#### Tagmachine learning

In my last post, I discussed how to get predictions from Gaussian Processes in STAN quickly using the analytical solution. I was able to get it down to 3.8 seconds, which is pretty quick. But I can do better.

One of this issues you may or may not have noticed is, using the model wherein posterior predictions are generated quantities, your computer bogs down with anything over 1000 iterations (I know mine froze up). The issue here is that generated quantities saves all the variables into memory, including those large matrices $$\boldsymbol{K}_{obs}, \boldsymbol{K}_{obs}^{*}, \boldsymbol{K}^{*}$$ . We can resolve this issue, and speed things up, using functions to calculate the predictive values while storing the matrices only as local variables within the function (they are not saved in memory).

One obvious solution is to not calculate them within STAN, but to do it externally. Python can do this relatively quickly with judicious use of numpy (whose linear algebra functions are all written in C and so very fast) and numba/just-in-time compilers. However, since STAN compiles everything into C++,  and many of its linear algebra functions are optimized for speed, it makes sense to do it all in STAN. Fortunately, we can create our own functions within STAN to do this, which are then compiled and executed in C++. The code is as follows:

import numpy as np
import matplotlib.pyplot as plt
import pystan as pyst

X = np.arange(-5, 6)
Y_m = np.sin(X)
Y = Y_m + np.random.normal(0, 0.5, len(Y_m))

X_pred = np.linspace(-5, 7, 100)

gp_pred = """
functions{
// radial basis, or square exponential, function
matrix rbf(int Nx, vector x, int Ny, vector y, real eta, real rho){
matrix[Nx,Ny] K;
for(i in 1:Nx){
for(j in 1:Ny){
K[i,j] = eta*exp(-rho*pow(x[i]-y[j], 2));
}
}
return K;
}

// all of the posterior calculations are now done within this function
vector post_pred_rng(real eta, real rho, real sn, int No, vector xo, int Np, vector xp, vector yobs){
matrix[No,No] Ko;
matrix[Np,Np] Kp;
matrix[No,Np] Kop;
matrix[Np,No] Ko_inv_t;
vector[Np] mu_p;
matrix[Np,Np] Tau;
matrix[Np,Np] L2;
vector[Np] Yp;

// note the use of matrix multiplication for the sn noise, to remove any for-loops
Ko = rbf(No, xo, No, xo, eta, rho) + diag_matrix(rep_vector(1, No))*sn ;
Kp = rbf(Np, xp, Np, xp, eta, rho) + diag_matrix(rep_vector(1, Np))*sn ;
Kop = rbf(No, xo, Np, xp, eta, rho) ;

Ko_inv_t = Kop' / Ko;
mu_p = Ko_inv_t * yobs;
Tau = Kp - Ko_inv_t * Kop;
L2 = cholesky_decompose(Tau);
Yp = mu_p + L2*rep_vector(normal_rng(0,1), Np);
return Yp;
}
}
data{
int<lower=1> N1;
int<lower=1> N2;
vector[N1] X;
vector[N1] Y;
vector[N2] Xp ;
}
transformed data{
vector[N1] mu;
for(n in 1:N1) mu[n] = 0;
}
parameters{
real<lower=0> eta_sq;
real<lower=0> inv_rho_sq;
real<lower=0> sigma_sq;
}
transformed parameters{
real<lower=0> rho_sq;
rho_sq = inv(inv_rho_sq);
}
model{
matrix[N1,N1] Sigma;
matrix[N1,N1] L_S;

// can actually use those functions here, too!!
Sigma = rbf(N1, X, N1, X, eta_sq, rho_sq) + diag_matrix(rep_vector(1, N1))*sigma_sq;

L_S = cholesky_decompose(Sigma);
Y ~ multi_normal_cholesky(mu, L_S);

eta_sq ~ cauchy(0,5);
inv_rho_sq ~ cauchy(0,5);
sigma_sq ~ cauchy(0, 5);
}
generated quantities{
vector[N2] Ypred;

