scripts/test_ffpp.sh

#!/usr/bin/env bash
# Run test-only inference on the FaceForensics++ dataset.
#
# Usage:
#   bash scripts/test_ffpp.sh <method> [checkpoint_path] [extra hydra overrides...]
#
# Examples:
#   bash scripts/test_ffpp.sh cta
#   bash scripts/test_ffpp.sh psm outputs/psm/checkpoints/last.ckpt
#   bash scripts/test_ffpp.sh radr /path/to/custom.ckpt backbone=timesformer
#
# Notes:
#   * Uses the standalone FFPPTestDataset (configs/data/fairtalking_ffpp.yaml),
#     which scans <FFPP_ROOT>/manipulated_sequences/<gen>/c23/videos/*.mp4
#     directly. No symlink tree is needed.
#   * Real videos under <FFPP_ROOT>/original_sequences/youtube/c23/videos
#     are picked up automatically if present (for AUC).
set -eo pipefail
cd "$(dirname "$0")/.."

# load .env if present
[ -f .env ] && set -a && . ./.env && set +a

if [ $# -lt 1 ]; then
    echo "Usage: $0 <method> [checkpoint_path] [extra hydra overrides...]"
    echo "Available methods: cta, psm, radr"
    exit 1
fi

METHOD="$1"
shift
CKPT="${1:-outputs/${METHOD}/checkpoints/last.ckpt}"
if [ $# -ge 1 ]; then shift; fi

case "$METHOD" in
    cta|psm|radr) ;;
    *)
        echo "Error: Invalid method '$METHOD'. Available: cta, psm, radr"
        exit 1
        ;;
esac

if [ ! -f "$CKPT" ]; then
    echo "[test_ffpp] checkpoint not found: $CKPT"
    echo "[test_ffpp] Please train the model first or provide a valid checkpoint path."
    exit 1
fi

FFPP_ROOT_RESOLVED="${FFPP_ROOT:-/apdcephfs_gy4/share_303628665/joywu/dataset/FF++}"

echo "[test_ffpp] Method:      $METHOD"
echo "[test_ffpp] Dataset:     fairtalking_ffpp (FaceForensics++)"
echo "[test_ffpp] Checkpoint:  $CKPT"
echo "[test_ffpp] FFPP_ROOT:   $FFPP_ROOT_RESOLVED"

python3 src/train.py \
    method="$METHOD" \
    data=fairtalking_ffpp \
    trainer=ddp \
    backbone=timesformer \
    +test_only=true \
    +test_ckpt="$CKPT" \
    "$@"