import pystan
import numpy as np
import os
import time
theta = 10
n = 50
y = np.random.exponential(10, n)
data = {'n': y.size, 'y': y}
seed = 1
include_files = ["external_manual.hpp"]
start = time.time()
sm = pystan.StanModel(file="model.stan",
allow_undefined=True,
includes=include_files,
include_dirs=["."],
verbose=True)
print("Compilation took {0:.1f} seconds".format(time.time() - start))
fit = sm.vb(data=data, iter=1000, algorithm='meanfield', seed=seed)
print(fit['mean_pars'])
stan_data["obs_mBx1c_cov"] = np.array(
[list(item) for item in stan_data["obs_mBx1c_cov"]], dtype=np.float64)
print(stan_data["obs_mBx1c_cov"])
stan_data["allow_alpha_S_N"] = 1
print("SKEW NORMAL!!!!" * 100)
pfl_name = subprocess.getoutput("hostname") + "_1DGP.pickle"
f = open("../stan_code_1DGP.txt", 'r')
lines = f.read()
f.close()
try:
[sm, sc] = pickle.load(open(pfl_name, 'rb'))
if sc != lines:
raise_time
except:
sm = pystan.StanModel(file="../stan_code_1DGP.txt")
pickle.dump([sm, lines], open(pfl_name, 'wb'))
fit = sm.sampling(data=stan_data, iter=1000, chains=4, refresh=10, init=initfn)
print(fit)
fit_params = fit.extract(permuted=True)
pickle.dump((stan_data, fit_params), open("results.pickle", 'wb'))
parameters {
real beta0;
real beta1;
real<lower=0> sigma;
}
model {
vector[nobs] mu;
mu = beta0 + beta1 * x;
y ~ normal(mu, sigma); // Likelihood function
}
"""
# compile model
model = pystan.StanModel(model_code=stan_code)
# perform fit
fit = model.sampling(data=toy_data, iter=5000, chains=3, verbose=False, n_jobs=3)
# Output
nlines = 8 # number of lines in screen output
output = str(fit).split('\n')
for item in output[:nlines]:
print(item)
# Plot
import pylab as plt
import matplotlib.pyplot as plt
import numpy as np
from simulate_data import simulate_data
import ranking as rk
import pystan
import patsy
import pickle
try:
gen = pickle.load(open("thurstonian_gen.pkl", 'rb'))
except FileNotFoundError:
gen = pystan.StanModel(file="thurstonian_cov_generate.stan")
with open('thurstonian_gen.pkl', 'wb') as f:
pickle.dump(gen, f)
try:
sm = pickle.load(open("thurstonian.pkl", 'rb'))
except FileNotFoundError:
sm = pystan.StanModel(file="thurstonian_cov.stan")
with open('thurstonian_cov.pkl', 'wb') as f:
pickle.dump(sm, f)
y_mean = x_full @ W
y_full = y_mean + np.random.normal(0, noise_sigma, (N_train + N_test, M))
Y = y_full[:N_train]
Y_test = y_full[N_train:]
model_data = {
'N': N_train,
'K': K,
'y': Y[:, 0],
'X': X,
'sigma': noise_sigma
}
try:
sm = pickle.load(open('model_linear_reg_chains192.pkl', 'rb'))
except:
sm = pystan.StanModel(model_code=linear_reg_fixed_variance_code)
with open('model_linear_reg_chains192.pkl', 'wb') as f:
pickle.dump(sm, f)
