生成模型 第25课/共35课

第25课:DCGAN

📖 课程概述

本课学习DCGAN的核心原理与实现。我们将从数学基础出发,通过代码实践真正理解每一个算法的运作机制。

📑 本课目录

1. DCGAN架构2. 转置卷积3. BatchNorm在GAN中4. 训练技巧5. 插值生成6. 向量运算7. DCGAN实战

1. DCGAN架构

DCGAN用CNN替代全连接

import torch, torch.nn as nn
G = nn.Sequential(nn.ConvTranspose2d(100,256,4,1,0,bias=False), nn.BatchNorm2d(256), nn.ReLU(), nn.ConvTranspose2d(256,128,4,2,1,bias=False), nn.BatchNorm2d(128), nn.ReLU(), nn.ConvTranspose2d(128,1,4,2,1,bias=False), nn.Tanh())
z = torch.randn(4,100,1,1)
print(f'DCGAN G: {z.shape} -> {G(z).shape}')

2. 转置卷积

转置卷积=可学习的上采样

import torch, torch.nn as nn
conv_t = nn.ConvTranspose2d(256, 128, kernel_size=4, stride=2, padding=1)
x = torch.randn(1, 256, 7, 7)
print(f'转置卷积: {x.shape} -> {conv_t(x).shape}')
print('vs bilinear: 固定插值, 不可学习')

3. BatchNorm在GAN中

BatchNorm稳定GAN训练

import torch, torch.nn as nn
# G用BN, D不用BN(或用LN/SN)
print('DCGAN规则:')
print('G: 用BatchNorm(输出层除外)')
print('D: 用LayerNorm或谱归一化')
print('G: ReLU(输出Tanh)')
print('D: LeakyReLU(0.2)')

4. 训练技巧

训练技巧

print('DCGAN训练技巧:')
print('1. 所有权重N(0,0.02)初始化')
print('2. Adam lr=0.0002, beta1=0.5')
print('3. G用ReLU, D用LeakyReLU')
print('4. 不用MaxPool, 用stride卷积')
print('5. 不用FC层, 用全局池化')

5. 插值生成

隐空间插值

import numpy as np
# 球面线性插值(slerp)
def slerp(z1, z2, t):
    omega = np.arccos(np.dot(z1/(np.linalg.norm(z1)+1e-8), z2/(np.linalg.norm(z2)+1e-8)))
    if omega < 1e-10: return (1-t)*z1 + t*z2
    return np.sin((1-t)*omega)/np.sin(omega)*z1 + np.sin(t*omega)/np.sin(omega)*z2
z1, z2 = np.random.randn(100), np.random.randn(100)
for t in [0, 0.25, 0.5, 0.75, 1.0]:
    z = slerp(z1, z2, t)
    print(f't={t:.2f}: |z|={np.linalg.norm(z):.2f}')

6. 向量运算

向量运算

import numpy as np
# 微笑女人 - 中性女人 + 中性男人 = 微笑男人
z_smile_woman = np.random.randn(100)
z_neutral_woman = z_smile_woman + np.random.randn(100)*0.3
z_neutral_man = np.random.randn(100)
smile_vector = z_smile_woman - z_neutral_woman
z_smile_man = z_neutral_man + smile_vector
print(f'微笑向量范数: {np.linalg.norm(smile_vector):.2f}')
print('GAN向量运算有语义含义!')

7. DCGAN实战

DCGAN实战

import torch, torch.nn as nn
class DCGANG(nn.Module):
    def __init__(self, nz=100, nc=1, ngf=64):
        super().__init__()
        self.main = nn.Sequential(nn.ConvTranspose2d(nz,ngf*8,4,1,0,bias=False), nn.BatchNorm2d(ngf*8), nn.ReLU(True),
                                   nn.ConvTranspose2d(ngf*8,ngf*4,4,2,1,bias=False), nn.BatchNorm2d(ngf*4), nn.ReLU(True),
                                   nn.ConvTranspose2d(ngf*4,nc,4,2,1,bias=False), nn.Tanh())
    def forward(self, x): return self.main(x)
netG = DCGANG()
z = torch.randn(4,100,1,1)
print(f'DCGAN: {z.shape} -> {netG(z).shape}')

🔑 DCGAN核心要点

📝 课后练习

  1. 实现本课所有算法并对比效果
  2. 在不同参数下测试算法的鲁棒性
  3. 将本课方法应用到自己的数据集
  4. 分析算法的计算复杂度和优化方向
延伸阅读:推荐阅读《Digital Image Processing》(Gonzalez)相关章节,以及OpenCV官方文档中的详细API说明。实际项目中,建议先在简单数据上验证算法,再迁移到复杂场景。

