eval.py

import argparse
import torch
from tqdm import tqdm
import numpy as np
from sklearn.metrics import average_precision_score, roc_auc_score, accuracy_score, precision_recall_curve
import pandas as pd
import os
import json
from datetime import datetime

from model import FusionModel
from utils import seed_run

def convert_numpy_types(obj):
    """Convert NumPy types to Python native types for JSON serialization"""
    if isinstance(obj, (np.integer, np.int64)):
        return int(obj)
    elif isinstance(obj, (np.floating, np.float64)):
        return float(obj)
    elif isinstance(obj, np.ndarray):
        return obj.tolist()
    elif isinstance(obj, dict):
        return {key: convert_numpy_types(value) for key, value in obj.items()}
    elif isinstance(obj, list):
        return [convert_numpy_types(item) for item in obj]
    else:
        return obj

def calculate_acc_at_eer(labels, scores):
    """Calculate accuracy at Equal Error Rate (EER)"""
    from sklearn.metrics import roc_curve
    
    fpr, tpr, thresholds = roc_curve(labels, scores)
    fnr = 1 - tpr
    
    # Find threshold where FPR = FNR (EER)
    eer_threshold_idx = np.nanargmin(np.abs(fpr - fnr))
    eer_threshold = thresholds[eer_threshold_idx]
    
    # Calculate accuracy at EER threshold
    binary_predictions = (scores >= eer_threshold).astype(int)
    acc_at_eer = accuracy_score(labels, binary_predictions)
    
    return acc_at_eer, eer_threshold

def calculate_metrics_for_subset(labels, scores, subset_name):
    """Calculate comprehensive metrics for a subset of data"""
    if len(labels) == 0:
        return None
    
    try:
        auc = roc_auc_score(labels, scores)
        ap = average_precision_score(labels, scores)
        binary_predictions = (scores > 0).astype(int)
        accuracy = accuracy_score(labels, binary_predictions)
        acc_at_eer, eer_threshold = calculate_acc_at_eer(labels, scores)
        
        return {
            'subset': subset_name,
            'samples': len(labels),
            'real_samples': np.sum(labels == 0),
            'fake_samples': np.sum(labels == 1),
            'accuracy': accuracy,
            'auc': auc,
            'average_precision': ap,
            'acc_at_eer': acc_at_eer,
            'eer_threshold': eer_threshold
        }
    except Exception as e:
        print(f"Error calculating metrics for {subset_name}: {e}")
        return None

def extract_method_from_filename(filename):
    """Extract method name from filename"""
    # Extract method name from filename pattern: {id}_Fake_{MethodName}.npz
    if 'Fake_' in filename:
        # Extract method name after 'Fake_'
        method_part = filename.split('Fake_')[1]
        method_name = method_part.replace('.npz', '')
        return method_name
    elif 'Real' in filename:
        return 'real'
    else:
        return 'unknown'

def classify_video_method(video_path):
    """Classify video into diffusion or non-diffusion methods"""
    # Extract filename from path
    filename = os.path.basename(video_path)
    
    # Extract method name from filename
    method_name = extract_method_from_filename(filename)
    
    # Define method categories
    diffusion_methods = ['AniPortrait', 'Ditto', 'Hallo', 'JoyVASA', 'Sonic']
    non_diffusion_methods = ['EDTalk', 'Float', 'SadTalk']
    
    if method_name == 'real':
        return 'real'
    elif method_name in diffusion_methods:
        return 'diffusion'
    elif method_name in non_diffusion_methods:
        return 'non_diffusion'
    else:
        return method_name

def process_video(data, fusion_model, device, invert_score=False):
    visual_tensor = torch.from_numpy(data["visual"]).to(device)
    audio_tensor = torch.from_numpy(data["audio"]).to(device)

    # L2 norm
    visual_tensor = visual_tensor / (torch.linalg.norm(visual_tensor, ord=2, dim=-1, keepdim=True))
    audio_tensor = audio_tensor / (torch.linalg.norm(audio_tensor, ord=2, dim=-1, keepdim=True))

    output = fusion_model(visual_tensor, audio_tensor)
    score = torch.logsumexp(-output, dim=0).detach().cpu().squeeze()

    # 如果训练得到的 ckpt 分数方向与官方相反(我们 train_diffusion_only.py 训出来的
    # 模型,给 fake 打的 logsumexp(-output) 更低、给 real 打的更高,正好和官方反),
    # 加 --invert_score 翻一下符号,让 AUC > 0.5 时反映"正确判别"。
    if invert_score:
        score = -score

    return score

def main(args):
    seed_run()

    print(f"Evaluating AVH-Align on {args.dataset} with pretrained weights saved at {args.checkpoint_path} ...")