// this is where the magic happens. note that now we only store Ypred,
// instead of all those extraneous matrices
Ypred = post_pred_rng(eta_sq, rho_sq, sigma_sq,
N1, X, N2, Xp, Y);
}
"""

gp1 = pyst.StanModel(model_code=gp_pred)

data1 = {'N1': len(X), 'X': X, 'Y': Y, 'Xp': X_pred, 'N2': len(X_pred)}
fit_gp1 = gp1.sampling(data1, iter=1000)

There are a couple of tricks here. I used a diagonal matrix $$\boldsymbol{I}\sigma_n$$ to remove a for-loop in calculating the noise variance.  It has the added benefit of cleaning up the code (I find for-loops to be messy to read, I like things tidy). Second, by using functions, the generated quantities only stores the predictions, not all the extra matrices, freeing up tons of memory.

One other thing to note is that the prediction function ends with the suffix ‘_rng’. This is because it is a random number generator, drawing random observations from the posterior distribution. Using the ‘_rng’ suffix allows the function to access other ‘_rng’ functions, chiefly the normal_rng function which draws random normal deviates. That’s necessary for the cholesky trick to turn N(0,1) numbers into the posterior distribution (see the last post). However, normal_rng only returns ONE number, so you have to repeat it for however many observations you need, hence the rep_vector( ) wrapper.

This code is extremely fast. After compilation, it executes 4 chains, 1000 iterations in roughly 1 second (compared to just over three from the previous post). Further, since I no longer have memory issues, I can run more iterations. Whereas 5000 iterations froze my computer before, now it executes in 5 seconds. 10,000 iterations, previously unimaginable on my laptop, runs in 10 seconds.

Pretty great.

As described in an earlier post, Gaussian process models are a fast, flexible tool for making predictions. They’re relatively easy to program if you happen to know the parameters of your covariance function/kernel, but what if you want to estimate them from the data? There are several methods available, but my favorite so far is STAN. True, it requires programming the kernel by hand, but I actually find this easier to understand than trying to parse out the kernel functions from, say, scikit-learn.

STAN can fit GP models quickly, but there are certain tricks you can do that make it lightning fast and accurate. I’ve had trouble getting scikit to converge on a stable/accurate solution, but STAN does this with no problem. Plus, the Hamiltonian Monte-Carlo sampler is very quick for GPs (see the STAN User Manual for more).

Here’s a quick tutorial on how to fit GPs in STAN, and how to speed them up. First, let’s import our modules and simulate some fake data:

import numpy as np
import scipy.stats as st
import matplotlib.pyplot as plt

X = np.arange(-5, 6)
Y_m = np.sin(X)
Y = Y_m + np.random.normal(0, 0.5, len(Y_m))

Now, let’s develop the STAN model. Here’s the boilerplate model for a GP using a squared-exponential model (taken right from the STAN manual).

gp_no_pred = """
data{
int<lower=1> N;
vector[N] X;
vector[N] Y;
}
transformed data{
vector[N] mu;

for(n in 1:N) mu[n] = 0;
}
parameters{
real<lower=0> s_f;
real<lower=0> inv_rho;
real<lower=0> s_n;
}
transformed parameters{
real<lower=0> rho;

rho = inv(inv_rho);
}
model{
matrix[N,N] Sigma;

for(i in 1:(N-1)){
for(j in (i+1):N){
Sigma[i,j] = s_f * exp(-rho * pow(X[i]-X[j], 2));
Sigma[j,i] = Sigma[i,j];
}
}

for(k in 1:N){
Sigma[k,k] = s_f + s_n;
}

Y ~ multi_normal(mu, Sigma);
s_f ~ cauchy(0,5);
inv_rho ~ cauchy(0,5);
s_n ~ cauchy(0,5);

}
"""

This model fits the parameters of the kernel. On my computer, it does so in about  0.15 seconds total (1000 iterations).

