Guillaume Lagrange

Since the original issue has been resolved, closing this issue in favor of #2192.

@rubenjr0 is this still an issue?

Thanks @sdroege, appreciate you offering to check it out! I am not familiar with GStreamer, so that would be very helpful. @vesuvisian I see you're pulling burn from git, I...

Btw you can still write on discord, just embed the logs as a file. Would love to help, but this will be difficult to reproduce without a minimal example.

Ahhh ok that might be the case. We have an open issue for that (#1802).

While the lack of an equivalent to `retain_graph` is really a missing feature, the issue you're encountering could be a bug. Without `retain_graph` you can't calculate the gradients multiple times...

> I think eventually I would prefer onnx-ir tracked shape information. However, I do not want to burden Burn onnx implementers right now. This was also my first concern regarding...

Another issue just resurfaced that motivates the changes proposed in this PR to better track shapes. Since most shapes are set to `None` (in favor of runtime shapes), there is...

> We don't need to track the output shapes statically because the Burn API tracks and calculates dynamically. Burn needs only rank output. Indeed, Burn doesn't need to track that....

Sorry for the delayed response! > I think .set_require_grad may be coercing to a Float tensor? As you probably noticed, the [current implementation](https://github.com/tracel-ai/burn/blob/main/crates/burn-import/src/burn/node/constant.rs#L125) for constant node uses `Initializer::Zeros::init(...)` https://github.com/tracel-ai/burn/blob/d9418adcd4e429c17c8a72d782d0db481f879eb0/crates/burn-core/src/nn/initializer.rs#L79-L85 This...