for n in range(N_sim):
w_mean_n = w_mean_vi_list[n]
num_proposal_samples = 4000
try:
fit_hmc
except NameError:
fit_hmc = sm.sampling(data=model_data, iter=2000)
try:
fit_vb
except NameError:
fit_vb = sm.vb(data=model_data,
def test_threading_support(self):
# Dont test with Windows
if sys.platform.startswith("win"):
return
# Set up environmental variable
os.environ['STAN_NUM_THREADS'] = "2"
# Enable threading
extra_compile_args = ['-pthread', '-DSTAN_THREADS']
stan_code = """
functions {
vector bl_glm(vector mu_sigma, vector beta,
real[] x, int[] y) {
vector[2] mu = mu_sigma[1:2];
vector[2] sigma = mu_sigma[3:4];
real lp = normal_lpdf(beta | mu, sigma);
real ll = bernoulli_logit_lpmf(y | beta[1] + beta[2] * to_vector(x));
return [lp + ll]';
}
}
data {
int<lower = 0> K;
int<lower = 0> N;
vector[N] x;
int<lower = 0, upper = 1> y[N];
}
transformed data {
int<lower = 0> J = N / K;
real x_r[K, J];
int<lower = 0, upper = 1> x_i[K, J];
{
int pos = 1;
for (k in 1:K) {
int end = pos + J - 1;
x_r[k] = to_array_1d(x[pos:end]);
x_i[k] = y[pos:end];
pos += J;
}
}
}
parameters {
vector[2] beta[K];
vector[2] mu;
vector<lower=0>[2] sigma;
}
model {
mu ~ normal(0, 2);
sigma ~ normal(0, 2);
target += sum(map_rect(bl_glm, append_row(mu, sigma),
beta, x_r, x_i));
}
"""
stan_data = dict(K=4,
N=12,
x=[
1.204, -0.573, -1.35, -1.157, -1.29, 0.515, 1.496,
0.918, 0.517, 1.092, -0.485, -2.157
],
y=[1, 1, 0, 0, 1, 1, 0, 0, 0, 1, 1, 1])
stan_model = pystan.StanModel(model_code=stan_code,
extra_compile_args=extra_compile_args)
fit = stan_model.sampling(data=stan_data, chains=2, n_jobs=1)
self.assertIsNotNone(fit)
fit2 = stan_model.sampling(data=stan_data, chains=2, n_jobs=2)
self.assertIsNotNone(fit2)
draw = fit.extract(pars=fit.model_pars + ['lp__'], permuted=False)
lp = {
key: values[-1, 0]
for key, values in draw.items() if key == 'lp__'
}['lp__']
draw = {
key: values[-1, 0]
for key, values in draw.items() if key != 'lp__'
}
draw = fit.unconstrain_pars(draw)
self.assertEqual(fit.log_prob(draw), lp)
draw2 = fit2.extract(pars=fit2.model_pars + ['lp__'], permuted=False)
lp2 = {
key: values[-1, 0]
for key, values in draw2.items() if key == 'lp__'
}['lp__']
draw2 = {
key: values[-1, 0]
for key, values in draw2.items() if key != 'lp__'
}
draw2 = fit2.unconstrain_pars(draw2)
self.assertEqual(fit2.log_prob(draw2), lp2)
len_obs = len(obs_no)
print(len_obs)
print(obs_no)
print(y)
stan_data = {
'T': T,
'len_obs': len_obs,
'y': y,
'obs_no': obs_no
}
if os.path.exists('5-3-2-local-level-interpolation.pkl'):
# sm = pickle.load(open('5-3-2-local-level-interpolation.pkl', 'rb'))
sm = pystan.StanModel(file='5-3-2-local-level-interpolation.stan')
else:
