Crate kn_graph

A neural network inference graph intermediate representation, with surrounding utilities.

The core type of this crate is Graph, see its documentation for how to manually build and compose graphs.

An example demonstrating some of the features of this crate:

// load an onnx file
let graph = load_graph_from_onnx_path("test.onnx", false)?;

// optimize the graph
let graph = optimize_graph(&graph, Default::default());

// render the graph as an svg file
graph_to_svg("test.svg", &graph, false, false)?;

// execute the graph on the CPU
let batch_size = 8;
let inputs = [DTensor::F32(Tensor::zeros(IxDyn(&[batch_size, 16])))];
let outputs = cpu_eval_graph(&graph, batch_size, &inputs);

This crate is part of the Kyanite project, see its readme for more information.

Re-exports

pub use ndarray;

Modules

cpu

CPU graph execution.

dot

Graph visualization as a dot or svg file.

dtype

The DType enum.

graph

The core graph datastructure.

onnx

Onnx file loading.

optimizer

Graph optimization.

shape

The Shape type and utilities.

visualize

Hidden activations visualization.

wrap_debug

Macros

dispatch_dtensor

dispatch_dtensor_pair

dispatch_dtype

map_dscalar_pair

map_dtensor

map_dtensor_pair

shape