DeepSharp.Export.Onnx
Writes a DeepSharp network as an ONNX graph that any ONNX runtime answers through: new OnnxExport(each).Write(network, loss, stream). Each step of the network is written as the operators of the oldest operator set that holds them — a dense layer as a MatMul and the Add of its bias, a low-rank pair as the extra products it is, the normalisations, activations, convolutions and joins of a network whose paths meet, an LSTM and a GRU, and attention and the encoder block as the nodes of the oldest sets that hold them — its numbers as the graph's initializers named by the slot's path, and the output activation its loss applies itself as the last node. The graph is the one DeepSharp's ONNX reader reads back into the same numbers. A network written as code has no declared order and is refused by name, and so is every layer that has no ONNX operator to be written as.
Activity
- Latest release
- 3d ago
- Total releases
- 4
- Cadence
- ~2 days
- Last 12 months
- 4
Details
- License
- MIT
- First release
- Oct 21, 2026
| Version | Released | |
|---|---|---|
0.17.0
minor
| ||
0.16.0
minor
| ||
0.15.0
minor
| ||
0.14.0
initial
|