Inference¶

Basic usage¶

python scripts/predict.py \
    --config config/predict_ast.yaml \
    --data-root $SCRATCH/Audio-XAI/audio_xai/data/external \
    --checkpoint $SCRATCH/Audio-XAI/runs/ast/version_5/checkpoints/epoch=4-step=9740.ckpt

Models¶

AST¶

python scripts/predict.py \
    --model-type ast \
    --checkpoint runs/ast/version_5/checkpoints/epoch=4-step=9740.ckpt \
    --data-root /path/to/data --split test

VGGish¶

python scripts/predict.py \
    --model-type vggish \
    --checkpoint runs/vggish/version_3/checkpoints/epoch-epoch=009.ckpt \
    --data-root /path/to/data --split test

Sonics (HuggingFace)¶

python scripts/predict.py \
    --model-type sonics \
    --model-id awsaf49/sonics-spectttra-gamma-5s \
    --data-root /path/to/data --split test

Key flags¶

Flag	Default	Description
`--model-type`	—	`ast`, `vggish`, or `sonics`
`--model-id`	—	HuggingFace repo (Sonics only)
`--checkpoint`	—	Path to `.ckpt` file (AST / VGGish)
`--data-root`	—	Dataset root directory
`--split`	`test`	`train`, `valid`, or `test`
`--n-samples`	`None`	Cap to N balanced samples (None = full split)
`--clip-seconds`	`5.0`	Clip duration
`--batch-size`	`16`	Inference batch size
`--log-dir`	`runs/predict`	TensorBoard output directory
`--device`	`cuda`	`cuda` or `cpu`

Output¶

Results are saved to --log-dir:

TensorBoard scalars: accuracy, AUROC, F1, EER (per-sample and aggregate)
CSV: results.csv with per-sample predictions, probabilities, and true labels

Metrics computed¶

Metric	Description
Accuracy	Fraction of correct predictions
AUROC	Area under ROC curve
F1 (macro)	Macro-averaged F1 score
EER	Equal Error Rate — the standard deepfake metric
Sensitivity	True positive rate (fake detected as fake)
Specificity	True negative rate (real detected as real)

SLURM¶

sbatch sbatch/predict/ast.sbatch
sbatch sbatch/predict/vggish.sbatch
sbatch sbatch/predict/spectttra.sbatch

Override the checkpoint or split with environment variables before submitting:

AST_CKPT=/path/to/other.ckpt sbatch sbatch/predict/ast.sbatch
DEVICE=cpu sbatch sbatch/predict/ast.sbatch