Now that we know what the hyperparameters are, we’d like to predict new values, but we also want to do so incorporating full uncertainty in the hyperparameters. The way the STAN manual says to go about this is to make a second vector containing the locations you’d like to predict, paste those together to the X values you have, make a vector of prediction points as parameters, paste those to the Y values you have, and feed those into the multivariate normal distribution as one big mush. The issue here is that the covariance matrix, Sigma, gets very large very fast. Large covariance matrices take a while to invert in the multivariate probability density, and so slows down the sampler.

Here’s the model, after making a vector of 100 prediction points to get a smooth line:

X_pred = np.linspace(-5, 7, 100)

gp_pred1 = """
data{
int<lower=1> N1;
int<lower=1> N2;
vector[N1] X;
vector[N1] Y;
vector[N2] Xp;
}
transformed data{
int<lower=1> N;
vector[N1+N2] Xf;
vector[N1+N2] mu;

N = N1+N2;
for(n in 1:N) mu[n] = 0;
for(n in 1:N1) Xf[n] = X[n];
for(n in 1:N2) Xf[N1+n] = Xp[n];
}
parameters{
real<lower=0> s_f;
real<lower=0> inv_rho;
real<lower=0> s_n;
vector[N2] Yp;
}
transformed parameters{
real<lower=0> rho;

rho = inv(inv_rho);
}
model{
matrix[N,N] Sigma;
vector[N] Yf;

for(n in 1:N1) Yf[n] = Y[n];
for(n in 1:N2) Yf[N1+n] = Yp[n];

for(i in 1:(N-1)){
for(j in (i+1):N){
Sigma[i,j] = s_f * exp(-rho * pow(Xf[i]-Xf[j], 2));
Sigma[j,i] = Sigma[i,j];
}
}

for(k in 1:N){
Sigma[k,k] = s_f + s_n;
}

Yf ~ multi_normal(mu, Sigma);

s_f ~ cauchy(0,5);
inv_rho ~ cauchy(0,5);
s_n ~ cauchy(0, 5);
}
"""

This is extremely slow, taking about 81 seconds on my computer (up from less than one second). Taking the cholesky decompose of Sigma an using multi_normal_cholesky didn’t speed things up, either.

We can speed this up by taking advantage of the analytical form of the solution. That is, once we know the hyperparameters of the kernel from the observed data, we can directly calculate the multivariate normal distribution of the predicted data:

$$P(Y_p | X_p, X_o, Y_o) \sim Normal(\boldsymbol{K}_{obs}^{*’} \boldsymbol{K}_{obs}^{-1} \boldsymbol{y}_{obs}, \boldsymbol{K}^{*} – \boldsymbol{K}_{obs}^{*’} \boldsymbol{K}_{obs}^{-1} \boldsymbol{K}_{obs}^{*})$$

We can calculate those quantities directly within STAN. Then, as an added trick, we can take advantage of the Cholesky decomposition to generate random samples of $$Y_p$$ within STAN as well. Here’s the annotated model:

data{
int<lower=1> N1;
int<lower=1> N2;
vector[N1] X;
vector[N1] Y;
vector[N2] Xp;
}
transformed data{
vector[N1] mu;

for(n in 1:N1) mu[n] = 0;
}
parameters{
real<lower=0> s_f;
real<lower=0> inv_rho;
real<lower=0> s_n;
// This is going to be just a generic (0,1) vector
// for use in generating random samples
vector[N2] z;
}
transformed parameters{
real<lower=0> rho;

rho = inv(inv_rho);
}
model{
// kernel for only the observed data
matrix[N1,N1] Sigma;

for(i in 1:(N1-1)){
for(j in (i+1):N1){
Sigma[i,j] = s_f * exp(-rho * pow(X[i]-X[j], 2));
Sigma[j,i] = Sigma[i,j];
}
}

