feat(pnnx): relax F_interpolate_onnx Resize pattern matching

tools/pnnx/src/pass_level2/F_interpolate.cpp

@@ -819,13 +819,17 @@ pnnx.Output             output      1 0 out
 
         if (captured_params.find("op_0.nearest_mode") != captured_params.end())
         {
-            if (captured_params.at("op_0.nearest_mode").type != 4 || captured_params.at("op_0.nearest_mode").s != "floor")
+            if (captured_params.at("op_0.nearest_mode").type != 4)
+                return false;
+            const std::string& nearest_mode = captured_params.at("op_0.nearest_mode").s;
+            if (nearest_mode != "floor" && nearest_mode != "round_prefer_floor")
                 return false;
         }
 
+        // roi may be fused as identity [1,1,...,1] by fuse_constant_as_attribute, accept any float array
         if (captured_params.find("op_0.roi") != captured_params.end())
         {
-            if (captured_params.at("op_0.roi").type != 6 || !captured_params.at("op_0.roi").af.empty())
+            if (captured_params.at("op_0.roi").type != 6)
                 return false;
         }
 
@@ -1045,7 +1049,7 @@ pnnx.Output             output      1 0 out
 
         if (captured_params.find("op_0.roi") != captured_params.end())
         {
-            if (captured_params.at("op_0.roi").type != 6 || !captured_params.at("op_0.roi").af.empty())
+            if (captured_params.at("op_0.roi").type != 6)
                 return false;
         }
 
@@ -1063,7 +1067,10 @@ pnnx.Output             output      1 0 out
 
         if (captured_params.find("op_0.nearest_mode") != captured_params.end())
         {
-            if (captured_params.at("op_0.nearest_mode").type != 4 || captured_params.at("op_0.nearest_mode").s != "floor")
+            if (captured_params.at("op_0.nearest_mode").type != 4)
+                return false;
+            const std::string& nearest_mode = captured_params.at("op_0.nearest_mode").s;
+            if (nearest_mode != "floor" && nearest_mode != "round_prefer_floor")
                 return false;
         }
 

tools/pnnx/tests/onnx/CMakeLists.txt

@@ -13,6 +13,7 @@ pnnx_onnx_add_test(onnx_normalize_ops)
 pnnx_onnx_add_test(onnx_pool_ops)
 pnnx_onnx_add_test(onnx_dense_ops)
 pnnx_onnx_add_test(onnx_rnn_ops)
+pnnx_onnx_add_test(onnx_resize_with_roi)
 pnnx_onnx_add_test(onnx_conv_ops)
 pnnx_onnx_add_test(onnx_opset21_ops)
 

tools/pnnx/tests/onnx/test_onnx_resize_with_roi.py

@@ -0,0 +1,60 @@
+# Copyright 2025 Tencent
+# SPDX-License-Identifier: BSD-3-Clause
+
+import sys
+if sys.version_info < (3, 9):
+    sys.exit(0)
+
+import torch
+import onnx
+import numpy as np
+import onnxruntime as ort
+from onnxscript import FLOAT, script
+from onnxscript import opset19 as op
+
+@script()
+def Model(x2d: FLOAT["N","C","H","W"], x1d: FLOAT["N","C","L"]):
+    roi2d = op.Constant(value=onnx.numpy_helper.from_array(np.array([1,1,1,1,1,1,1,1], dtype=np.float32)))
+    scales2d = op.Constant(value=onnx.numpy_helper.from_array(np.array([1.0,1.0,2.0,2.0], dtype=np.float32)))
+    y2d = op.Resize(x2d, roi2d, scales2d, None, mode='nearest', coordinate_transformation_mode='half_pixel', nearest_mode='round_prefer_floor')
+
+    roi1d = op.Constant(value=onnx.numpy_helper.from_array(np.array([1,1,1,1,1,1], dtype=np.float32)))
+    scales1d = op.Constant(value=onnx.numpy_helper.from_array(np.array([1.0,1.0,2.0], dtype=np.float32)))
+    y1d = op.Resize(x1d, roi1d, scales1d, None, mode='nearest', coordinate_transformation_mode='half_pixel', nearest_mode='round_prefer_floor')
+
+    return (y2d, y1d)
+
+def test():
+    # save onnx
+    onnx.save(Model.to_model_proto(), "test_onnx_resize_with_roi.onnx")
+
+    torch.manual_seed(0)
+    x2d = torch.rand(1, 3, 8, 8)
+    x1d = torch.rand(1, 3, 16)
+
+    # ort inference
+    sess = ort.InferenceSession("test_onnx_resize_with_roi.onnx")
+    a = tuple(torch.from_numpy(out) for out in sess.run(None, {"x2d": x2d.numpy(), "x1d": x1d.numpy()}))
+
+    # onnx to pnnx and ncnn
+    import os
+    os.system("pnnx test_onnx_resize_with_roi.onnx inputshape=[1,3,8,8],[1,3,16] fp16=0")
+
+    # pnnx inference
+    import test_onnx_resize_with_roi_pnnx
+    b = test_onnx_resize_with_roi_pnnx.test_inference()
+
+    # ncnn inference
+    import test_onnx_resize_with_roi_ncnn
+    c = test_onnx_resize_with_roi_ncnn.test_inference()
+
+    for a0, b0, c0 in zip(a, b, c):
+        if not torch.allclose(a0, b0, 1e-3, 1e-3) or not torch.allclose(a0, c0, 1e-3, 1e-3):
+            return False
+    return True
+
+if __name__ == "__main__":
+    if test():
+        exit(0)
+    else:
+        exit(1)