Guillaume Lagrange
Guillaume Lagrange
Currently we only have an example using the `MNISTDataset` via the huggingface loader. Would be nice to include an example to load a custom image dataset from disk, similar to...
While working on the Llama-3 implementation I stumbled upon a memory issue when importing pytorch weights with `PyTorchFileRecorder`. When I profiled the memory usage for ResNet-152 (checkpoint is 252MB on...
Implement cumulative sum across an input dimension. Example: ```rust let x = Tensor::arange(0..6 as i64, &device); x.cumsum(0) // [ 0, 1, 3, 6, 10, 15] ```
### Related Issues/PRs Closes #1521 ### Changes Added ResNet-50 inference in backend-comparison - New build script to fetch the [resnet-burn](https://github.com/tracel-ai/models/tree/main/resnet-burn) implementation and apply a diff patch to be compatible with...
Follow-up to PR #1488. Related to #1490 . The default implementation loops could be parallelized to improve performance on backends that don't directly implement the operations. *Note: we have run_par,...
As a follow-up to the ResNet tutorial, we discussed adding a fine-tuning example. I suggest a multi-label classification since this is not addressed yet (two birds, one stone).
### Feature description Default implementation in #1488 requires `async` and isn't optimal. Would be great to remove the `async` methods with a wgpu implementation.
Current implementation supports `Tensor`, but there are other use-cases such as multi-label classification that require support for inputs (and multi-encoded targets) of type `Tensor`.
### Feature description Add common linear algebra operations. ### Feature motivation We've had a couple of users ask about support for linalg ops like SVD, QR decomposition and matrix exponential....
### Feature description Right now we support LR schedulers but only provide a constant schedule or the Noam schedule. Would be nice if we added a couple of common/popular LR...