fix roi align symbolic function in onnx opset>=16 (#2428)

Author: CescMessi

mmdeploy/mmcv/ops/roi_align.py

@@ -58,23 +58,38 @@ def roi_align_default(g, input: Tensor, rois: Tensor, output_size: List[int],
     else:
         from torch.onnx.symbolic_opset9 import _cast_Long
         from torch.onnx.symbolic_opset11 import add, select
-        batch_indices = _cast_Long(
-            g,
-            g.op(
-                'Squeeze',
-                select(
-                    g, rois, 1,
-                    g.op(
-                        'Constant',
-                        value_t=torch.tensor([0], dtype=torch.long))),
-                axes_i=[1]), False)
+        ir_cfg = get_ir_config(ctx.cfg)
+        opset_version = ir_cfg.get('opset_version', 11)
+        if opset_version < 13:
+            batch_indices = _cast_Long(
+                g,
+                g.op(
+                    'Squeeze',
+                    select(
+                        g, rois, 1,
+                        g.op(
+                            'Constant',
+                            value_t=torch.tensor([0], dtype=torch.long))),
+                    axes_i=[1]), False)
+        else:
+            axes = g.op(
+                'Constant', value_t=torch.tensor([1], dtype=torch.long))
+            batch_indices = _cast_Long(
+                g,
+                g.op(
+                    'Squeeze',
+                    select(
+                        g, rois, 1,
+                        g.op(
+                            'Constant',
+                            value_t=torch.tensor([0], dtype=torch.long))),
+                    axes), False)
         rois = select(
             g, rois, 1,
             g.op(
                 'Constant',
                 value_t=torch.tensor([1, 2, 3, 4], dtype=torch.long)))
-        ir_cfg = get_ir_config(ctx.cfg)
-        opset_version = ir_cfg.get('opset_version', 11)
+
         if opset_version < 16:
             # preprocess rois to make compatible with opset 16-
             # as for opset 16+, `aligned` get implemented inside onnxruntime.
@@ -96,6 +111,10 @@ def roi_align_default(g, input: Tensor, rois: Tensor, output_size: List[int],
                 sampling_ratio_i=sampling_ratio,
                 mode_s=pool_mode)
         else:
+            if aligned:
+                coordinate_transformation_mode = 'half_pixel'
+            else:
+                coordinate_transformation_mode = 'output_half_pixel'
             return g.op(
                 'RoiAlign',
                 input,
@@ -106,4 +125,5 @@ def roi_align_default(g, input: Tensor, rois: Tensor, output_size: List[int],
                 spatial_scale_f=spatial_scale,
                 sampling_ratio_i=sampling_ratio,
                 mode_s=pool_mode,
-                aligned_i=aligned)
+                coordinate_transformation_mode_s=coordinate_transformation_mode
+            )