scripts/BATCH_TESTING_README.md

Batch Testing Script Documentation

Overview

The batch_test_all.sh script automates the testing of all trained models (CTA, PSM, RADR) on all available test datasets (ours, THB, FF++, HDTF). It tests the top-3 checkpoints (based on validation AUC) and the last checkpoint for each method.

Quick Start

# Run the full batch test
bash scripts/batch_test_all.sh

# Check the configuration first
bash scripts/test_batch_script.sh

What It Tests

Methods

• CTA (Cross-Temporal Attention)
• PSM (Pairwise Similarity Matching)
• RADR (Random Adversarial Regularization)

Datasets

• ours: Original FairTalking test set
• thb: TalkingHeadBench test set
• ff++: FaceForensics++ test set
• hdtf: HDTF-paired test set

Checkpoints per Method

• Top-3 checkpoints (highest validation AUC)
• Last checkpoint (final training state)

Total tests: 3 methods × 4 datasets × 4 checkpoints = 48 test combinations

Output Format

The script generates a CSV file with the following columns:

method,checkpoint,dataset,test_acc,test_auc,timestamp
psm,epoch24-valauc0.9963.ckpt,ours,0.9553,0.9935,2026-04-30_11:20:28
psm,epoch24-valauc0.9963.ckpt,thb,0.8721,0.9456,2026-04-30_11:21:15
...

File Locations

Input Checkpoints

• outputs/cta/checkpoints/ - CTA model checkpoints
• outputs/psm/checkpoints/ - PSM model checkpoints
• outputs/radr/checkpoints/ - RADR model checkpoints

Test Scripts

• scripts/test_{method}.sh - Ours dataset testing
• scripts/test_{method}_thb.sh - THB dataset testing
• scripts/test_ffpp.sh - FF++ dataset testing
• scripts/test_hdtf_paired.sh - HDTF dataset testing

Output Files

• batch_test_results_YYYYMMDD_HHMMSS.csv - Main results file
• outputs/{method}/test_predictions_*.csv - Per-sample predictions

Example Usage

# Test a single method on a single dataset
bash scripts/test_psm.sh outputs/psm/checkpoints/last.ckpt

# Test all methods and datasets (full batch)
bash scripts/batch_test_all.sh

# Check results
cat batch_test_results_20260430_112000.csv | column -t -s,

Metrics Explanation

• test/acc: Test accuracy (binary classification accuracy)
• test/auc: Test AUC (Area Under ROC Curve)

Higher values indicate better performance. AUC is generally more informative for imbalanced datasets.

Troubleshooting

Common Issues

1. Script not found: Check that all test scripts exist in the scripts/ directory
2. Checkpoint not found: Ensure models have been trained and checkpoints exist
3. Metrics not extracted: The script uses regex to extract metrics from stdout

Manual Testing

If the batch script fails, you can test individual combinations manually:

# Test PSM on THB dataset
bash scripts/test_psm_thb.sh outputs/psm/checkpoints/last.ckpt

# Test CTA on FF++ dataset
bash scripts/test_ffpp.sh cta outputs/cta/checkpoints/last.ckpt

Performance Notes

• The full batch test takes approximately 30-60 minutes to complete
• Each individual test takes 1-3 minutes depending on dataset size
• Results are saved incrementally, so you can interrupt and resume
• The script uses DDP (Distributed Data Parallel) for faster inference

File Name Fix

The script automatically handles the checkpoint filename issue where val_auc was showing as 0.0000 due to template mismatch. New checkpoints will have correct AUC values in their filenames.

Best Model Selection

The "best" model for each method is the checkpoint with the highest validation AUC among the top-3 saved checkpoints. This is automatically selected by PyTorch Lightning's ModelCheckpoint callback.