Expected behavior
Resize with coordinate_transformation_mode=half_pixel (or pytorch_half_pixel, asymmetric) and non-integer scale factors (e.g., 2.5x) should compute source coordinates as:
src = (dst + 0.5) / scale - 0.5
and produce results matching ONNX Runtime.
Actual behavior
TVM computes incorrect source coordinates, leading to wrong pixel mapping. For a 3×3 input with scale=2.5, the max absolute difference vs ORT is 4.0 (nearest mode) and 0.63 (linear mode).
Reproduction
import numpy as np
import onnx
from onnx import helper, TensorProto, numpy_helper
import onnxruntime as ort
import tvm
from tvm import relax
from tvm.relax.frontend.onnx import from_onnx
x = np.arange(9, dtype=np.float32).reshape(1, 1, 3, 3)
X = helper.make_tensor_value_info("X", TensorProto.FLOAT, [1, 1, 3, 3])
Y = helper.make_tensor_value_info("Y", TensorProto.FLOAT, None)
roi = numpy_helper.from_array(np.array([], dtype=np.float32), "roi")
sc = numpy_helper.from_array(np.array([1, 1, 2.5, 2.5], dtype=np.float32), "scales")
node = helper.make_node("Resize", ["X", "roi", "scales"], ["Y"],
mode="nearest",
coordinate_transformation_mode="half_pixel",
nearest_mode="floor")
graph = helper.make_graph([node], "test", [X], [Y], initializer=[roi, sc])
model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 18)])
model = onnx.shape_inference.infer_shapes(model)
# ORT
sess = ort.InferenceSession(model.SerializeToString())
ort_out = sess.run(None, {"X": x})[0]
# TVM
mod = from_onnx(model)
exe = tvm.relax.build(
tvm.ir.transform.Sequential([relax.transform.LegalizeOps()])(mod), target="llvm")
vm = tvm.relax.VirtualMachine(exe, device=tvm.cpu())
tvm_out = vm["main"](tvm.runtime.tensor(x, device=tvm.cpu())).numpy()
print("Max diff:", np.abs(ort_out - tvm_out).max()) # 4.0Affected configurations
| mode |
coordinate_transformation_mode |
max_diff vs ORT |
| nearest |
half_pixel |
4.0 |
| nearest |
pytorch_half_pixel |
4.0 |
| linear |
half_pixel |
0.63 |
| linear |
asymmetric |
0.46 |
| linear |
pytorch_half_pixel |
0.63 |
Trigger condition: non-integer scale factor (e.g., 2.5) on odd spatial dims (e.g., 3×3). Integer scales (2x, 3x) appear correct.
Root cause
The coordinate transformation for half_pixel mode applies floor((dst + 0.5) / scale - 0.5) but the implementation appears to use a different rounding or offset, causing the source index to be off by one pixel for certain output positions.
Environment
- TVM: 0.24.dev0, commit 0b0afd8 (2026-04-24)
- Python: 3.11
- OS: Linux
cc @KJlaccHoeUM9l @junrushao
Expected behavior
Resizewithcoordinate_transformation_mode=half_pixel(orpytorch_half_pixel,asymmetric) and non-integer scale factors (e.g., 2.5x) should compute source coordinates as:and produce results matching ONNX Runtime.
Actual behavior
TVM computes incorrect source coordinates, leading to wrong pixel mapping. For a 3×3 input with scale=2.5, the max absolute difference vs ORT is 4.0 (nearest mode) and 0.63 (linear mode).
Reproduction
Affected configurations
Trigger condition: non-integer scale factor (e.g., 2.5) on odd spatial dims (e.g., 3×3). Integer scales (2x, 3x) appear correct.
Root cause
The coordinate transformation for
half_pixelmode appliesfloor((dst + 0.5) / scale - 0.5)but the implementation appears to use a different rounding or offset, causing the source index to be off by one pixel for certain output positions.Environment
cc @KJlaccHoeUM9l @junrushao