shapes

The shapes sub-namespace of onnx_optim hosts the per-operator shape-inference functions (ComputeShape*). Each function consumes a ShapesContext (a name → OptimTensor map), a NodeProto and the names of the input values to read, and writes the descriptors of the node’s outputs back into the context.

Concrete functions are organized per operator domain.

• shapes_context.h
• shape_broadcast.h
• shape_check.h
• shape_inference.h
• controlflow
• generator
• logical
• math
• nn
• optional
• preview
• quantization
• reduction
• sequence
• tensor
• text
• traditionalml
• training