# a model using prior for mu and sigma.
sm = pystan.StanModel(file='5-3-2-local-level-interpolation.stan')
control = {
'adapt_delta': 0.8,
'max_treedepth': 10
}
mcmc_result = sm.sampling(
data=stan_data,
seed=1,
chains=4,
iter=2000,
warmup=1000,
def model_impact_causal(df,
kind='self',
prior_beta=1.,
prior_delta=1.,
prior_alpha=1.,
prior_mu_beta=1.,
prior_mu_alpha=1.,
code_only=False,
iters=NUM_SAMPLES,
adapt_delta=ADAPT_DELTA,
max_treedepth=M_TREEDEPTH):
# Model: Eqs. 1 & 2, with `f=0`
# Results: Tables 1, S2; Figure 4
import pystan as st
model_code = '''
data {
int<lower=1> n; //number of data points
int<lower=1> m; //number of conditions
int<lower=2> k; //number of outcomes
int<lower=1,upper=m> x_cond[n]; //treatment group
int<lower=1,upper=k> y_pre[n]; //pre-exposure outcome
int<lower=1,upper=k> y_post[n]; //post-exposure outcome
}
parameters {
real mu_beta;
real beta[m];
vector<lower=0>[k-1] mu_delta;
simplex[k-1] delta[m];
vector[k-1] mu_alpha;
ordered[k-1] alpha[m];
ordered[k-1] alpha_pre;
}
model {
mu_beta ~ normal(0, %f);
mu_alpha ~ normal(0, %f);
mu_delta ~ exponential(%f);
for (i in 1:m){
beta[i] ~ normal(mu_beta, %f);
alpha[i] ~ normal(mu_alpha, %f);
delta[i] ~ dirichlet(mu_delta);
}
alpha_pre ~ normal(0, %f);
for (i in 1:n){
y_pre[i] ~ ordered_logistic(0, alpha_pre);
y_post[i] ~ ordered_logistic(beta[x_cond[i]]*sum(delta[x_cond[i]][:y_pre[i]-1]), alpha[x_cond[i]]);
}
}
''' % (prior_mu_beta, prior_mu_alpha, prior_delta, prior_beta,
prior_alpha, prior_alpha)
data = {
'n': df.shape[0],
'm': 2,
'k': 4,
'x_cond': df['Treatment'].values + 1,
'y_pre': df['Vaccine Intent for %s (Pre)' % kind].values,
'y_post': df['Vaccine Intent for %s (Post)' % kind].values
}
if code_only: return model_code
model = st.StanModel(model_code=model_code)
fit = model.sampling(data=data,
iter=iters,
control=dict(adapt_delta=adapt_delta,
max_treedepth=max_treedepth))
return fit
def model_socdem(df,
dd,
atts=[],
model_name='causal',
group=None,
kind='self',
decrease=0,
prior_beta=1.,
prior_delta=1.,
prior_alpha=1.,
prior_mu_beta=1.,
prior_mu_alpha=1.,
code_only=False,
iters=NUM_SAMPLES,
adapt_delta=ADAPT_DELTA,
max_treedepth=M_TREEDEPTH):
# Model: Eqs. 1 & 2, with `f` modeling a linear combination of socio-demographics and other covariates
# Results: Tables S3, S4, S5, S6; Figures 5, S1, S2, S3
import pystan as st
import numpy as np
from .bayesoc import Dim #we define some helper classes to extract posterior samples easily
cats = [
'Age', 'Gender', 'Education', 'Employment', 'Religion', 'Political',
'Ethnicity', 'Income'
]
if isinstance(atts, str): atts = [atts]
for att in atts:
cats += [x for x in list(df) if x[:len(att)] == att]
outs = [
'Vaccine Intent for %s (Pre)' % kind,
'Vaccine Intent for %s (Post)' % kind, 'Treatment'
]
df = df[cats + outs].dropna()
if group is not None: model_name = 'post'
causal = int(model_name in ['causal', 'causaldiff'])
dims = [
Dim(pi=len(dd[cat]),
out=causal + 1,
prior_beta=prior_beta,
prior_mu_beta=prior_mu_beta,
value=dd[cat].keys(),
name=cat) for cat in cats
]
stan = [
d.get_stan(outcome_size='m',
outcome_index='x_cond[i]',
hierarchical=True) for d in dims
]
