,5 @@
 codebase_config = dict(
     type='mmdet3d', task='VoxelDetection', model_type='end2end')
 onnx_config = dict(
-    input_names=['voxels', 'num_points', 'coors'],
-    output_names=['scores', 'bbox_preds', 'dir_scores'])
+    input_names=['placeholder', 'num_points', 'voxels', 'coors'],
+    output_names=['cls_score', 'bbox_pred', 'dir_cls_pred'])

docs/en/07-developer-guide/architecture.md

@@ -24,7 +24,7 @@ $ tree -L 1
 ├── service                    # Some small boards not support python, we use C/S mode for model conversion, here is server code
 ├── tests                         # unittest
 ├── third_party           # 3rd party dependencies required by SDK and FFI
-└── tools                        # Tools are also the entrance to all functions, such as onnx2xx.py, profile.py, test.py, etc.
+└── tools                        # Tools are also the entrance to all functions, such as onnx2xx.py, profiler.py, test.py, etc.
 ```
 
 It should be clear

docs/zh_cn/07-developer-guide/architecture.md

@@ -24,7 +24,7 @@ $ tree -L 1
 ├── service                    # 有些小板子不能跑 python，模型转换用的 C/S 模式。这个目录放 Server
 ├── tests                         # 单元测试
 ├── third_party           # SDK 和 ffi 要的第三方依赖
-└── tools                        # 工具，也是一切功能的入口。除了 deploy.py 还有 onnx2xx.py、profile.py 和 test.py
+└── tools                        # 工具，也是一切功能的入口。除了 deploy.py 还有 onnx2xx.py、profiler.py 和 test.py
 ```
 
 这样大致应该清楚了

mmdeploy/apis/pytorch2onnx.py

@@ -60,15 +60,12 @@ def torch2onnx(img: Any,
     task_processor = build_task_processor(model_cfg, deploy_cfg, device)
 
     torch_model = task_processor.build_pytorch_model(model_checkpoint)
-    data, model_inputs = task_processor.create_input(
-        img,
-        input_shape,
-        data_preprocessor=getattr(torch_model, 'data_preprocessor', None))
+    data, model_inputs = task_processor.create_input(img, input_shape)
     if not isinstance(model_inputs, torch.Tensor) and len(model_inputs) == 1:
         model_inputs = model_inputs[0]
     data_samples = data['data_samples']
     patch_metas = {'data_samples': data_samples}
-    input_metas = {'data_samples': data_samples, 'mode': 'predict'}
+    input_metas = {'data_samples': data_samples, 'mode': task_processor.mode()}
 
     # export to onnx
     context_info = dict()

mmdeploy/apis/visualize.py

@@ -68,13 +68,19 @@ def visualize_model(model_cfg: Union[str, mmengine.Config],
             model = task_processor.build_backend_model(model)
 
     model_inputs, _ = task_processor.create_input(img, input_shape)
+
     with torch.no_grad():
         result = model.test_step(model_inputs)[0]
 
+    data_samples = None
+    if 'data_samples' in model_inputs:
+        data_samples = model_inputs['data_samples']
+
     task_processor.visualize(
         image=img,
         model=model,
         result=result,
+        data_samples=data_samples,
         output_file=output_file,
         window_name=backend.value,
         show_result=show_result)

mmdeploy/backend/tensorrt/utils.py

@@ -1,5 +1,8 @@
 # Copyright (c) OpenMMLab. All rights reserved.
 import logging
+import os
+import re
+import sys
 from typing import Dict, Optional, Sequence, Union
 
 import onnx
@@ -38,6 +41,54 @@ def load(path: str) -> trt.ICudaEngine:
         return engine
 
 
+def search_cuda_version() -> str:
+    """try cmd to get cuda version, then try `torch.cuda`
+    Returns:
+        str: cuda version, for example 10.2
+    """
+
+    version = None
+
+    pattern = re.compile(r'[0-9]+\.[0-9]+')
+    platform = sys.platform.lower()
+
+    def cmd_result(txt: str):
+        cmd = os.popen(txt)
+        return cmd.read().rstrip().lstrip()
+
+    if platform == 'linux' or platform == 'darwin' or platform == 'freebsd':  # noqa E501
+        version = cmd_result(
+            " nvcc --version | grep  release | awk '{print $5}' | awk -F , '{print $1}' "  # noqa E501
+        )
+        if version is None or pattern.match(version) is None:
+            version = cmd_result(
+                " nvidia-smi  | grep CUDA | awk '{print $9}' ")
+
+    elif platform == 'win32' or platform == 'cygwin':
+        # nvcc_release = "Cuda compilation tools, release 10.2, V10.2.89"
+        nvcc_release = cmd_result(' nvcc --version | find "release" ')
+        if nvcc_release is not None:
+            result = pattern.findall(nvcc_release)
+            if len(result) > 0:
+                version = result[0]
+
+        if version is None or pattern.match(version) is None:
+            # nvidia_smi = "| NVIDIA-SMI 440.33.01    Driver Version: 440.33.01    CUDA Version: 10.2     |" # noqa E501
+            nvidia_smi = cmd_result(' nvidia-smi | find "CUDA Version" ')
+            result = pattern.findall(nvidia_smi)
+            if len(result) > 2:
+                version = result[2]
+
+    if version is None or pattern.match(version) is None:
+        try:
+            import torch
+            version = torch.version.cuda
+        except Exception:
+            pass
+
+    return version
+
+
 def from_onnx(onnx_model: Union[str, onnx.ModelProto],
               output_file_prefix: str,
               input_shapes: Dict[str, Sequence[int]],
@@ -127,6 +178,15 @@ def from_onnx(onnx_model: Union[str, onnx.ModelProto],
         profile.set_shape(input_name, min_shape, opt_shape, max_shape)
     config.add_optimization_profile(profile)
 
+    cuda_version = search_cuda_version()
+    if cuda_version is not None:
+        version_major = int(cuda_version.split('.')[0])
+        if version_major < 11:
+            # cu11 support cublasLt, so cudnn heuristic tactic should disable CUBLAS_LT # noqa E501
+            tactic_source = config.get_tactic_sources() - (
+                1 << int(trt.TacticSource.CUBLAS_LT))
+            config.set_tactic_sources(tactic_source)
+
     if fp16_mode:
         if version.parse(trt.__version__) < version.parse('8'):
             builder.fp16_mode = fp16_mode

mmdeploy/codebase/base/task.py

@@ -73,6 +73,7 @@ def build_data_preprocessor(self):
 
         from mmengine.registry import MODELS
         data_preprocessor = MODELS.build(preprocess_cfg)
+        data_preprocessor.to(self.device)
 
         return data_preprocessor
 
@@ -295,6 +296,15 @@ def visualize(self,
             show=show_result,
             out_file=output_file)
 
+    @abstractmethod
+    def mode(self) -> str:
+        """Get pytorch inference mode, it depends on codebase implementation.
+
+        Returns:
+            str: codebase inference mode
+        """
+        return 'predict'
+
     @staticmethod
     @abstractmethod
     def get_partition_cfg(partition_type: str, **kwargs) -> Dict:

mmdeploy/codebase/mmdet3d/deploy/__init__.py

@@ -1,6 +1,5 @@
 # Copyright (c) OpenMMLab. All rights reserved.
-from .mmdetection3d import MMDetection3d
-from .voxel_detection import VoxelDetection
+from .voxel_detection import MMDetection3d, VoxelDetection
 from .voxel_detection_model import VoxelDetectionModel
 
 __all__ = ['MMDetection3d', 'VoxelDetection', 'VoxelDetectionModel']