OnnxSharp 0.3.2

Large Language Model (LLM) generator. And Small as well, SML. Give data and generate your model. Fuuga implements a complete language model pipeline: tokenization, data ingestion, model training, fine-tuning, and text generation. Code in .NET (F#).

Large Language Model (LLM) generator. And Small as well, SML. Give data and generate your model. Fuuga implements a complete language model pipeline: tokenization, data ingestion, model training, fine-tuning, and text generation. Code in .NET (F#).

Large Language Model (LLM) generator. And Small as well, SML. Give data and generate your model. Fuuga implements a complete language model pipeline: tokenization, data ingestion, model training, fine-tuning, and text generation. Code in .NET (F#).

Reads an ONNX graph into a DeepSharp network: the one PyTorch's torch.onnx.export writes, by either of its exporters, the one Keras writes with model.export(format="onnx"), or the one tf2onnx converts a TensorFlow SavedModel into. A stack of dense layers — a Gemm, or a MatMul and the Add of its bias — convolutions and poolings along a series, over an image or through a volume, batch, layer, group, instance and root-mean-square normalisations, flattens and activations — the GELU, SiLU, mish, leaky, parametric and self-normalising ones and softmax among them, as the operators of the newer sets or as the several operators PyTorch's exporters spell them in for older ones — is lowered onto DeepSharp's layers as DeepSharp's words build them, the loss it answers through handed in. The numbers go in turned as each node declares them — a weight matrix outputs by inputs where Gemm says so, a kernel channels first, a series', an image's or a volume's values flattened channel by channel — and series, images and volumes go in and come out with their channels last, as TensorFlow's already are; they are widened from half precision, rounded from double, read from the file the exporter keeps beside the graph, strictly, every fault named at once where the graph holds it. Where paths meet — a value added back after the layers that worked on it, laid beside another along the last axis, or cut apart by columns — the graph is read as steps that take the steps that made their values. A series read by an LSTM or a GRU node, in one direction or both, to its last state or at every step, becomes the recurrent layer of the same name, and the Attention operator of operator set 23 over the steps a graph's products made becomes the layer that looks across them. Any node no network here is built of — a stride of its own for each axis, a dilated window, a pooling that rounds up, channels in groups — is refused at its node.

A C# ONNX backend implementation.