wuyii8941 opened a new issue, #19690:
URL: https://github.com/apache/tvm/issues/19690
## Expected behavior
ONNX `AffineGrid` defaults to `align_corners=0`. Both ONNX Runtime and
PyTorch implement the `align_corners=0` ("center of pixel") sampling convention.
## Actual behavior
The TVM converter rejects `align_corners=0`:
```
NotImplementedError: AffineGrid with align_corners=0 is not yet supported in
TVM
```
Because the ONNX default is 0 (the attribute is normally omitted from the
graph), **every** ONNX model that uses `AffineGrid` without explicitly setting
`align_corners=1` fails to import.
## Reproduction
```python
import numpy as np
import onnx
from onnx import helper, TensorProto, numpy_helper
import onnxruntime as ort
from tvm.relax.frontend.onnx import from_onnx
theta = helper.make_tensor_value_info("theta", TensorProto.FLOAT, [2, 2, 3])
Y = onnx.ValueInfoProto(); Y.name = "Y"
size = numpy_helper.from_array(np.array([2, 1, 4, 5], dtype=np.int64),
"size")
# align_corners defaults to 0 — attribute omitted from node, as ONNX
exporters do
node = helper.make_node("AffineGrid", ["theta", "size"], ["Y"])
g = helper.make_graph([node], "g", [theta], [Y], initializer=[size])
m = helper.make_model(g, opset_imports=[helper.make_opsetid("", 20)])
theta_v = np.random.randn(2, 2, 3).astype(np.float32)
print("ORT shape:", ort.InferenceSession(m.SerializeToString()).run(None,
{"theta": theta_v})[0].shape)
# ORT shape: (2, 4, 5, 2)
inf = onnx.shape_inference.infer_shapes(m)
mod = from_onnx(inf) # NotImplementedError: AffineGrid with align_corners=0
...
```
## Root cause
`python/tvm/relax/frontend/onnx/onnx_frontend.py`, `AffineGrid._impl_v20`:
```python
align_corners = attr.get("align_corners", 0)
if align_corners != 1:
raise NotImplementedError("AffineGrid with align_corners=0 is not yet
supported in TVM")
```
The underlying `relax.op.image.affine_grid` (and corresponding TOPI
implementation) currently only implements the `align_corners=1` formula.
For 2D, the difference is the source-coordinate mapping:
- `align_corners=1`: `x = (-1, +1) ⟼ (0, W-1)` — corners aligned
- `align_corners=0`: `x = (-1, +1) ⟼ (-0.5, W-0.5)` — pixel centers aligned
This is the standard PyTorch convention and is what ONNX Runtime implements.
## Suggested fix
Plumb `align_corners` into `relax.op.image.affine_grid` /
`topi.image.affine_grid` so both modes are supported, mirroring how
`image.resize2d` already takes `coordinate_transformation_mode`.
Minimal short-term workaround would be to emit the coordinate offsets in the
frontend itself (compute the grid via composing `arange + scale + translate`
with `align_corners=0` math), but the cleaner fix is in the op.
## Impact
Every ONNX export that uses `AffineGrid` without explicitly setting
`align_corners=1` (i.e., almost all of them) fails to import. This blocks
Spatial Transformer Networks and any geometric-warping model exported from
PyTorch with default arguments.
## Environment
- TVM: latest `main` (commit b172d5ea3)
- Python: 3.11
- ONNX Runtime: 1.24.4
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
