rten
rten copied to clipboard
robertknight
ONNX neural network inference engine
Add a naive text-to-phoneme converter based on a dictionary built with the piper-phonemize PyPI package. This converter is naive in the sense that it just ignores words that don't appear...
Hi, I think it would help a lot of we could have a utility for reading images as bytes in the `rten_imageio` crate. Currently, there's only reading from the path...
Unary operators (eg. sigmoid, tanh) are much less efficient with non-contiguous inputs. The problem is two-fold: - For SIMD-vectorized operators (eg. tanh), the fast path calls a SIMD function that...
A common operation in various ONNX operators is to broadcast multiple inputs, of dynamically determined rank, to a common shape, then apply some operation over matching elements from the broadcasted...
This is a tracking issue listing which [ONNX operators](https://onnx.ai/onnx/operators/) are currently supported. Some things to note: - Support for an operator does not mean that all attributes or data types...
Some ONNX operators are only available if certain rten crate features are enabled. If rten is compiled with these features turned off and a user tries to run a model...
The current APIs for running models in RTen are annoying to use if a model has many inputs or outputs. The current APIs require a list of `(NodeId, Input)` tuples...
To make using this library easier in downstream projects, it would help to have more documentation / examples etc. around converting to/from tensors/arrays in other libraries. In the meantime, here...
ONNX models currently have to be converted into the FlatBuffers-based `.rten` format to use them. The `.rten` format is intended to support efficient loading, and have a small code footprint...
Some packages (eg. rten-tensor) can be run under Miri to check for bugs. There are tests in the main `rten` crate which run much too slowly under Miri however. Instead...