for(k in 1:N1){
Sigma[k,k] = s_f + s_n;
}

// sampling statement for only the observed data
Y ~ multi_normal(mu, Sigma);
// generic sampling statement for z
z ~ normal(0,1);

s_f ~ cauchy(0,5);
inv_rho ~ cauchy(0,5);
s_n ~ cauchy(0, 5);
}
generated quantities{
matrix[N1,N1] Ko;
matrix[N2,N2] Kp;
matrix[N1,N2] Kop;
matrix[N1,N1] L;
matrix[N1, N1] Ko_inv;
vector[N2] mu_p;
matrix[N2,N2] Tau;
vector[N2] Yp;
matrix[N2,N2] L2;

// kernel for observed data
for(i in 1:(N1-1)){
for(j in (i+1):N1){
Ko[i,j] = s_f * exp(-rho * pow(X[i]-X[j], 2));
Ko[j,i] = Ko[i,j];
}
}

for(k in 1:N1){
Ko[k,k] = s_f + s_n;
}

// kernel for prediction data
for(i in 1:(N2-1)){
for(j in (i+1):N2){
Kp[i,j] = s_f * exp(-rho * pow(Xp[i]-Xp[j], 2));
Kp[j,i] = Kp[i,j];
}
}

for(k in 1:N2){
Kp[k,k] = s_f + s_n;
}

// kernel for observed-prediction cross
for(i in 1:N1){
for(j in 1:N2){
Kop[i,j] = s_f * exp(-rho * pow(X[i]-Xp[j], 2));
}
}

// Follow the algorithm 2.1 of Rassmussen and Williams
// cholesky decompose Ko
L = cholesky_decompose(Ko);
Ko_inv = inverse(L') * inverse(L);
// calculate the mean of the Y prediction
mu_p = Kop' * Ko_inv * Y;
// calculate the covariance of the Y prediction
Tau = Kp - Kop' * Ko_inv * Kop;

// Generate random samples from N(mu,Tau)
L2 = cholesky_decompose(Tau);
Yp = mu_p + L2*z;

}
"""

This model runs at 3.5 seconds, huge improvement over 81. If you want the predictions, just extract Yp, and that gives you predictions with full uncertainty. We can also compare the two outputs:

Note: I have no idea why, but upping the iterations from 1000 to 5000 causes the analytical solution model to freeze my computer. I can’t quite figure that one out.

Gaussian Processes for Machine Learning by Rasmussen and Williams has become the quintessential book for learning Gaussian Processes. They kindly provide their own software that runs in MATLAB or Octave in order to run GPs. However, I find it easiest to learn by programming on my own, and my language of choice is Python. This is the first in a series of posts that will go over GPs in Python and how to produce the figures, graphs, and results presented in Rasmussen and Williams.

• Note, Python as numerous excellent packages for implementing GPs, but here I will work on doing them myself “by hand”.

This post will cover the basics presented in Chapter 2. Specifically, we will cover Figures 2.2, 2.4, and 2.5.

• Note, I’m not covering the theory of GPs here (that’s the subject of the entire book, right?)

Before we get going, we have to set up Python:

import numpy as np
import scipy.stats as st
import matplotlib.pyplot as plt

# Figure 2.2

We want to make smooth lines to start, so make 100 evenly spaced $$x$$ values:

N_star = 101
x_star = np.linspace(-5, 5, N_star)

Next we have to calculate the covariances between all the observations and store them in the matrix $$\boldsymbol{K}$$. Here, we use the squared exponential covariance: $$\text{exp}[-\frac{1}{2}(x_i – x_j)^2]$$

K_star = np.empty((N_star, N_star))
for i in range(N_star):
for j in range(N_star):
K_star[i,j] = np.exp(-0.5 * (x_star[i] - x_star[j])**2)

