siyiweigeHEW opened a new issue, #20143:
URL: https://github.com/apache/tvm/issues/20143
### Expected behavior
A valid ONNX `ConvTranspose` model with **3D transposed convolution** (5D
input `X: (N, C, D1, D2, D3)`) should be imported successfully by
`tvm.relax.frontend.onnx.from_onnx`, just like the 1D and 2D cases. The ONNX
spec allows arbitrary N-D transposed convolution, and
`topi.nn.conv3d_transpose` already exists, so the capability is present.
### Actual behavior
`from_onnx` raises `NotImplementedError`:
```
NotImplementedError: Relax ConvTranspose3d not supported yet
```
raised at `python/tvm/relax/frontend/onnx/onnx_frontend.py:1352` in
`ConvTranspose._impl_v1`, which only handles `ndim == 3` (1D) and `ndim == 4`
(2D) and hard-codes `raise NotImplementedError` for `ndim == 5`.
The same model passes `onnx.checker` and runs correctly in onnxruntime
(output shape `(1, 3, 6, 6, 6)`), confirming it is a valid, runnable ONNX model.
### Environment
- OS: Linux
- TVM: v0.24.dev0 (main branch, commit `262c6d2e0`, built 2026-02-11)
- Python: 3.11
- onnx: 1.20.1
- onnxruntime: 1.24.1
### Steps to reproduce
```python
"""Repro: 3D ConvTranspose is a valid & runnable ONNX model, but TVM relax
frontend rejects it."""
import numpy as np
import onnx, onnxruntime
from onnx import helper, TensorProto
from tvm.relax.frontend.onnx import from_onnx
w = np.zeros((2, 3, 3, 3, 3), dtype="float32")
graph = helper.make_graph(
[helper.make_node("ConvTranspose", ["X", "W"], ["Y"])], "g",
[helper.make_tensor_value_info("X", TensorProto.FLOAT, [1, 2, 4, 4, 4])],
[helper.make_tensor_value_info("Y", TensorProto.FLOAT, [1, 3, 6, 6, 6])],
[helper.make_tensor("W", TensorProto.FLOAT, [2, 3, 3, 3, 3],
w.tolist())],
)
model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 18)])
model.ir_version = 8
onnx.checker.check_model(model) # (1)
valid ONNX model
x = np.random.rand(1, 2, 4, 4, 4).astype("float32")
y = onnxruntime.InferenceSession(model.SerializeToString()).run(None, {"X":
x})[0]
print("onnxruntime runs OK ->", y.shape) # (2)
reference impl works
from_onnx(model, shape_dict={"X": [1, 2, 4, 4, 4]}) # (3) TVM
rejects the same model
```
Actual output:
```
onnxruntime runs OK -> (1, 3, 6, 6, 6)
Traceback (most recent call last):
...
File "tvm/relax/frontend/onnx/onnx_frontend.py", line 1352, in _impl_v1
raise NotImplementedError("Relax ConvTranspose3d not supported yet")
NotImplementedError: Relax ConvTranspose3d not supported yet
```
### Triage
* needs-triage
* bug
* relax
* frontend/onnx
--
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]
