Neural-network estimation of ion-channel gating rates (plus baseline, amplitude and noise level) from single-channel current recordings, trained online on simulated traces.
| Path | What it holds |
|---|---|
scripts/simulation/ |
Python wrapper (simulator.py) and C API (simdwell_capi.cpp) around the 3D_Fit_unclamped simulator; build.sh compiles libsimdwell.so |
scripts/hist_pipeline/ |
dwell-time histogram networks trained online: train_hist_online.py, the self-calibrating train_selfcal_online.py, GPU histogram / noise / trace synthesis code |
scripts/refine_pipeline/ |
RAFT-style refinement by re-simulation on top of a self-calibrating stage 1 (train_refine_online.py), inference on recordings, diagnostics and comparison plots |
scripts/softdwell_pipeline/ |
SoftDwell network and its online trainer |
scripts/plotting/ |
result and comparison plots |
configs/sim/ |
simulator configs (COC, COCO, OCO; noise-file sets) |
*.sh (top level) |
Slurm job scripts for the NHR@FAU Alex cluster |
Results, checkpoints, noise files and the container image (pytorch.sif) are not in the repository.
How the pipelines developed, how to re-run each step and which branch is for what: docs/ROUTE.md.
The simulator core is a separate repository and is not copied here. Clone it into external/:
git clone -b 3DFit_unclamped_light git@github.com:vatsimanos/3D_Fit_unclamped.git external/3D_Fit_unclamped
git -C external/3D_Fit_unclamped checkout 8a6ec71 # the commit the current results were made withscripts/simulation/build.sh builds libsimdwell.so from it (set SIM_ROOT to use another location); the job
scripts build it on the compute node.
Every job script runs inside the Apptainer image pytorch.sif (definition: hohd_pytorch.def), e.g.
sbatch sim_selfcal_online.sh --sim-config configs/sim/COC_noise5.json --output-root outputs/<name>
sbatch sim_refine_online.sh --stage1-run <selfcal run>/seed1 --output-root outputs/refine_online/<name>