\
- 1.0.0 - 12/10/2022 - First embedded release.
- 1.1.0 - 21/09/2026 - Unified Windows/BF518 source tree, safer desktop CRT, improved allocation/error handling and common test entry.
- 1.2.0 - 21/09/2026 - otFILE is now the single filesystem abstraction on both Windows and BF518; identical dataset/save/load path.
- 1.3.0 - 21/09/2026 - Common application test with native VS2019 main() and otStudio/BF518 setup()/loop() entry points.
- 1.4.0 - 21/09/2026 - Restored separate I/O backends: standard CRT on Windows, otFILE on otStudio/BF518.
- 1.5.0 - 21/09/2026 - Added optional fast sigmoid LUT for embedded targets. BF518 defaults to the LUT implementation; Windows keeps the exact expf() implementation. The LUT spans +/-12 with 1/32 spacing and linear interpolation.
- 1.6.1 - 21/09/2026 - BF518 hot-path placement: forward/inference uses FAST_CODE_1; training/backprop uses FAST_CODE_2. No neural-network mathematics or public API changes.
- 1.7.0 - 21/09/2026 - Added optional Q15 inference engine. Float weights remain the master/training representation; prepareQ15() builds a runtime 16-bit copy and predictQ15() performs fixed-point MACs while retaining float sigmoid/output compatibility.
- 1.8.0 - 21/09/2026 - Added a true fixed-point inference core. prepareQ15() now builds power-of-two scaled weights and Q8 biases; predictQ15Raw() executes Q15 MACs plus integer sigmoid interpolation without floating-point operations. predictQ15() remains as a compatibility wrapper.
- 1.9.0 - 21/09/2026 - Added Q15 runtime memory policies and dual-bank fast-memory placement. AUTO prefers Bank 2 for weights and Bank 1 for scratch buffers, with transparent normal- heap fallback when the selected fast bank cannot fit a buffer. FAST_ONLY and NORMAL modes are also available.
- 1.10.0 - 21/09/2026 - Added stable programmatic Q15 memory reporting through otTinnQ15MemoryInfo/q15MemoryInfo(), including per-bank payload use, remaining fast-heap bytes and buffer placement.
- 1.11.0 - 21/09/2026 - Added exact float-bit training checkpoint model I/O used by resumable training sessions. The historical .tinn text format remains unchanged for normal model exchange.
- 1.12.0 - 22/09/2026 - v0.23 Step 1: topology representation moved into the neural core. Up to four hidden layers can now be described and inspected through a C++98 POD topology. The execution core intentionally still accepts exactly one hidden layer; legacy init()/properties() stay intact.
- 1.13.0 - 22/09/2026 - v0.23 Step 2: generic float forward propagation for one to four hidden layers. The historical one-hidden layout/order remains bit-compatible; training, legacy serialization and Q15 remain intentionally single-layer until their dedicated migration steps.
- 1.14.0 - 22/09/2026 - v0.23 Step 3: generic float back-propagation/training for one to four hidden layers. The validated historical single-hidden bprop remains a dedicated compatibility path; deeper networks use persistent heap delta storage. Q15 and existing checkpoint/file formats remain single-layer.
- 1.15.0 - 22/09/2026 - v0.23 Step 4: exact resumable checkpoints now support one to four hidden layers. OTNNCP1 remains byte-for-byte compatible for single-hidden models; OTNNCP2 stores generic topology plus exact float bits. The loader accepts both formats. Q15 and the historical human-readable .tinn exchange format remain single-layer.
- 1.16.0 - 22/09/2026 - v0.23 Step 5: generic Q15 inference now supports one to four hidden layers. Every connection matrix gets an independent power-of-two weight scale and every non-input layer keeps its own signed-Q8 scalar bias. The historical one-hidden Q15 arithmetic path remains dedicated and bit-identical. FAST/MIXED/NORMAL memory policy now sizes the complete multi-layer activation set.
- 1.17.0 - 22/09/2026 - v0.23 Step 9: knowledge-preserving Candidate initialization. Same-topology retries clone ACTIVE exactly; width-only expansions copy the complete old subnetwork and zero cross-connections from newly added neurons so float and Q15 inference are initially function-preserving. Deeper topologies remain freshly randomized rather than pretending an exact sigmoid identity transform exists.
- 1.18.0 - 22/09/2026 - v0.25 Step 2: structured PRUNE-COPY warm start for same-depth narrower Candidates. Neurons are selected deterministically by outgoing-weight salience and the corresponding rows/columns are copied coherently. This preserves useful trained structure but is deliberately reported as non-function-preserving. Depth reduction remains fresh because sigmoid layer removal has no exact mapping. \
footer
otStudio - Library Reference - (C) 2020-25 Officina Turini, All Rights Reserved
Document built with Doxygen 1.4.0