alnn_rewrite/config.toml

# Pipline Scenario to run
# "scen_train_all" - Train, evaluate (test+val), noise, combine, analyze
# "scen_load_all"  - Load, evaluate (test+val), noise, combine, analyze
# "scen_load_eval" - Load, evaluate (test+val), combine, analyze (no noise)
# "scen_analyze"   - Analyze existing evaluation files only (no training/eval)
# "scen_kfold"     - Patient-grouped k-fold CV: train+eval per fold, then OOF + analyze
scenario = "scen_kfold"


[run]
# Script/automation options. Leave both false for normal interactive use.
autorun = false          # start the pipeline automatically once the config loads
autoquit = false         # exit the app when the pipeline finishes (0/1 return code)
# In autorun, allow overwriting PROTECTED outputs (models/evaluations) without a
# prompt. Not needed for "scen_analyze", which only writes regeneratable analysis
# artifacts and therefore never triggers the overwrite guard.
force_overwrite = false

[data]
mri_files_path = "../data/PET_volumes_customtemplate_float32/"
xls_file_path = "../data/LP_ADNIMERGE.csv"
seed = 42
data_splits = [0.7, 0.2, 0.1] # train, validation, test
image_channels = 1
clin_data_channels = 2
num_classes = 2 # AD, NL


[training]
device = "cuda:0" # "cpu", "cuda", "mps"
batch_size = 32
ensemble_size = 30
droprate = 0.05
learning_rate = 0.0001
num_epochs = 25
deterministic = false # force deterministic cuDNN kernels (slower, exact reproduction)


[evaluation]
# Gaussian noise applied to images during noisy evaluation (step 6).
#
# With noise_relative = true (below) each sigma is a FRACTION OF THAT IMAGE'S own
# intensity standard deviation: sigma = 1.0 means the added noise is as strong as
# the image's signal variation (SNR ~ 1), and sigma = 8.0 is essentially pure
# noise. This scaling is required here because the PET volumes are unnormalised
# (std in the thousands) -- an absolute sigma of 0.2 perturbs them by ~0.002% and
# does nothing measurable.
#
# The grid is dense below 1.0 (where degradation begins) and spreads out above it
# so the curve is followed all the way into the unusable regime, giving the fit
# both plateaus (clean accuracy and chance level) plus the transition between.
# 0.0 is the clean baseline and must stay first; values must be unique.
noise_levels = [0.0, 0.05, 0.1, 0.2, 0.3, 0.5, 0.7, 1.0, 1.5, 2.0, 3.0, 5.0, 8.0]
# false => sigma is an absolute value in raw voxel units (not recommended here).
noise_relative = true
# Monte-Carlo forward passes used to estimate Bayesian predictive/model
# uncertainty (step 5/6). Ignored for the deterministic ensemble.
mc_passes = 30
# Splits pooled into the combined_*.nc files that analysis prefers. The splits
# are patient-disjoint, so pooling just enlarges the evaluation set (tighter
# accuracy/uncertainty estimates). Set to ["test"] for a test-only headline
# number -- note val is not a pure held-out set if it ever informed tuning.
combine_splits = ["test", "val"]


[kfold]
# Used only by the "scen_kfold" scenario. Each fold trains a full ensemble on
# ~(k-1)/k of the patients and evaluates on the held-out 1/k; out-of-fold results
# are pooled to cover every sample. NOTE: cost is ~k * a normal train+eval run.
k_folds = 5          # number of patient-grouped folds (each ~1/k held out)
val_fraction = 0.15  # fraction of each fold's non-test patients used for validation
include_noise = false # also run the noise sweep within each fold (much slower)