Guillaume Lagrange
Guillaume Lagrange
I agree that when running into issues with `tch`, the actual error is never really clear. What happens the most often is trying to use the CUDA version when the...
Had a glimpse at the issue with OP. The error comes from the [input data type](https://github.com/tracel-ai/burn/blob/main/crates/onnx-ir/src/proto_conversion.rs#L234-243): ``` name: images tensor: uint8[batch, 1080, 1920, 3] ``` We don't have uint8 tensor...
I've manually edited the graph to take a float input instead to check for any further incompatibilities. I stumbled upon the gather with scalar input issue as described in #1915....
Cross-posting my comment here as well: The problem you were having with `ToElement` is simply because you were trying to make everything go through another trait for complex conversion. But...
I think `onnx-parse` is simple and clear to the point. `onnx-unify` could also work but I think it's not as clear.
I was convinced I left a comment last time 😅 Thanks for bringing this back up on discord. Is the outer loop necessary to reproduce the issue? That's a lot...
Wow I reviewed this earlier this afternoon, I guess I forgot to submit 😅
Thanks for the quick change! Looks good to go 👌🏻
> During this I noticed that accessing an ID that is not present within the instance of GradientsParams leads to a panic. Hmm I'm not sure this is the cause...
Ahhhh I see the problem. The tensor container does not have the gradients on the autodiff backend, but on the inner backend. ```rust let grad_option: Option = grad.get((self.conv1).weight.id); ``` Your...