📊 方法对比总结

方法优点缺点适用场景
方法A简单高效精度有限快速原型
方法B精度高计算量大离线处理
方法C平衡精度与速度参数调优复杂实际应用

📐 关键公式速查

本课涉及的核心数学公式汇总,方便快速参考:

🔬 DCGAN进阶内容

本节深入探讨DCGAN的进阶话题和工程实践中的关键考虑。

工业实践要点

最新研究进展

# 性能基准测试模板
import time, numpy as np

def benchmark(func, *args, n=100, **kwargs):
    times = []
    for _ in range(n):
        t0 = time.time()
        result = func(*args, **kwargs)
        times.append(time.time() - t0)
    return np.mean(times)*1000, np.std(times)*1000

print(f"性能基准测试: 多次运行取均值和标准差")
print(f"注意: 首次运行可能较慢(JIT编译/缓存加载)")

⚠️ 常见陷阱与解决方案

📊 性能优化策略

优化方向方法精度影响加速比
模型压缩剪枝/蒸馏1-3%下降2-4x
量化INT8/FP16<1%下降2-8x
算子融合TensorRT/ONNX1.5-3x
批处理增大batch size线性

💻 DCGAN完整Pipeline代码

以下是一个完整的DCGAN处理管道,从数据准备到结果评估,包含所有关键步骤和参数调优建议:

import cv2
import numpy as np
import time

class DCGANPipeline:
    """完整的DCGAN处理管道"""
    
    def __init__(self, params=None):
        self.params = params or {}
        self.results = {}
    
    def preprocess(self, image):
        """预处理步骤"""
        # 1. 去噪
        if self.params.get('denoise', True):
            image = cv2.GaussianBlur(image, (3, 3), 1.0)
        
        # 2. 对比度增强
        if self.params.get('enhance', False):
            if len(image.shape) == 2:
                image = cv2.equalizeHist(image)
            else:
                lab = cv2.cvtColor(image, cv2.COLOR_BGR2LAB)
                lab[:,:,0] = cv2.equalizeHist(lab[:,:,0])
                image = cv2.cvtColor(lab, cv2.COLOR_LAB2BGR)
        
        # 3. 尺寸调整
        max_size = self.params.get('max_size', 1024)
        h, w = image.shape[:2]
        if max(h, w) > max_size:
            scale = max_size / max(h, w)
            image = cv2.resize(image, (int(w*scale), int(h*scale)))
        
        return image
    
    def process(self, image):
        """核心处理步骤"""
        t0 = time.time()
        preprocessed = self.preprocess(image)
        t1 = time.time()
        
        # 主处理逻辑
        result = self._main_process(preprocessed)
        t2 = time.time()
        
        self.results['timing'] = {
            'preprocess': (t1-t0)*1000,
            'process': (t2-t1)*1000,
            'total': (t2-t0)*1000
        }
        return result
    
    def _main_process(self, image):
        """主处理逻辑(子类可重写)"""
        return image
    
    def evaluate(self, result, ground_truth=None):
        """评估处理结果"""
        metrics = {}
        if ground_truth is not None:
            if len(result.shape) == 2:
                mse = np.mean((result.astype(float) - ground_truth.astype(float))**2)
                metrics['mse'] = mse
                metrics['psnr'] = 10 * np.log10(255**2 / (mse + 1e-10))
        metrics['timing'] = self.results.get('timing', {})
        return metrics
    
    def visualize(self, image, result):
        """可视化对比"""
        if len(image.shape) == 2:
            image = cv2.cvtColor(image, cv2.COLOR_GRAY2BGR)
        if len(result.shape) == 2:
            result = cv2.cvtColor(result, cv2.COLOR_GRAY2BGR)
        h1, w1 = image.shape[:2]
        h2, w2 = result.shape[:2]
        h, w = max(h1,h2), w1+w2+10
        vis = np.zeros((h, w, 3), dtype=np.uint8)
        vis[:h1, :w1] = image
        vis[:h2, w1+10:] = result
        cv2.putText(vis, 'Input', (10, 25), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0,255,0), 2)
        cv2.putText(vis, 'Output', (w1+20, 25), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0,255,0), 2)
        return vis


# 使用示例
pipeline = DCGANPipeline({'denoise': True, 'enhance': False, 'max_size': 512})
test_img = np.random.randint(0, 256, (256, 256, 3), dtype=np.uint8)
result = pipeline.process(test_img)
metrics = pipeline.evaluate(result)
print(f"处理时间: {metrics['timing'].get('total', 0):.1f}ms")

🔧 参数调优指南

不同参数对结果的影响:

建议从默认参数开始,根据结果逐步调整。先在少量数据上快速迭代,确定参数后再全量处理。

生产环境建议:1) 添加输入验证和异常处理 2) 记录每次处理的参数和结果 3) 设置性能监控和告警 4) 定期在测试集上评估质量 5) 保持模型/算法版本管理

📊 完整评估指标

import numpy as np

class Metrics:
    @staticmethod
    def mse(img1, img2):
        return np.mean((img1.astype(float) - img2.astype(float))**2)
    
    @staticmethod
    def psnr(img1, img2, max_val=255):
        mse_val = Metrics.mse(img1, img2)
        if mse_val == 0: return float('inf')
        return 10 * np.log10(max_val**2 / mse_val)
    
    @staticmethod
    def iou(mask1, mask2):
        inter = np.logical_and(mask1 > 0, mask2 > 0).sum()
        union = np.logical_or(mask1 > 0, mask2 > 0).sum()
        return inter / (union + 1e-10)
    
    @staticmethod
    def f1_score(pred, gt, threshold=0.5):
        pred_bin = (pred > threshold)
        gt_bin = (gt > 0)
        tp = np.logical_and(pred_bin, gt_bin).sum()
        precision = tp / (pred_bin.sum() + 1e-10)
        recall = tp / (gt_bin.sum() + 1e-10)
        return 2 * precision * recall / (precision + recall + 1e-10)

# 示例
img1 = np.random.randint(0, 256, (100,100), dtype=np.uint8)
img2 = img1 + np.random.randint(-10, 10, (100,100))
img2 = np.clip(img2, 0, 255).astype(np.uint8)

print(f"MSE: {Metrics.mse(img1,img2):.2f}")
print(f"PSNR: {Metrics.psnr(img1,img2):.2f} dB")
print(f"说明: PSNR>30dB为良好, >40dB为优秀")

🎯 实际应用场景与案例

本课所学技术在实际工程中有广泛的应用。以下是几个典型场景:

场景1:工业质检

在制造业中,计算机视觉技术被用于产品缺陷检测。关键挑战包括:缺陷样本少(需要数据增强或异常检测方法)、实时性要求高(流水线速度)、光照变化大。

# 工业质检示例
import cv2, numpy as np

# 模拟产品表面
product = np.full((200,200), 180, dtype=np.uint8)
# 添加划痕缺陷
cv2.line(product, (30,80), (170,120), 50, 2)

# 缺陷检测
blurred = cv2.GaussianBlur(product, (5,5), 1.0)
diff = cv2.absdiff(product, blurred)
_, defects = cv2.threshold(diff, 15, 255, cv2.THRESH_BINARY)
defect_area = np.count_nonzero(defects)
print(f"缺陷面积: {defect_area}px, 缺陷率: {defect_area/product.size:.4%}")
is_defective = defect_area > 50
print(f"检测结果: {'不合格' if is_defective else '合格'}")

场景2:自动驾驶

自动驾驶需要实时处理多种视觉任务:车道检测、目标检测、语义分割。延迟要求<50ms,且需要处理各种天气和光照条件。

# 车道检测示例
import cv2, numpy as np

road = np.zeros((480,640), dtype=np.uint8)
# 模拟道路
cv2.fillPoly(road, [np.array([[200,480],[440,480],[350,200],[290,200]])], 100)
# 模拟车道线
cv2.line(road, (280,480), (320,200), 255, 3)
cv2.line(road, (360,480), (320,200), 255, 3)

# 车道检测
edges = cv2.Canny(road, 50, 150)
lines = cv2.HoughLinesP(edges, 1, np.pi/180, 50, minLineLength=100, maxLineGap=50)
print(f"检测到车道线: {len(lines) if lines is not None else 0} 条")

场景3:医学影像

医学影像分析需要高精度和高可靠性。常见应用包括:CT/MRI分割、X光异常检测、病理图像分析。关键要求:误诊率极低、可解释性强、符合医疗法规。

# 简单医学分割
import cv2, numpy as np

# 模拟CT扫描
ct = np.zeros((256,256), dtype=np.uint8)
cv2.ellipse(ct, (128,128), (60,50), 0, 0, 360, 150, -1)  # 器官
cv2.circle(ct, (140,120), 15, 200, -1)  # 病灶

# Otsu分割
_, mask = cv2.threshold(ct, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
print(f"检测到 {len(contours)} 个区域")
for c in contours:
    area = cv2.contourArea(c)
    if area > 100:
        print(f"  区域面积: {area}, 疑似病灶: {area < 1000}")

✅ 实机验证

🏆

DCGAN实战家

你已经掌握了DCGAN的核心知识,继续前进!