Python Example
Included in the static site for reading. Run locally from the source repository when needed.
"""
第19章：Siamese Network 与 Triplet Loss
Chapter 19: Siamese Network & Triplet Loss
"""
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
class EmbeddingNet(nn.Module):
"""共享权重的编码网络"""
def __init__(self, input_dim=128, embed_dim=64):
super().__init__()
self.fc = nn.Sequential(
nn.Linear(input_dim, 256),
nn.ReLU(),
nn.Linear(256, embed_dim)
)
def forward(self, x):
return self.fc(x)
# Triplet Loss
def triplet_loss(anchor, positive, negative, margin=0.2):
"""
L = max(||f(A)-f(P)||² - ||f(A)-f(N)||² + margin, 0)
"""
d_pos = F.pairwise_distance(anchor, positive, p=2)
d_neg = F.pairwise_distance(anchor, negative, p=2)
loss = torch.relu(d_pos**2 - d_neg**2 + margin)
return loss.mean()
# 模拟训练
net = EmbeddingNet()
optimizer = torch.optim.Adam(net.parameters(), lr=1e-3)
for epoch in range(100):
anchor = torch.randn(16, 128)
positive = anchor + torch.randn(16, 128) * 0.1 # 相似样本（同一人）
negative = torch.randn(16, 128) # 不同样本（不同人）
emb_a = net(anchor)
emb_p = net(positive)
emb_n = net(negative)
loss = triplet_loss(emb_a, emb_p, emb_n)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if epoch % 20 == 0:
print(f"Epoch {epoch}: Loss={loss.item():.4f}")
# 验证
with torch.no_grad():
test_a = net(torch.randn(1, 128))
test_p = net(test_a + torch.randn(1, 128) * 0.1)
test_n = net(torch.randn(1, 128))
d_pos = torch.dist(test_a, test_p).item()
d_neg = torch.dist(test_a, test_n).item()
print(f"\n验证距离:")
print(f" 同一人距离 (A-P): {d_pos:.4f}")
print(f" 不同人距离 (A-N): {d_neg:.4f}")
print(f" 距离差: {d_neg - d_pos:.4f} (应 > 0)")