    # Init model
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    fusion_model_weights = torch.load(args.checkpoint_path, weights_only=False)

    # Load model
    fusion_model = FusionModel().to(device)

    # Handle multi-GPU saved weights (remove 'module.' prefix)
    state_dict = fusion_model_weights["state_dict"]
    if list(state_dict.keys())[0].startswith('module.'):
        # Remove 'module.' prefix from keys
        new_state_dict = {k.replace('module.', ''): v for k, v in state_dict.items()}
        fusion_model.load_state_dict(new_state_dict)
    else:
        fusion_model.load_state_dict(state_dict)

    fusion_model.eval()

    # Load metadata for access to labels
    metadata = pd.read_csv(args.metadata)

    # Optional: filter metadata to a subset of fake-generation methods (always
    # keeping all real samples). This makes it easy to evaluate, with the same
    # video-level logsumexp aggregation as the full run, on the same test
    # subset that eval_diffusion_only.py uses (e.g. SadTalk,EDTalk,Float).
    methods_filter = None
    if args.methods is not None and args.methods.strip().lower() not in ("", "all"):
        methods_filter = [m.strip() for m in args.methods.split(",") if m.strip()]
        original_n = len(metadata)
        # Build a row-wise mask: keep if real OR if filename matches any method
        def _row_keep(path):
            fname = os.path.basename(path)
            spec = extract_method_from_filename(fname)
            return spec == 'real' or spec in methods_filter
        keep_mask = metadata["path"].map(_row_keep)
        metadata = metadata[keep_mask].reset_index(drop=True)
        print(f"[INFO] --methods filter active: {methods_filter}")
        print(f"[INFO] metadata filtered: {original_n} -> {len(metadata)} (kept all real + only listed fake methods)")

    outputs = []
    ground_truths = []
    video_paths = []
    video_methods = []
    specific_methods = []

    print(f"Processing {len(metadata)} videos...")
    for _, row in tqdm(metadata.iterrows()):
        data = np.load(os.path.join(args.features_path, row["path"].replace(".mp4", ".npz")), allow_pickle=True)
        label = row["label"]
        score = process_video(data, fusion_model, device, invert_score=args.invert_score)
        outputs.append(score)
        ground_truths.append(label)
        video_paths.append(row["path"])

        # Extract method information
        filename = os.path.basename(row["path"])
        specific_method = extract_method_from_filename(filename)
        category_method = classify_video_method(row["path"])

        video_methods.append(category_method)
        specific_methods.append(specific_method)

    outputs = np.array(outputs)
    ground_truths = np.array(ground_truths)

    # Calculate overall metrics
    overall_metrics = calculate_metrics_for_subset(ground_truths, outputs, "Overall")
    
    # Calculate metrics by method type
    # Note: For binary classification metrics (AUC, AP, ACC@EER), each subset must
    # include both real samples and fake samples. So we combine real + each method's fake.
    method_metrics = {}
    
    real_mask = ground_truths == 0
    fake_mask = ground_truths == 1
    
    video_methods_arr = np.array(video_methods)
    specific_methods_arr = np.array(specific_methods)
    
    # Sample distribution stats
    print(f"\n--- Sample Distribution ---")
    print(f"Real samples: {np.sum(real_mask)}")
    unique_methods, method_counts = np.unique(specific_methods_arr[fake_mask], return_counts=True)
    for m, c in zip(unique_methods, method_counts):
        print(f"  {m} fake samples: {c}")
    
    # Diffusion methods (real + all diffusion fakes)
    diffusion_fake_mask = (video_methods_arr == 'diffusion') & fake_mask
    diffusion_subset_mask = real_mask | diffusion_fake_mask
    if np.sum(diffusion_fake_mask) > 0 and np.sum(real_mask) > 0:
        method_metrics['diffusion'] = calculate_metrics_for_subset(
            ground_truths[diffusion_subset_mask], outputs[diffusion_subset_mask],
            "Diffusion Methods (real + diffusion fakes)"
        )
    
    # Non-diffusion methods (real + all non-diffusion fakes)
    non_diffusion_fake_mask = (video_methods_arr == 'non_diffusion') & fake_mask
    non_diffusion_subset_mask = real_mask | non_diffusion_fake_mask
    if np.sum(non_diffusion_fake_mask) > 0 and np.sum(real_mask) > 0:
        method_metrics['non_diffusion'] = calculate_metrics_for_subset(
            ground_truths[non_diffusion_subset_mask], outputs[non_diffusion_subset_mask],
            "Non-Diffusion Methods (real + non-diffusion fakes)"
        )
    
    # Individual methods (all 8 generators): each = real + that method's fakes
    all_specific_methods = sorted(set(specific_methods))
    for method in all_specific_methods:
        if method == 'real' or method == 'unknown':
            continue
        method_fake_mask = (specific_methods_arr == method) & fake_mask
        method_subset_mask = real_mask | method_fake_mask
        if np.sum(method_fake_mask) > 0 and np.sum(real_mask) > 0:
            method_metrics[method] = calculate_metrics_for_subset(
                ground_truths[method_subset_mask], outputs[method_subset_mask],
                f"{method} (real + {method} fakes)"
            )