We now have our prior distribution with a mean of 0 and a covariance matrix of $$\boldsymbol{K}$$. We can draw samples from this prior distribution

priors = st.multivariate_normal.rvs(mean=[0]*N_star, cov=K_star, size=1000)

for i in range(5):
plt.plot(x_star, priors[i])

plt.fill_between(x_star, np.percentile(priors, 2.5, axis=0), np.percentile(priors, 97.5, axis=0), alpha=0.2)
plt.show()

Next, let’s add in some observed data:

x_obs = np.array([-4.5, -3.5, -0.5, 0, 1])
y_obs = np.array([-2, 0, 1, 2, -1])

We now need to calculate the covariance between our unobserved data (x_star) and our observed data (x_obs), as well as the covariance among x_obs points as well. The first for loop calculates observed covariances. The second for loop calculates observed-new covariances.

N_obs = 5
K_obs = np.empty((N_obs, N_obs))
for i in range(N_obs):
for j in range(N_obs):
K_obs[i,j] = np.exp(-0.5 * (x_obs[i] - x_obs[j])**2)

K_obs_star = np.empty((N_obs, N_star))
for i in range(N_obs):
for j in range(N_star):
K_obs_star[i,j] = np.exp(-0.5*(x_obs[i] - x_star[j])**2)

We can then get our posterior distributions:

$$\boldsymbol{\mu} = \boldsymbol{K}_{obs}^{*’} \boldsymbol{K}_{obs}^{-1} \boldsymbol{y}_{obs}$$
$$\boldsymbol{\Sigma} = \boldsymbol{K}^{*} – \boldsymbol{K}_{obs}^{*’} \boldsymbol{K}_{obs}^{-1} \boldsymbol{K}_{obs}^{*}$$

and simulate from this posterior distribution.

post_mean = K_obs_star.T.dot(np.linalg.pinv(K_obs)).dot(y_obs)
post_var = K_star - K_obs_star.T.dot(np.linalg.pinv(K_obs)).dot(K_obs_star)

posteriors = st.multivariate_normal.rvs(mean=post_mean, cov=post_var, size=1000)

for i in range(1000):
plt.plot(x_star, posteriors[i], c='k', alpha=0.01)

plt.plot(x_obs, y_obs, 'ro')
plt.show()

This may not look exactly like the Rasmussen and Williams Fig. 2.2b because I guessed at the data points and they may not be quite right. As the authors point out, we can actually plot what the covariance looks like for difference x-values, say $$x=-1,2,3$$.

plt.plot(x_star, post_var[30], label='x= = -2')
plt.plot(x_star, post_var[60], label='x = 1')
plt.plot(x_star, post_var[80], label='x = 3')
plt.ylabel('Posterior Covariance')
plt.legend()
plt.show()

In this case, however, we’ve forced the scale to be equal to 1, that is you have to be at least one unit away on the x-axis before you begin to see large changes $$y$$. We can incorporate a scale parameter $$\lambda$$ to change that. We can use another parameter $$\sigma_f^2$$ to control the noise in the signal (that is, how close to the points does the line have to pass) and we can add further noise by assuming measurement error $$\sigma_n^2$$.

$$cov(x_i, x_j) = \sigma_f^2 \text{exp}[-\frac{1}{2\lambda^2} (x_i – x_j)^2] + \delta_{ij}\sigma_n^2$$

Let’s make some new data:

x_obs = np.array([-7, -5.5, -5, -4.8, -3, -2.8, -1.5,
-1, -0.5, 0.25, 0.5, 1, 2.5, 2.6, 4.5, 4.6, 5, 5.5, 6])
y_obs = np.array([-2, 0, 0.1, -1, -1.2, -1.15, 0.5, 1.5, 1.75,
0, -0.9, -1, -2.5, -2, -1.5, -1, -1.5, -1.1, -1])
x_star = np.linspace(-7, 7, N_star)