code = {'data': [], 'parameters': [], 'model': [], 'output': []}
for key in code:
for d in stan:
code[key].append(d[key])
model_code = {
'pre':
'''
data {
int<lower=1> n; //number of data points
int<lower=2> k; //number of outcomes
int<lower=1,upper=k> y_pre[n]; //pre-exposure
%s
}
parameters {
%s
ordered[k-1] alpha;
}
model {
%s
alpha ~ normal(0, %f);
for (i in 1:n)
y_pre[i] ~ ordered_logistic(%s, alpha);
}
''' %
('\n'.join(code['data']), '\n'.join(code['parameters']), '\n'.join(
code['model']), prior_alpha, ' + '.join(code['output'])),
'post':
'''
data {
int<lower=1> n; //number of data points
int<lower=2> k; //number of outcomes
int<lower=1,upper=k> y_pre[n]; //pre-exposure
int<lower=1,upper=k> y_post[n]; //post-exposure
%s
}
parameters {
%s
real beta;
simplex[k-1-%i] delta;
ordered[k-1] alpha;
}
model {
%s
beta ~ normal(0, %f);
{
vector[k-1-%i] u;
for (i in 1:(k-1-%i))
u[i] = 1;
delta ~ dirichlet(%f*u);
}
alpha ~ normal(0, %f);
for (i in 1:n)
y_post[i] ~ ordered_logistic(beta*sum(delta[:y_pre[i]-1-%i]) + %s, alpha);
}
''' %
('\n'.join(code['data']), '\n'.join(code['parameters']), decrease,
'\n'.join(code['model']), prior_beta, decrease, decrease, prior_delta,
prior_alpha, decrease, ' + '.join(code['output'])),
'causal':
'''
data {
int<lower=1> n; //number of data points
int<lower=1> m; //number of conditions
int<lower=2> k; //number of outcomes
int<lower=1,upper=m> x_cond[n]; //treatment group
int<lower=1,upper=k> y_pre[n]; //pre-exposure outcome
int<lower=1,upper=k> y_post[n]; //post-exposure outcome
%s
}
parameters {
%s
real mu_beta;
real beta[m];
vector<lower=0>[k-1-%i] mu_delta;
simplex[k-1-%i] delta[m];
vector[k-1] mu_alpha;
ordered[k-1] alpha[m];
}
model {
%s
mu_beta ~ normal(0, %f);
mu_alpha ~ normal(0, %f);
mu_delta ~ exponential(%f);
for (i in 1:m){
beta[i] ~ normal(mu_beta, %f);
alpha[i] ~ normal(mu_alpha, %f);
delta[i] ~ dirichlet(mu_delta);
}
for (i in 1:n)
y_post[i] ~ ordered_logistic(beta[x_cond[i]]*sum(delta[x_cond[i]][:y_pre[i]-1-%i]) + %s, alpha[x_cond[i]]);
}
''' %
('\n'.join(code['data']), '\n'.join(code['parameters']), decrease,
decrease, '\n'.join(code['model']), prior_mu_beta, prior_mu_alpha,
prior_delta, prior_beta, prior_alpha, decrease, ' + '.join(
code['output'])),
'causaldiff':
'''
data {
int<lower=1> n; //number of data points
int<lower=1> m; //number of conditions
int<lower=2> k; //number of outcomes
int<lower=1,upper=m> x_cond[n]; //treatment group
int<lower=1,upper=k> y_diff[n]; //post-pre difference in outcomes
%s
}
parameters {
%s
vector[k-1] mu_alpha;
ordered[k-1] alpha[m];
}
model {
%s
mu_alpha ~ normal(0, %f);
for (i in 1:m)
alpha[i] ~ normal(mu_alpha, %f);
for (i in 1:n)
y_diff[i] ~ ordered_logistic(%s, alpha[x_cond[i]]);
}
''' % ('\n'.join(code['data']), '\n'.join(
code['parameters']), '\n'.join(code['model']), prior_mu_alpha,
prior_alpha, ' + '.join(code['output']))
}
data = {}
if causal:
data['m'] = 2
data['x_cond'] = df['Treatment'].values + 1
if model_name == 'causal':
data['k'] = 4
data['y_pre'] = df['Vaccine Intent for %s (Pre)' % kind].values
data['y_post'] = df['Vaccine Intent for %s (Post)' % kind].values
else:
data['k'] = 3
tmp = df['Vaccine Intent for %s (Post)' %
