[TVM Discuss] [Questions] ONNX type mismatch when building with opt level 1

[jonso via TVM Discuss <no...@discuss.tvm.ai>]

Hey all,

I am working on a model that is written in PyTorch and exported to ONNX. During `relay.build` with opt level = 1, I ran into a type mismatch. The error does not occur when opt level = 0:

```
TypeError: Check failed: a.dtype() == b.dtype(): mismatched types
Error during compile function
-----------------------------
v0.0.4
fn (%p0: Tensor[(7, 1, 32, 1536), float32], Primitive=1) -> Tensor[(7, 1, 1536), float32] {
  %0 = reshape(%p0, newshape=[7, 1, -1, 1536]) /* ty=Tensor[(7, 1, 32, 1536), float32] */;
  %1 = take(%0, 0 /* ty=int64 */, axis=2) /* ty=Tensor[(7, 1, 1536), float32] */;
  reshape(%1, newshape=[-1, 1, 1536]) /* ty=Tensor[(7, 1, 1536), float32] */
}
```

This is in `expr.h`, in the function `BinaryOpNode::make`. The issue seems to be coming from the fact that the second input of `take` is an int64. I was able to workaround the issue by casting this argument to int32 in the ONNX frontend, but that solution really isn't ideal.

I traced the `take` call back to this line of PyTorch:
```
new_data = all_data[:,:,0]
```

This is not the first time I've seen small type mismatch errors like this when importing from ONNX, and I don't think that making a change on the PyTorch side is the right way to fix it (I'm not even sure there is a way to fix it, given the simplicitly of this line).

Is there a way we can fix this in TVM? For example, can we cast the int64 to int32? Also, why does it only show up when opt level = 1? Presumably this has to do with fusion, but I haven't been able to figure out the root cause.

Thanks!

cc @jwfromm @masahi





---
[Visit Topic](https://discuss.tvm.ai/t/onnx-type-mismatch-when-building-with-opt-level-1/6368/1) to respond.

You are receiving this because you enabled mailing list mode.

To unsubscribe from these emails, [click here](https://discuss.tvm.ai/email/unsubscribe/25704735ce84009c661318070ff2a296f3cfbe41a975a00b8cb0d220659ffa83).

[TVM Discuss] [Questions] ONNX type mismatch when building with opt level 1

[masahi via TVM Discuss <no...@discuss.tvm.ai>]

yeah, I also remember being annoyed by this int32 vs int64 issue.  I sent some PRs below but I don't have a good solution.

https://github.com/apache/incubator-tvm/pull/4573
https://github.com/apache/incubator-tvm/pull/4528

Fortunately now that we have PyTorch frontend, and I don't need to deal with other frameworks, I'm not using ONNX any more. I recommend switching to Torch frontend if that is also an option for you.





---
[Visit Topic](https://discuss.tvm.ai/t/onnx-type-mismatch-when-building-with-opt-level-1/6368/2) to respond.

You are receiving this because you enabled mailing list mode.

To unsubscribe from these emails, [click here](https://discuss.tvm.ai/email/unsubscribe/39da3ab83952957170a1490ccfc99bb4e6ed4b10eee00a60ede4bdd25dbe4fe4).