LiRunGuo opened a new pull request, #20406:
URL: https://github.com/apache/tvm/pull/20406
Models exported from PyTorch with `F.interpolate(x, size=...)` currently
fail to import through `from_onnx`, and the failure comes from how the Resize
converter unpacks its optional inputs rather than from the resize itself. There
are two separate triggers.
At opset 12 and below, the PyTorch exporter does not omit `roi` and `scales`
when `sizes` is used; it passes them as empty constant tensors. The converter
only treated a missing input as "not provided", so it saw both `scales` and
`sizes` and failed the `Only one of scales and sizes can be provided in Resize`
assertion. The ONNX spec treats an empty tensor here the same as an omitted
input, so this PR maps zero-element constants to `None` before the check.
At opset 13 and above, when the model has a dynamic batch dimension, the
exporter builds `sizes` as `Concat(Shape(x)[:2], [H_out, W_out])`. That reaches
the converter as a `ShapeExpr` whose leading entries are symbolic, and
`int(val.value)` raised `AttributeError: 'Var' object has no attribute
'value'`. Only the spatial entries are used, and
`relax.op.image.resize2d/resize3d` already accept symbolic output extents, so
the converter now keeps non-constant entries as `PrimExpr`s instead of forcing
them to Python ints. As a side effect, resizing to another tensor's runtime
height and width (sizes taken from a symbolic `Shape`) also imports.
A minimal reproduction:
```python
import io, torch, torch.nn.functional as F, onnx
from tvm.relax.frontend.onnx import from_onnx
class M(torch.nn.Module):
def forward(self, x):
return F.interpolate(x, size=(8, 8), mode="nearest")
for opset, dynamic_axes in [(11, {}), (13, {"x": {0: "N"}})]:
buf = io.BytesIO()
torch.onnx.export(M(), torch.randn(1, 3, 4, 4), buf, opset_version=opset,
input_names=["x"], dynamic_axes=dynamic_axes,
dynamo=False)
from_onnx(onnx.load_from_string(buf.getvalue()))
# opset 11: AssertionError: Only one of scales and sizes can be provided in
Resize.
# opset 13: AttributeError: 'Var' object has no attribute 'value'
```
The new tests cover the empty-`scales` pattern with static and symbolic
batch, `sizes` computed from `Shape(x)` with a symbolic batch, and spatial
sizes taken from a second input with symbolic height and width. Each checks the
imported output shape and compares against onnxruntime; all four fail without
this change. I also checked PyTorch exports of nearest and bilinear
`F.interpolate` at opsets 11, 13 and 17 with batch sizes 1 and 3, which match
onnxruntime to within 2.4e-7.
--
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]