kind].values - df['Vaccine Intent for %s (Pre)' %
kind].values
tmp[tmp > 0] = 1
tmp[tmp < 0] = -1
data['y_diff'] = tmp + 2
elif model_name == 'post':
data['k'] = 4
df = df.loc[df['Treatment'] == group]
data['y_pre'] = df['Vaccine Intent for %s (Pre)' % kind].values
data['y_post'] = df['Vaccine Intent for %s (Post)' % kind].values
else:
data['k'] = 4 - decrease
data['y_pre'] = df['Vaccine Intent for %s (Pre)' %
kind].values - decrease
data['n'] = df.shape[0]
print('Dataframe of size:', df.shape)
for i in range(len(cats)):
name = dims[i].name
data['k_%s' % name] = len(dd[cats[i]])
data[name] = np.array(df[cats[i]].values, dtype=int)
if data[name].min() == 0: data[name] += 1
if code_only: return model_code[model_name]
model = st.StanModel(model_code=model_code[model_name])
fit = model.sampling(data=data,
iter=iters,
control=dict(adapt_delta=adapt_delta,
max_treedepth=max_treedepth))
return fit
def model_image_impact(df,
group=1,
kind='self',
prior_beta=1.,
prior_delta=1.,
prior_gamma=1.,
prior_alpha=1.,
code_only=False,
iters=NUM_SAMPLES,
adapt_delta=ADAPT_DELTA,
max_treedepth=M_TREEDEPTH):
# Model: Eq. 7
# Results: Table S7
import pystan as st
import numpy as np
model_code = '''
data {
int<lower=1> n; //number of data points
int<lower=1> p; //number of images
int<lower=1> m; //number of metrics
int<lower=2> k; //number of outcomes
int<lower=1,upper=k> y_pre[n]; //pre-exposure
int<lower=1,upper=k> y_post[n]; //post-exposure
matrix[p,m] x_img[n]; //image metrics
}
parameters {
real beta;
vector[m] beta_img;
simplex[p] gamma;
simplex[k-1] delta;
ordered[k-1] alpha;
}
model {
beta ~ normal(0, %f);
beta_img ~ normal(0, %f);
{
vector[p] u_img;
for (i in 1:p)
u_img[i] = 1;
gamma ~ dirichlet(%f*u_img);
}
{
vector[k-1] u;
for (i in 1:(k-1))
u[i] = 1;
delta ~ dirichlet(%f*u);
}
alpha ~ normal(0, %f);
for (i in 1:n)
y_post[i] ~ ordered_logistic(beta*sum(delta[:y_pre[i]-1]) + to_row_vector(gamma)*x_img[i]*beta_img, alpha);
}
''' % (prior_beta, prior_beta, prior_gamma, prior_delta,
prior_alpha)
metrics = ['Vaccine Intent', 'Agreement', 'Trust', 'Fact-check', 'Share']
df = df.loc[df['Treatment'] == group]
x = np.dstack([
df[['Image %i:%s' % (i + 1, m) for i in range(5)]].values
for m in metrics
])
data = {
'n': df.shape[0],
'p': 5,
'm': len(metrics),
'k': 4,
'x_img': x,
'y_pre': df['Vaccine Intent for %s (Pre)' % kind].values,
'y_post': df['Vaccine Intent for %s (Post)' % kind].values
}
if code_only: return model_code
model = st.StanModel(model_code=model_code)
fit = model.sampling(data=data,
iter=iters,
control=dict(adapt_delta=adapt_delta,
max_treedepth=max_treedepth))
return fit
def model_similar_content(df,
model_name='seen',
kind='self',
prior_beta=1.,
prior_alpha=1.,
prior_delta=1.,
prior_mu_beta=1.,
prior_mu_alpha=1.,
code_only=False,
iters=NUM_SAMPLES,
adapt_delta=ADAPT_DELTA,
max_treedepth=M_TREEDEPTH):
import pystan as st
import numpy as np
model_code = {
'pre':
'''
data {
int<lower=1> n; //number of data points
int<lower=1> m; //number of conditions
int<lower=2> k; //number of outcomes
int<lower=1,upper=m> x_cond[n]; //treatment group
int<lower=0,upper=1> x_seen[n]; //seen images
