Fix some bugs on dev1.x (#1390)

* fix onnx export unused param

* add cfgoptions in reg test

mmdeploy/apis/onnx/export.py

@@ -25,7 +25,6 @@ def export(model: torch.nn.Module,
            dynamic_axes: Optional[Dict] = None,
            verbose: bool = False,
            keep_initializers_as_inputs: Optional[bool] = None,
-           patch_metas: Dict = {},
            optimize: bool = False):
     """Export a PyTorch model into ONNX format. This is a wrap of
     `torch.onnx.export` with some enhancement.
@@ -67,7 +66,6 @@ def export(model: torch.nn.Module,
         verbose (bool): Enable verbose model on `torch.onnx.export`.
         keep_initializers_as_inputs (bool): Whether we should add inputs for
             each initializer.
-        patch_meta (Dict): The information used to patch the model.
         optimize (bool): Perform optimize on model.
     """
     output_path = output_path_prefix + '.onnx'
@@ -117,8 +115,9 @@ def export(model: torch.nn.Module,
             assert isinstance(
                 input_metas, dict
             ), f'Expect input_metas type is dict, get {type(input_metas)}.'
-            model_forward = model.forward
+            model_forward = patched_model.forward
-            model.forward = partial(model.forward, **input_metas)
+            patched_model.forward = partial(patched_model.forward,
+                                            **input_metas)
 
         torch.onnx.export(
             patched_model,
@@ -133,4 +132,4 @@ def export(model: torch.nn.Module,
             verbose=verbose)
 
         if input_metas is not None:
-            model.forward = model_forward
+            patched_model.forward = model_forward

mmdeploy/apis/pytorch2onnx.py

@@ -67,7 +67,6 @@ def torch2onnx(img: Any,
     if isinstance(model_inputs, list) 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'}
 
     # export to onnx
@@ -107,5 +106,4 @@ def torch2onnx(img: Any,
             dynamic_axes=dynamic_axes,
             verbose=verbose,
             keep_initializers_as_inputs=keep_initializers_as_inputs,
-            patch_metas=patch_metas,
             optimize=optimize)

requirements/runtime.txt

@@ -2,6 +2,7 @@ aenum
 grpcio
 h5py
 matplotlib
+mmengine
 multiprocess
 numpy
 onnx>=1.8.0

tests/regression/mmdet.yml

@@ -328,3 +328,14 @@ models:
     pipelines:
       - *pipeline_seg_ort_dynamic_fp32
       - *pipeline_seg_trt_dynamic_fp32
+
+  - name: RTMDet
+    metafile: configs/rtmdet/metafile.yml
+    model_configs:
+      - configs/rtmdet/rtmdet_s_8xb32-300e_coco.py
+    pipelines:
+      - *pipeline_ort_dynamic_fp32
+      - deploy_config: configs/mmdet/detection/detection_tensorrt_static-640x640.py
+        convert_image: *convert_image
+        backend_test: *default_backend_test
+        sdk_config: *sdk_dynamic

tools/regression_test.py

@@ -487,6 +487,14 @@ def get_backend_fps_metric(deploy_cfg_path: str, model_cfg_path: Path,
     ]
 
     codebase_name = get_codebase(str(deploy_cfg_path)).value
+    # to stop Dataloader OOM in docker CI
+    if codebase_name not in ['medit', 'mmocr']:
+        cfg_options = 'test_dataloader.num_workers=0 ' \
+                      'test_dataloader.persistent_workers=False ' \
+                      'val_dataloader.num_workers=0 ' \
+                      'val_dataloader.persistent_workers=False '
+        cmd_lines.append(f'--cfg-options {cfg_options}')
+
     # Test backend
     return_code = run_cmd(cmd_lines, log_path)
     fps, backend_metric, test_pass = get_fps_metric(return_code,