Each export now writes a small <model>.ts.pt.meta.json so the native side
(tp_dnn.cpp / CuasDnnLocal, pieces 3-4) reads the model's params from a bundled
sidecar instead of hardcoding them — single source of truth.
  L1: {N, P, vr, out_ch, velocity_mode, vmax}
  L2: {ch_hidden, vmax, grid, ch_in}
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Piece 1 of the native-JNA DNN path (no Python server). Adds export_l2_torchscript.py:
wraps a trained Layer2Net's cell+head into a single-step L2Step module
forward(x,h)->(h_new,det,vel) — exactly infer_server's per-scene recurrence
(h=cell(x,h); det,vel=decode(h)) — so the C++ side just carries h and calls it
per scene. Size-agnostic (circular pad + 1x1 head), runs on the full field.

Validated: scripted==eager exact (0.0); C++ LibTorch (libtorch_probe/l2_probe)
loads it on Blackwell CUDA and replays the recurrence with hidden-state match
9.5e-7. Required disabling the TorchScript JIT fuser (nvrtc element-wise fusion
fails on Blackwell -arch; production wants no runtime nvrtc) — folds into the
native lib startup in piece 2.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

infer_server decode batched GPU_CHUNK scenes (p=[chunk,121,H,W]); 16 -> ~2.5GB -> OOM on the 16GB 5060 Ti
(shared with Eyesis JCuda). Make GPU_CHUNK env-tunable (default 16 for big GPUs); run_infer_local.sh sets
GPU_CHUNK=4 (+ expandable_segments) so local L1+L2 fits. By Claude 06/27/2026.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Run the PyTorch L1+L2 server on the workstation 5060 Ti; Java pipeline points at 127.0.0.1:5577.
Verified: server loads L1(weighted9_pm_s)+L2(l2_v1) on CUDA, warm-up + pyramid build OK.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Per Andrey: future DNN companions to imagej-elphel (all PyTorch, not Java) live here too. README + per-file
header tagline generalized accordingly.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Validated in PoC: L1 (weighted9_pm_s) -> TorchScript -> C++/CUDA on Blackwell matches PyTorch (7.6e-4).
Writes raw-f32 reference vectors for the native probe.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

L1 RawFCN + L2 ConvGRU(torus), synthetic data gen, training/eval, infer_server,
and export_torchscript.py (self-contained TorchScript for native LibTorch inference).
GPLv3 (Elphel norm); headers on all .py/.sh; LICENSE = GPLv3. runs/ checkpoints untracked.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>