int<lower=1,upper=k> y_pre[n]; //pre-exposure outcome
}
parameters {
real mu_beta;
real beta[m];
vector[k-1] mu_alpha;
ordered[k-1] alpha[m];
real<lower=0,upper=1> theta[m];
}
model {
mu_beta ~ normal(0, %f);
mu_alpha ~ normal(0, %f);
for (i in 1:m){
beta[i] ~ normal(mu_beta, %f);
alpha[i] ~ normal(mu_alpha, %f);
}
theta ~ beta(1, 1);
for (i in 1:n){
x_seen[i] ~ bernoulli(theta[x_cond[i]]);
y_pre[i] ~ ordered_logistic(beta[x_cond[i]]*x_seen[i], alpha[x_cond[i]]);
}
}
''' %
(prior_mu_beta, prior_mu_alpha, prior_beta, prior_alpha),
'causal':
'''
data {
int<lower=1> n; //number of data points
int<lower=1> m; //number of conditions
int<lower=2> k; //number of outcomes
int<lower=1,upper=m> x_cond[n]; //treatment group
int<lower=0,upper=1> x_seen[n]; //seen images
int<lower=1,upper=k> y_pre[n]; //pre-exposure outcome
int<lower=1,upper=k> y_post[n]; //post-exposure outcome
}
parameters {
real mu_beta;
real beta[m];
real mu_beta_pre;
real beta_pre[m];
real mu_beta_post;
real beta_post[m];
vector<lower=0>[k-1] mu_delta;
simplex[k-1] delta[m];
vector[k-1] mu_alpha_pre;
ordered[k-1] alpha_pre[m];
vector[k-1] mu_alpha_post;
ordered[k-1] alpha_post[m];
real<lower=0,upper=1> theta[m];
}
model {
mu_beta ~ normal(0, %f);
mu_beta_pre ~ normal(0, %f);
mu_beta_post ~ normal(0, %f);
mu_alpha_pre ~ normal(0, %f);
mu_alpha_post ~ normal(0, %f);
mu_delta ~ exponential(%f);
for (i in 1:m){
beta[i] ~ normal(mu_beta, %f);
beta_pre[i] ~ normal(mu_beta_pre, %f);
beta_post[i] ~ normal(mu_beta_post, %f);
alpha_pre[i] ~ normal(mu_alpha_pre, %f);
alpha_post[i] ~ normal(mu_alpha_post, %f);
delta[i] ~ dirichlet(mu_delta);
}
theta ~ beta(1, 1);
for (i in 1:n){
x_seen[i] ~ bernoulli(theta[x_cond[i]]);
y_pre[i] ~ ordered_logistic(beta_pre[x_cond[i]]*x_seen[i], alpha_pre[x_cond[i]]);
y_post[i] ~ ordered_logistic(beta[x_cond[i]]*sum(delta[x_cond[i]][:y_pre[i]-1])+beta_post[x_cond[i]]*x_seen[i], alpha_post[x_cond[i]]);
}
}
''' %
(prior_mu_beta, prior_mu_beta, prior_mu_beta, prior_mu_alpha,
prior_mu_alpha, prior_delta, prior_beta, prior_beta, prior_beta,
prior_alpha, prior_alpha),
'seen_ordinal':
'''
data {
int<lower=1> n; //number of data points
int<lower=1> m; //number of conditions
int<lower=2> k; //number of outcomes
int<lower=1,upper=m> x_cond[n]; //treatment group
int<lower=0,upper=1> x_seen[n]; //seen images
int<lower=1,upper=k> y_pre[n]; //pre-exposure outcome
}
parameters {
real mu_beta;
real beta[m];
vector<lower=0>[k-1] mu_delta;
simplex[k-1] delta[m];
real mu_alpha;
real alpha[m];
}
model {
mu_beta ~ normal(0, %f);
mu_alpha ~ normal(0, %f);
mu_delta ~ exponential(%f);
for (i in 1:m){
beta[i] ~ normal(mu_beta, %f);
alpha[i] ~ normal(mu_alpha, %f);
delta[i] ~ dirichlet(mu_delta);
}
for (i in 1:n)
x_seen[i] ~ bernoulli_logit(beta[x_cond[i]]*sum(delta[x_cond[i]][:y_pre[i]-1]) + alpha[x_cond[i]]);
}
''' %
(prior_mu_beta, prior_mu_alpha, prior_delta, prior_beta, prior_alpha),
'seen':
'''
data {
int<lower=1> n; //number of data points
int<lower=1> m; //number of conditions
int<lower=2> k; //number of outcomes
int<lower=1,upper=m> x_cond[n]; //treatment group
int<lower=0,upper=1> x_seen[n]; //seen images
int<lower=1,upper=k> y_pre[n]; //pre-exposure outcome
