Expand description
Core tensor operations, static shape checking, and autograd for Incin.
Re-exports§
pub use typenum;
Modules§
- advanced
- Structural shape cursors retained for low-level and legacy integrations. Ordinary application code should prefer signed numeric axis selectors.
- autograd
- Backend-owned gradient handles produced by autograd passes.
- backend_
authoring - Extension traits and operation descriptor contracts for backend authors.
- compiled
- Experimental compiled graph representation, IR capture, and execution planning.
- dist
- Placement and distributed-execution contracts.
- distributions
- Probability distribution sampling with custom distribution trait support.
- error
- Structured errors returned by core and its extension crates.
- exec
- Proof-carrying operation lowering.
- experimental
- Unstable APIs that carry no compatibility guarantee.
- graph
- Capture-level graph representation shared by import and compiled previews.
- io
- Bounded model-file readers (GGUF metadata, format inspection).
- loss
- Loss functions and reduction definitions.
- metrics
- Evaluation metrics module.
- nn
- Neural network layers, modules, and utilities.
- onnx
- ONNX import and export helpers for graph interchange.
- optim
- Generic parameter optimization: update-rule traits, gradient clipping, and the concrete optimizers built on them.
- prelude
- Core prelude re-exporting common types, neural network modules, shapes, and backend traits.
- resource
- Resource limits and bounded I/O checks for model and data parsers (
SEC-007). - serialization
- Named serialization surface for applications that do not use the prelude.
- shapes
- Type-level shape definitions, dimension arithmetic, and shape verification traits.
- tensor
- The tensor frontend: construction, operations, and device plumbing.
- types
- Expert type-level shape vocabulary kept outside the normal prelude.
Macros§
- SeqTy
- Names the nested
Sequentialtype thatseq!would build a value of, from the same flat list of layer types. - axis
- Build an arbitrary static or runtime axis selector.
- best_
device - Names the best device this build can target.
- dim
- Generates zero-sized semantic tags for named tensor dimensions.
Runtime extents remain in
ShapeBuf, while the generated tags provide stable schema identity for compile-time metadata and lookup. - paranoid_
audit - Assert a descriptor’s invariants when
paranoid-validationis on. - seq
- A macro to easily build Sequential models with many layers.
seq!(L1, L2, L3)expands toSequential(L1, Sequential(L2, L3)).