T-SNE源码剖析(python版)
T-SNE算法是用于可视化的算法中效果最好的算法之一,相信大家也对T-SNE算法略有耳闻,本文参考T-SNE作者 Laurens van der Maaten 给出的源代码自己实现T-SNE算法代码,以此来加深对T-SNE的理解。先简单介绍一下T-SNE算法,T-SNE将数据点变换映射到概率分布上。
1.在高维空间中构建概率分布
p_{j|i} = \frac{exp(-||x_i-x_j||^2/2\sigma_i^2)}{\sum_{k \neq l}exp(-||x_i-x_k||^2/2\sigma_i^2)}
p_{ji} = \frac{p_{i|j}+p_{j|i}}{2}
2.在低维空间中构建概率分布
q_{ji} = \frac{(1+||y_i-y_j||^{-1})}{\sum_{k \neq i}(1+||y_i-y_k||^{-1})}
3.使用相对熵(KL散度)作为损失函数
C = KL(P||Q) = \sum_{i}\sum_{j}p_{ij}\log \frac{p_{ij}}{q_{ij}}
4.采用梯度下降法来最小化损失函数
\frac{\delta C}{\delta y_i} = 4\sum_{j}(p_{ij} - q_{ij})(y_i - y_j)
T-SNE的详细推导可以参考
那么我们需要解决如下几个问题:
a.对每个每个样本点,需要计算 \sigma_i ,对此我们需要计算困惑度Perp,困惑度可以解释为一个点附近的有效近邻点个数,对于点i来说,有
Perp(x_i)=2^{P(i)}
其中
P(i) = -\sum_{j = 1}^{N}p_{j|i} \log^{p_{j|i}}
b.获得 矩阵 P 的值,其中 P(i,j) = p_{ij}
c.获得矩阵 Q 的值,其中 Q(i,j) = q_{ij}
d.使用梯度下降法最小化损失函数
Laurens van der Maaten 在代码中使用了一些很巧妙的方法来进行计算。贴一下我参考源代码,主要是在源代码上进行了一些修改,删去了一些 Laurens van der Maaten 不必要的优化,使得代码更加符合公式。
import numpy as np
import matplotlib.pyplot as plt
# 输入为(n*m)的矩阵,表示n个样本,m个属性
# 返回一个距离矩阵
def cal_pairwise_dist(x):
# '''计算pairwise 距离, x是matrix
# (a-b)^2 = a^2 + b^2 - 2*a*b
# '''
sum_x = np.sum(np.square(x), 1)
# print -2 * np.dot(x, x.T)
# print np.add(-2 * np.dot(x, x.T), sum_x).T
dist = np.add(np.add(-2 * np.dot(x, x.T), sum_x).T, sum_x)
#返回任意两个点之间距离的平方
return dist
# 计算困惑度,最终会选择合适的beta,也就是每个点的方差啦
def cal_perplexity(dist, idx=0, beta=1.0):
# '''计算perplexity, D是距离向量,
# idx指dist中自己与自己距离的位置,beta是高斯分布参数
# 这里的perp仅计算了熵,方便计算
# '''
prob = np.exp(-dist * beta)
# 设置自身prob为0
prob[idx] = 0
sum_prob = np.sum(prob)
if sum_prob == 0:
prob = np.maximum(prob, 1e-12)
perp = -12
else:
prob /= sum_prob
perp = 0
for pj in prob:
if pj != 0:
perp += -pj*np.log(pj)
# 困惑度和pi\j的概率分布
return perp, prob
def seach_prob(x, tol=1e-5, perplexity=30.0):
# '''二分搜索寻找beta,并计算pairwise的prob
# '''
# 初始化参数
print("Computing pairwise distances...")
(n, d) = x.shape
dist = cal_pairwise_dist(x)
pair_prob = np.zeros((n, n))
beta = np.ones((n, 1))
# 取log,方便后续计算
base_perp = np.log(perplexity)
for i in range(n):
if i % 500 == 0:
print("Computing pair_prob for point %s of %s ..." %(i,n))
betamin = -np.inf
betamax = np.inf
#dist[i]需要换不能是所有点
perp, this_prob = cal_perplexity(dist[i], i, beta[i])
# 二分搜索,寻找最佳sigma下的prob
perp_diff = perp - base_perp
tries = 0
while np.abs(perp_diff) > tol and tries < 50:
if perp_diff > 0:
betamin = beta[i].copy()
if betamax == np.inf or betamax == -np.inf:
beta[i] = beta[i] * 2
else:
beta[i] = (beta[i] + betamax) / 2
else:
betamax = beta[i].copy()
if betamin == np.inf or betamin == -np.inf:
beta[i] = beta[i] / 2
else:
beta[i] = (beta[i] + betamin) / 2
# 更新perb,prob值
perp, this_prob = cal_perplexity(dist[i], i, beta[i])
perp_diff = perp - base_perp
tries = tries + 1
# 记录prob值
pair_prob[i,] = this_prob
print("Mean value of sigma: ", np.mean(np.sqrt(1 / beta)))
#每个点对其他点的条件概率分布pi\j
return pair_prob
def tsne(x, no_dims=2, initial_dims=50, perplexity=30.0, max_iter=800):
"""Runs t-SNE on the dataset in the NxD array x
to reduce its dimensionality to no_dims dimensions.
The syntaxis of the function is Y = tsne.tsne(x, no_dims, perplexity),
where x is an NxD NumPy array.
# Check inputs
if isinstance(no_dims, float):
print("Error: array x should have type float.")
return -1
if round(no_dims) != no_dims:
print("Error: number of dimensions should be an integer.")
return -1
(n, d) = x.shape
print (x.shape)
eta = 500
# 随机初始化Y
y = np.random.randn(n, no_dims)
# dy梯度
dy = np.zeros((n, no_dims))
# 对称化
P = seach_prob(x, 1e-5, perplexity)
P = P + np.transpose(P)
P = P / np.sum(P) #pij
# early exaggeration
# pi\j
P = P * 4
P = np.maximum(P, 1e-12)
# Run iterations
for iter in range(max_iter):
# Compute pairwise affinities
sum_y = np.sum(np.square(y), 1)
num = 1 / (1 + np.add(np.add(-2 * np.dot(y, y.T), sum_y).T, sum_y))
num[range(n), range(n)] = 0
Q = num / np.sum(num) #qij
Q = np.maximum(Q, 1e-12) #X与Y逐位比较取其大者
# Compute gradient
#pij-qij
PQ = P - Q
#梯度dy
for i in range(n):
dy[i,:] = np.sum(np.tile(PQ[:,i] * num[:,i], (no_dims, 1)).T * (y[i,:] - y), 0)
# 更新y
y = y - eta*dy
# 减去均值
y = y - np.tile(np.mean(y, 0), (n, 1))
# Compute current value of cost function
if (iter + 1) % 50 == 0:
if iter > 100:
C = np.sum(P * np.log(P / Q))
else:
C = np.sum( P/4 * np.log( P/4 / Q))
print("Iteration ", (iter + 1), ": error is ", C)
# Stop lying about P-values
if iter == 100:
P = P / 4
print("finished training!")
return y
if __name__ == "__main__":
print("Run Y = tsne.tsne(X, no_dims, perplexity) to perform t-SNE on your dataset.")