Robert Knight
Robert Knight
It would probably make sense to have a variant that accepts any type implementing `std::io::Read`, which includes `&[u8]`. This would then use [image::io::Reader](https://docs.rs/image/latest/image/io/struct.Reader.html#). In the meantime, I suggest you copy...
I haven't gotten around to this yet. My previous comment stands - copy and paste the `read_image` code from https://github.com/robertknight/rten/blob/665ea51d0ee80fed28fe249d484004e3e9571539/rten-imageio/src/lib.rs#L24 into your project, and change it to use [`image::load_from_memory`](https://docs.rs/image/latest/image/fn.load_from_memory.html) instead...
In the context of Ocrs and its `OcrEngine::prepare_input` API, this problem has been solved by re-designing the API for loading images, which can now take a byte buffer directly. See...
A related issue is the case where a model was loaded with a subset of the normal operators available via `Model::load_with_ops`, and a user tries to load a model which...
Strings that look like ONNX operator names (match `"[A-Z][A-Za-z]+"'`) grepped from `torch/onnx/symbolic_opset*.py`. This gives a rough idea of which ONNX operators models exported from PyTorch might actually use. ``` Abs...
Hello @Caellian, which operators did you run into issues with? Given that you mentioned enabling a feature, I'm guessing it was one of the `Random*` ops? There is [an issue](https://github.com/robertknight/rten/issues/154)...
Had a look at Rust protocol buffers runtimes: - [Prost](https://github.com/tokio-rs/prost) has the widest adoption, but it has downsides: - It allocates `Vecs` for each repeated field, with an exception for...
As an extension of this, it would be useful to be able to pass owned tensors as inputs to graph execution, rather than views, so that their buffers can be...
This was done for sharing between the main graph and subgraphs in https://github.com/robertknight/rten/pull/312. This is simpler because the interpreter loop for a subgraph runs on the same thread as the...
There is no support for quantization yet. This is something I plan to add in future. Currently f32 / i32 are the only data types supported (i64 and bool tensors...