    # Print detailed results
    print("\n=== Evaluation Results ===")
    print(f"Dataset: {args.dataset}")
    print(f"Total videos: {len(outputs)}")
    
    # Print overall metrics
    if overall_metrics:
        print(f"\n--- Overall Performance ---")
        print(f"ACC: {overall_metrics['accuracy']:.4f}")
        print(f"AUC: {overall_metrics['auc']:.4f}")
        print(f"AP: {overall_metrics['average_precision']:.4f}")
        print(f"ACC@EER: {overall_metrics['acc_at_eer']:.4f}")
        print(f"Real samples: {overall_metrics['real_samples']}, Fake samples: {overall_metrics['fake_samples']}")
    
    # Print method-specific metrics
    print(f"\n--- Performance by Method Type ---")
    for key, metrics in method_metrics.items():
        if metrics:
            print(f"\n{metrics['subset']}:")
            print(f"  Samples: {metrics['samples']}")
            print(f"  ACC: {metrics['accuracy']:.4f}")
            print(f"  AUC: {metrics['auc']:.4f}")
            print(f"  AP: {metrics['average_precision']:.4f}")
            print(f"  ACC@EER: {metrics['acc_at_eer']:.4f}")

    # Create results directory if it doesn't exist
    results_dir = "results"
    os.makedirs(results_dir, exist_ok=True)
    
    # Generate timestamp for unique filenames
    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
    checkpoint_name = os.path.basename(args.checkpoint_path).replace('.pt', '')
    
    # Save detailed predictions to CSV with method classification
    results_df = pd.DataFrame({
        'video_path': video_paths,
        'method_category': video_methods,
        'specific_method': specific_methods,
        'predicted_score': outputs,
        'predicted_label': (outputs > 0).astype(int),
        'ground_truth': ground_truths,
        'correct': ((outputs > 0).astype(int) == ground_truths).astype(int)
    })
    
    predictions_file = os.path.join(results_dir, f"{args.dataset}_{checkpoint_name}_detailed_predictions_{timestamp}.csv")
    results_df.to_csv(predictions_file, index=False)
    print(f"\nDetailed predictions saved to: {predictions_file}")
    
    # Save evaluation summary to JSON (convert NumPy types first)
    eval_summary = {
        'dataset': args.dataset,
        'checkpoint_path': args.checkpoint_path,
        'features_path': args.features_path,
        'metadata': args.metadata,
        'methods_filter': methods_filter,
        'invert_score': bool(args.invert_score),
        'overall_metrics': overall_metrics,
        'method_metrics': method_metrics,
        'total_videos': int(len(outputs)),
        'predictions_file': predictions_file,
        'evaluation_time': datetime.now().isoformat()
    }
    
    # Convert NumPy types to Python native types for JSON serialization
    eval_summary_converted = convert_numpy_types(eval_summary)
    
    summary_file = os.path.join(results_dir, f"{args.dataset}_{checkpoint_name}_detailed_summary_{timestamp}.json")
    with open(summary_file, 'w') as f:
        json.dump(eval_summary_converted, f, indent=2)
    
    print(f"Evaluation summary saved to: {summary_file}")
    
    print("\n=== Final Summary ===")
    print(f"Dataset: {args.dataset}")
    print(f"Total videos: {len(outputs)}")
    if overall_metrics:
        print(f"ACC: {overall_metrics['accuracy']:.4f}")
        print(f"AUC: {overall_metrics['auc']:.4f}")
        print(f"AP: {overall_metrics['average_precision']:.4f}")
        print(f"ACC@EER: {overall_metrics['acc_at_eer']:.4f}")

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Evaluate Fusion Model on Deepfake Dataset")

    parser.add_argument("--checkpoint_path", type=str, default="checkpoints/AVH-Align_AV1M.pt",
                        help="Path to the pretrained fusion model checkpoint.")
    parser.add_argument("--features_path", type=str,
                        default=f"av1m_features/val/",
                        help="Path to the root folder of test data.")
    parser.add_argument("--metadata", type=str,
                        default="av1m_metadata/test_metadata.csv",
                        help="CSV file containing ground truth labels.")
    parser.add_argument("--dataset", type=str, default="AV1M",
                        help="Dataset name")
    parser.add_argument("--methods", type=str, default=None,
                        help="Comma-separated subset of fake-generation methods to evaluate "
                             "(real samples are always kept). Example: 'SadTalk,EDTalk,Float'. "
                             "Default = None (use all methods).")
    parser.add_argument("--invert_score", action="store_true",
                        help="Negate the per-video score before computing metrics. "
                             "Use this for in-house ckpts whose score direction is opposite "
                             "to the official AVH-Align_AV1M.pt convention.")

    args = parser.parse_args()
    main(args)