monkey-w1n5t0n
9490e20a7a
feat(ml)!: P3 core — geometric dislike in nisps/, jolt/OU + geo ABI
...
Geometric dislike (rl-feedback-design §2.1/§4; upstream InterfaceRL @
0a541cc ported verbatim, constants included):
- nisps/ml/replay.hpp: ReplayView over storage-owned buffers — deepen-or-
store negatives (dedup 0.05, clamp -16), k-NN positive centroid with
deterministic index tie-break + fixed accumulation order, proportional
decay (0.0025*max(|r|,1)) + eviction, order-preserving compaction.
- nisps/ml/geo_push.hpp: push-away target (pushStep clamp(|avgNeg|,.25,1)
*0.5, taper /(1+len), useRandom on len<=1e-4 via nisps::Rng — the single
deliberate divergence from libc rand()), negLRRatio 0.5-0.4*negFraction.
- mlp.hpp: train_targets(input, computed-target, lr, out_mask) — trains
toward computed targets (negative lr = cold-start train-away); solo/
focus gating zeroes masked derivs.
- feedback.hpp: AvoidStyle {Geometric (new default), Diffuse (legacy
move_weights, kept for A/B)}; dislike_geometric() collapses upstream's
press+optimise into one synchronous call; on_up in geometric Avoid
feeds the positive centroid; dislike-multiplier bookkeeping. Storage
gains replay buffers (Fixed: ReplayCap=32 firmware default ≈ +8KB SRAM;
Dynamic arena: cap 64).
- bindings: nisps_ml_feedback_{dislike_geometric,store_positive,
positive_count,negative_count,set_avoid_style} + P3.2 jolt/OU ABI:
nisps_ml_jolt_{press,step,release,active,lr_scale,tick_lr_ramp},
nisps_ml_explore_{intensity,get_intensity,apply} (OUNoise<4096>
over-provisioned; same code the firmware ModeBase runs).
- parity v4: Stage 6 scripted geometric session (2 likes → 2 dislikes,
f32-exact heard vectors via Math.fround) — 961 floats PASS at 2.4e-7.
- tests: test_mlp_geo_dislike.cpp (replay dedup/deepen/clamp, centroid
tie-break, push direction/taper/mask/clamp, cold-start inertness +
train-away, determinism, Diffuse legacy); legacy Avoid test pinned to
Diffuse per the ADR's deliberate-break note.
Firmware: PAFSynth .text/.data unchanged (geometric path not referenced
by current glue). NOTE: discovered pre-existing bug 10c3e55c — the
explore/place wiring is linker-GC'd out of the PAFSynth ELF (predates
this refactor; evidence in the ergo task).
2026-07-14 04:16:21 +02:00
monkey-w1n5t0n
8a19e5b52c
refactor(ml)!: P2.1 storage-policy split — MLPCore<Storage>, fixed + dynamic models
...
Algorithms (forward, backprop/SGD, init, move_weights, diagnostics) now live
once in MLPCore<Storage> (nisps/ml/mlp.hpp). Storage models:
- FixedStorage (storage.hpp): template-sized std::array, zero heap. The
classic MLP<NIn,H1,H2,H3,NOut,...> is an alias preserving kInput/kHidden*/
kOutput/kNumLayers/weight_count() constexpr — firmware + bindings + modes
compile unchanged.
- DynamicStorage (dynamic_storage.hpp): runtime dims, ONE arena allocation
at construction, nothing per-call. #error under NISPS_TARGET_EMBEDDED
(new macro in core/perf.hpp); sole lint-cpp.sh heap-allowlist entry, plus
a lint check that fails if the #error guard disappears.
Verification:
- new ctest test_mlp_storage_parity: fixed↔dynamic BIT-identical across
init/draw/inference/train(FIFO)/move_weights(pin mask)/eval_loss/
layer_stats/set_weights/infer_batch/reset; invalid+moved-from inert
- golden ML vectors (pre-refactor constants) pass → bit-stable refactor
- native↔WASM parity PASS, max delta unchanged (2.4e-7)
- chokepoint B compile: PAFSynth .text 122324→122692 (+0.30%, ±1% budget);
RAM +416B (eval scratch)
- fix: firmware-common.sh used bare 'python' (absent here) → ${PYTHON:-python3}
Part of one-core-engine-refactor P2. nisps_ml_create ABI untouched (P2.2 is
an operator stop-point).
2026-07-13 23:47:03 +02:00
w1n5t0n
825ed6ad33
feat(nisps/ml): MLP library with fixed-architecture template + spread-aware RL (meml-wmh)
...
Stream 2 of the clean-slate rewrite: nisps/ml/ replaces src/memlp/ with a
header-only, heap-free MLP that satisfies nisps::core::MLEngine.
Files (nisps/ml/):
- activations.hpp — ReLU (leaky 0.01 for parity), sigmoid, tanh
- loss.hpp — MSE per-sample (fixes meml-ues double-scaling: returns the
sample's MSE without an extra 1/N multiplication; the training loop
averages explicitly)
- init.hpp — uniform/Xavier/spread-aware weight init
- training.hpp — gradient clip helper (±10.0 matches legacy)
- rl.hpp — move_weights with per-layer Xavier scaling, weight decay
(10% * spread), gaussian noise via the deterministic Rng (matches the
legacy JS sum-of-three-uniforms shape); draw_weights also spread-aware
- stats.hpp — per-layer mean/max/dead/saturating diagnostics
- mlp.hpp — 4-layer (3 hidden + sigmoid output) MLP class with
std::array-backed weights, biases, gradient accumulators, dataset
ring buffer (default 128 examples), loss history (default 4096 iters).
Bias is a separate per-layer parameter — no input-vector mutation.
Flat get_weights/set_weights layout: weights all layers (row-major,
layer order), then biases all layers.
Tests (tests/cpp/, all 50 passing under -Wall -Wextra -Werror -Wpedantic):
- test_mlp_init.cpp — deterministic seeding, spread regimes,
static_assert MLEngine concept satisfied
- test_mlp_inference.cpp — golden hand-computed forward pass match,
sigmoid output range, set_input bounds
- test_mlp_training.cpp — XOR convergence (loss < 0.01 in <2k iters),
ring-buffer eviction
- test_mlp_loss.cpp — meml-ues regression test: reported loss equals
hand-computed average MSE without extra 1/N scaling; sample weights
honoured
- test_mlp_rl.cpp — move_weights respects output_pin_mask (final-layer
rows + biases preserved); spread regimes; grad clear after draw_weights
- test_mlp_serialize.cpp — get_weights/set_weights round-trip preserves
inference exactly; eval_loss is non-mutating; infer_batch matches
individual inference
Verification:
- Clean build, no warnings
- 50 tests pass (22 prior + 28 new)
- No std::vector / new / malloc in nisps/ml/
- All float literals .f-suffixed in code (comments excepted)
2026-04-29 15:55:43 +03:00