Next, make a couple of functions to calculate $$\boldsymbol{K}_{obs}$$, $$\boldsymbol{K}^{*}$$, and $$\boldsymbol{K}_{obs}^{*}$$.

def k_star(l, sf, sn, x_s):
N_star = len(x_s)
K_star = np.empty((N_star, N_star))
for i in range(N_star):
for j in range(N_star):
K_star[i,j] = sf**2*np.exp(-(1.0 / (2.0*l**2)) * (x_s[i] - x_s[j])**2)
return K_star

def k_obs(l, sf, sn, x_o):
N_obs = len(x_o)
K_obs = np.empty((N_obs, N_obs))
for i in range(N_obs):
for j in range(N_obs):
K_obs[i,j] = sf**2*np.exp(-(1.0 / (2.0*l**2)) * (x_o[i] - x_o[j])**2)
for i in range(N_obs):
K_obs[i,i] += sn**2
return K_obs

def k_obs_star(l, sf, sn, x_o, x_s):
N_obs = len(x_o)
N_star = len(x_s)
K_obs_star = np.empty((N_obs, N_star))
for i in range(N_obs):
for j in range(N_star):
K_obs_star[i,j] = sf**2*np.exp(-(1.0 / (2.0*l**2)) *(x_o[i] - x_s[j])**2)
return K_obs_star

Then run the code for the various sets of parameters. Let’s start with (1, 1, 0.1):

K_star = k_star(1,1,0.2,x_star)
K_obs = k_obs(1,1,0.2, x_obs)
K_obs_star = k_obs_star(1,1,0.2,x_obs, x_star)

post_mean = K_obs_star.T.dot(np.linalg.pinv(K_obs)).dot(y_obs)
post_var = K_star - K_obs_star.T.dot(np.linalg.pinv(K_obs)).dot(K_obs_star)

posteriors = st.multivariate_normal.rvs(mean=post_mean, cov=post_var, size=1000)

for i in range(1000):
plt.plot(x_star, posteriors[i], c='k', alpha=0.01)

plt.plot(x_star, posteriors.mean(axis=0))
plt.plot(x_obs, y_obs, 'ro')
plt.show()

Then try (0.3, 1.08, 0.00005):

K_star = k_star(0.3, 1.08, 0.00005, x_star)
K_obs = k_obs(0.3, 1.08, 0.00005,  x_obs)
K_obs_star = k_obs_star(0.3, 1.08, 0.00005, x_obs, x_star)

post_mean = K_obs_star.T.dot(np.linalg.pinv(K_obs)).dot(y_obs)
post_var = K_star - K_obs_star.T.dot(np.linalg.pinv(K_obs)).dot(K_obs_star)

posteriors = st.multivariate_normal.rvs(mean=post_mean, cov=post_var, size=1000)

for i in range(1000):
plt.plot(x_star, posteriors[i], c='k', alpha=0.01)

plt.plot(x_star, posteriors.mean(axis=0))

plt.plot(x_obs, y_obs, 'ro')
plt.show()

Then finally (3, 1.16, 0.89):

K_star = k_star(3, 1.16, 0.89, x_star)
K_obs = k_obs(3, 1.16, 0.89,  x_obs)
K_obs_star = k_obs_star(3, 1.16, 0.89, x_obs, x_star)

post_mean = K_obs_star.T.dot(np.linalg.pinv(K_obs)).dot(y_obs)
post_var = K_star - K_obs_star.T.dot(np.linalg.pinv(K_obs)).dot(K_obs_star)

posteriors = st.multivariate_normal.rvs(mean=post_mean, cov=post_var, size=1000)

for i in range(1000):
plt.plot(x_star, posteriors[i], c='k', alpha=0.01)

plt.plot(x_star, posteriors.mean(axis=0))

plt.plot(x_obs, y_obs, 'ro')
plt.show()

And there you have it! Figs 2.2, 2.4, and 2.5 from Rasmussen and Williams.