[Fix]: fix pplnn empty output error (open-mmlab#320)

* fix ppl problems

* fix roialign

mmdeploy/backend/pplnn/wrapper.py

@@ -180,7 +180,13 @@ def forward(self, inputs: Dict[str,
         for i in range(self.runtime.GetOutputCount()):
             out_tensor = self.runtime.GetOutputTensor(i).ConvertToHost()
             name = self.output_names[i]
-            outputs[name] = torch.from_numpy(np.array(out_tensor, copy=False))
+            if out_tensor:
+                outputs[name] = np.array(out_tensor, copy=False)
+            else:
+                out_shape = self.runtime.GetOutputTensor(
+                    i).GetShape().GetDims()
+                outputs[name] = np.random.rand(*out_shape)
+            outputs[name] = torch.from_numpy(outputs[name])
         return outputs
 
     @TimeCounter.count_time()

mmdeploy/mmcv/ops/roi_align.py

@@ -4,6 +4,7 @@
 from torch import Tensor
 
 from mmdeploy.core import SYMBOLIC_REWRITER
+from mmdeploy.utils import Backend, get_backend
 
 
 # Here using mmcv.ops.roi_align.__self__ to find
@@ -36,9 +37,13 @@ def roi_align_default(ctx, g, input: Tensor, rois: Tensor,
     Returns:
         MMCVRoiAlign op for onnx.
     """
-
+    backend = get_backend(ctx.cfg)
+    if backend == Backend.PPLNN:
+        domain = 'mmcv'
+    else:
+        domain = 'mmdeploy'
     return g.op(
-        'mmdeploy::MMCVRoiAlign',
+        f'{domain}::MMCVRoiAlign',
         input,
         rois,
         output_height_i=output_size[0],