validate_model.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-
"""
简化版模型验证脚本
"""
import torch
from models import resnet_face18
import numpy as np
import os
import cv2
from config.yaml_config import Config
import sys

def load_image(img_path):
    """加载LFW图片"""
    image = cv2.imread(img_path, 0)
    if image is None:
        return None
    image = np.dstack((image, np.fliplr(image)))
    image = image.transpose((2, 0, 1))
    image = image[:, np.newaxis, :, :]
    image = image.astype(np.float32, copy=False)
    image -= 127.5
    image /= 127.5
    return image

def cosin_metric(x1, x2):
    """计算余弦相似度"""
    return np.dot(x1, x2) / (np.linalg.norm(x1) * np.linalg.norm(x2))

def main():
    print("="*80)
    print("模型验证报告")
    print("="*80)

    # 1. 加载配置和模型
    print("\n[1/4] 加载配置和模型...")
    opt = Config('config.yaml')
    device = torch.device("cuda" if opt.use_gpu and torch.cuda.is_available() else "cpu")
    print(f"  设备: {device}")
    print(f"  模型路径: {opt.test_model_path}")

    if not os.path.exists(opt.test_model_path):
        print(f"  ✗ 模型文件不存在")
        return

    model = resnet_face18(use_se=opt.use_se)
    model.load_state_dict(torch.load(opt.test_model_path, map_location=device))
    model.to(device)
    model.eval()
    print(f"  ✓ 模型加载成功")

    # 2. LFW配对验证
    print(f"\n[2/4] LFW配对验证...")
    print(f"  测试列表: {opt.lfw_test_list}")

    if not os.path.exists(opt.lfw_test_list):
        print(f"  ✗ 测试列表不存在")
        return

    with open(opt.lfw_test_list, 'r') as f:
        pairs = f.readlines()

    print(f"  总配对数: {len(pairs)}")

    sims = []
    labels = []
    failed = 0
    processed = 0

    with torch.no_grad():
        for idx, pair in enumerate(pairs):
            splits = pair.strip().split()
            if len(splits) != 3:
                continue

            img_path1 = os.path.join(opt.lfw_root, splits[0])
            img_path2 = os.path.join(opt.lfw_root, splits[1])
            label = int(splits[2])

            # 加载图片
            image1 = load_image(img_path1)
            image2 = load_image(img_path2)

            if image1 is None or image2 is None:
                failed += 1
                continue

            # 转换为tensor
            data1 = torch.from_numpy(image1).to(device)
            data2 = torch.from_numpy(image2).to(device)

            # 提取特征
            output1 = model(data1)
            output2 = model(data2)

            # 合并原图和翻转图的特征
            fe_1 = output1.data.cpu().numpy()
            fe_2 = output2.data.cpu().numpy()
            feature1 = np.hstack((fe_1[0], fe_1[1]))
            feature2 = np.hstack((fe_2[0], fe_2[1]))

            # 计算余弦相似度
            sim = cosin_metric(feature1, feature2)
            sims.append(sim)
            labels.append(label)
            processed += 1

            # 进度显示
            if (idx + 1) % 1000 == 0:
                print(f"  处理进度: {idx+1}/{len(pairs)}", end='\r')
                sys.stdout.flush()

    print(f"\n  ✓ 处理完成: {processed}个有效配对")
    if failed > 0:
        print(f"  ✗ 失败: {failed}个配对")

    # 3. 计算准确率
    print(f"\n[3/4] 计算准确率...")

    sims = np.asarray(sims)
    labels = np.asarray(labels)

    # 找最佳阈值
    best_acc = 0
    best_th = 0
    thresholds = sorted(sims)

    for th in thresholds[::10]:  # 每隔10个采样
        y_test = (sims >= th)
        acc = np.mean((y_test == labels).astype(int))
        if acc > best_acc:
            best_acc = acc
            best_th = th

    print(f"  最佳准确率: {best_acc*100:.2f}%")
    print(f"  最佳阈值: {best_th:.4f}")

    # 4. 统计分析
    print(f"\n[4/4] 相似度分布分析...")

    pos_sims = sims[labels == 1]
    neg_sims = sims[labels == 0]

    print(f"  同一人 (正样本):")
    print(f"    数量: {len(pos_sims)}")
    print(f"    相似度: {pos_sims.mean():.4f} ± {pos_sims.std():.4f}")
    print(f"    范围: [{pos_sims.min():.4f}, {pos_sims.max():.4f}]")

    print(f"  不同人 (负样本):")
    print(f"    数量: {len(neg_sims)}")
    print(f"    相似度: {neg_sims.mean():.4f} ± {neg_sims.std():.4f}")
    print(f"    范围: [{neg_sims.min():.4f}, {neg_sims.max():.4f}]")

    margin = pos_sims.mean() - neg_sims.mean()
    print(f"\n  判别间距: {margin:.4f}")

    # 评估判别性
    if margin > 0.2:
        quality = "优秀"
    elif margin > 0.15:
        quality = "良好"
    elif margin > 0.1:
        quality = "一般"
    else:
        quality = "较差"

    print(f"  特征判别性: {quality}")

    # 最终总结
    print("\n" + "="*80)
    print("验证总结")
    print("="*80)
    print(f"✓ LFW准确率: {best_acc*100:.2f}%")
    print(f"✓ 最佳阈值: {best_th:.4f}")
    print(f"✓ 判别间距: {margin:.4f}")
    print(f"✓ 特征质量: {quality}")
    print("="*80)

    return best_acc

if __name__ == '__main__':
    acc = main()