EricLouchez.Shorokoo.Core 0.1.7-dev

Shorokoo

Define, train, and run neural networks in pure C# — no Python required.

• Define models as C# classes — strongly typed Tensor<float32>, Scalar<int64>, shapes checked as you build.
• Train them — reverse-mode autodiff, optimizers, learning-rate schedules, checkpointing.
• Run them fast — execution backed by ONNX Runtime on CPU or GPU.
• Interoperate — export models as .onnx, load pretrained weights from .safetensors.

Getting started

Install this package plus one backend for your platform, and (recommended) the source generator for the [Module] syntax:

dotnet add package Shorokoo
dotnet add package Shorokoo.LinuxCPU # or Shorokoo.LinuxGPU / Shorokoo.WinCPU / Shorokoo.WinGPU
dotnet add package Shorokoo.CodeGen

For ready-made layers, losses, and optimizers also add:

dotnet add package Shorokoo.Modules

using Shorokoo;
using Shorokoo.Modules;
using static Shorokoo.Globals;
using static Shorokoo.NN;

[Module]
public partial class Dense
{
 public static Tensor<float32> Inline(
 [Hyper] Scalar<int64> outFeatures,
 Tensor<float32> x)
 {
 // ... build the layer from tensor ops ...
 }
}

Documentation

Guides, API reference, and samples: https://github.com/Eric-Louchez/Shorokoo