kirad/augmentations.py:190 |
assumes 16 kHz audio; maxF=8000 is its Nyquist |
kirad/base_models.py:93 |
above 10,000 scores, Sn/Sp thresholds use a 10,000-point linspace grid |
kirad/base_models.py:103 |
0% / 40% indet budgets |
kirad/base_models.py:194 |
per-window loss in train; loss of the per-recording mean in val/test |
kirad/base_models.py:391 |
1000 interpolation points |
kirad/base_models.py:849 |
English-only Whisper (DAM 1-3) |
kirad/base_models.py:867 |
1500 encoder positions = 30 s of audio (Whisper default) |
kirad/base_models.py:892 |
default model id |
kirad/base_models.py:922 |
English BERT (DAM 2 text) |
kirad/base_models.py:947 |
default model id |
kirad/base_models.py:950 |
LLaMA truncated to its first 15 layers |
kirad/constants.py:13 |
audio must already be 16 kHz; files are not resampled. |
kirad/constants.py:21 |
VAD silence/speech minimums (ms); VAD itself runs upstream, not in this repo. |
kirad/constants.py:36 |
questionnaire maxima; set the ordinal range of each task. |
kirad/constants.py:40 |
binary cuts for metrics/thresholds; label counts cutoffs exceeded (strict >). |
kirad/constants.py:42 |
CORAL target sets per task; exact28/exact22 use every score boundary. |
kirad/constants.py:141 |
Kintsugi remote storage URIs, blanked to None in this copy. |
kirad/constants.py:147 |
local cache under ~/.kintsugi and ~/.cache/kintsugi. |
kirad/constants.py:170 |
fixed 80-bin log-mel mean profile; WhisperTorchFeatureExtractor shifts each |
kirad/dataset_utils.py:118 |
audio files must be <filename>.wav, transcripts <filename>.json |
kirad/dataset_utils.py:183 |
download via gsutil from a private GCS URI in constants.py |
kirad/dataset_utils.py:490 |
10 random splits tried; the lowest-KL one is kept |
kirad/dataset_utils.py:663 |
Kintsugi CSV schema; missing columns are skipped, new ones are ignored. |
kirad/dataset_utils.py:691 |
columns kept as per-row lists after merging |
kirad/dataset_utils.py:701 |
duration columns summed after merging |
kirad/datasets_and_dataloaders.py:54 |
30 s windows (= Whisper input length); also the default min_duration filter |
kirad/datasets_and_dataloaders.py:56 |
default overlap only for direct calls; datasets pass max_overlap_frac (0.0) |
kirad/datasets_and_dataloaders.py:61 |
task -> CSV label column (depression = PHQ, anxiety = GAD) |
kirad/datasets_and_dataloaders.py:259 |
random_state=1 for the fraction subsample |
kirad/datasets_and_dataloaders.py:379 |
HF feature extractor name; mar_2024/config.yaml sets false (log-mel on GPU) |
kirad/datasets_and_dataloaders.py:485 |
only channel 0 is used (stereo files lose the other channel) |
kirad/datasets_and_dataloaders.py:545 |
tokenizer name (DAM 2 text branch); a W&B artifact URL is also accepted |
kirad/feature_extraction_whisper.py:83 |
16 kHz input, mel filters up to 8 kHz (Whisper setup) |
kirad/feature_extraction_whisper.py:90 |
fixed per-bin target energies from kirad/constants.py (80 values) |
kirad/feature_extraction_whisper.py:132 |
Whisper's 8 (log10) = 80 dB dynamic range and (x + 4) / 4 scaling |
kirad/launcher_utils.py:119 |
default seed |
kirad/launcher_utils.py:120 |
default patience (validations) |
kirad/launcher_utils.py:328 |
CuBLAS workspace setting required for deterministic GPU ops. |
kirad/plotting.py:73 |
ROC budgets 0%..70%, step 10% |
kirad/utils.py:206 |
URL layout host/entity/project/artifacts/type/name[/tag]; tag defaults to "best" |
kirad/utils.py:340 |
looks for baseline_const_0..3 only (tasks with up to 4 classes) |
kirad/utils.py:478 |
test scores are logged once, as version v0 |
kirad/utils.py:588 |
budgets 0%..90% |
kirad/utils.py:600 |
task names; edit for other outcomes |
scripts/launcher.py:88 |
DDP with find_unused_parameters=True; fp16 on GPU, fp32 on CPU. |
scripts/launcher.py:171 |
W&B entity/project are None, so every run raises until you set them. |
scripts/launcher.py:202 |
checkpoint root; /mnt/chkpts when JOB_NAME (Kubernetes) is set. |
scripts/launcher.py:254 |
"kipy" is a private Kintsugi package; remove it if unavailable. |
scripts/launcher.py:448 |
Kubernetes namespace "default". |
scripts/launcher.py:476 |
DDP master port 12355. |
scripts/run_whisper_asr.py:35 |
ASR model "large-v3"; reloaded on every call. |
scripts/run_whisper_asr.py:65 |
English only; change for Congo/Colombia/Jordan/Ukraine data. |
scripts/run_whisper_asr.py:70 |
decoding params (beam 5, best_of 5, temperature fallback). |
research/stable/ordinal_regression_mtl_aug_2024/model.py:182 |
default indeterminate budgets 0% and 40% |
research/stable/ordinal_regression_mtl_jun_2024/model.py:42 |
default indeterminate budgets 0% and 40% |
research/stable/ordinal_regression_mtl_mar_2024/model.py:56 |
default indeterminate budgets 0% and 40% |
research/stable/ordinal_regression_mtl_mar_2024/model.py:70 |
1500 encoder positions = 30 s windows (Whisper default) |
research/stable/ordinal_regression_mtl_mar_2024/model.py:73 |
320 samples per encoder position (hop 160 x conv stride 2 at 16 kHz); |