}
parameters {
vector[k] mu_beta;
vector[k] beta[m];
}
model {
mu_beta ~ normal(0, %f);
for (i in 1:m)
beta[i] ~ normal(mu_beta, %f);
for (i in 1:n)
x_seen[i] ~ bernoulli_logit(beta[x_cond[i]][y_pre[i]]);
}
''' % (prior_mu_beta, prior_beta)
}
df = df.loc[df['Seen such online content'] != 3] #ignoring do-not-know's
data = {
'n': df.shape[0],
'm': 2,
'k': 4,
'x_cond': df['Treatment'].values + 1,
'y_pre': df['Vaccine Intent for %s (Pre)' % kind].values,
'x_seen': [i % 2 for i in df['Seen such online content'].values]
} #"yes":1, "no":0
if model_name == 'causal':
data['y_post'] = df['Vaccine Intent for %s (Post)' % kind].values
if code_only: return model_code
model = st.StanModel(model_code=model_code[model_name])
fit = model.sampling(data=data,
iter=iters,
control=dict(adapt_delta=adapt_delta,
max_treedepth=max_treedepth))
return fit
def time_stan():
return pystan.StanModel(model_code=model, verbose=True)
3] = 0.41 # Replace subject C12, C26, and C31's implausibly low minimum RT with the lowest RT from all other subjects' data
model_rt_mins[14] = 0.41
model_rt_mins[20] = 0.41
model_data = {
'NS': 30,
'NP': 20,
'NT': 30,
'correct': model_data_correct,
'feedback': model_data_feedback,
'bid_congruence': model_data_bid_congruence,
'rt': model_data_rt,
'rt_mins': model_rt_mins,
'item_popularity': model_data_item_popularity
}
model_code_obj = pystan.StanModel(
file='model_dual_insight_bayesian_noeye_non_social.stan.cpp',
model_name='model_dual_insight_bayesian_noeye_non_social'
) # Specific to model
fit = model_code_obj.sampling(data=model_data, iter=2000, chains=4, refresh=10)
with open('pickles/model_dual_insight_bayesian_noeye_non_social.pkl',
'wb') as f: # Specific to model
pickle.dump(model_code_obj, f)
with open('pickles/fit_dual_insight_bayesian_noeye_non_social.pkl',
'wb') as f: # Specific to model
pickle.dump(fit, f)
print(fit)
####################################################
def save(obj, filename):
"""Save compiled models for reuse."""
import pickle
with open(filename, 'w') as f:
pickle.dump(obj, f, protocol=pickle.HIGHEST_PROTOCOL)
def load(filename):
"""Reload compiled models for reuse."""
import pickle
return pickle.load(open(filename, 'r'))
####################################################
model = pystan.StanModel(model_code = pyfitfull)
save(model, 'pyfitfull_model')
new_model = load('pyfitfull_model')
####################################################
## part six
## fit stan model with saved pyfitfull_model
####################################################
print("\n fit pystan model")
## fit=pystan.stan(model_code="pystancode.stan", data=Mtable, iter=1000, chains=4)
def setUpClass(cls):
model_code = 'parameters {real y;} model {y ~ normal(0,1);}'
cls.model = pystan.StanModel(model_code=model_code,
model_name="normal1",
verbose=True,
obfuscate_model_name=False)
def main(argv):
if len(argv) > 1:
raise app.UsageError("Too many command-line arguments.")
x, y = LoadCloud(FLAGS.cloud_path)
code = """
functions {
matrix cov_exp_quad_ARD(vector[] x,
real alpha,
vector rho,
real delta) {
int N = size(x);
matrix[N, N] K;
real neg_half = -0.5;
real sq_alpha = square(alpha);
for (i in 1:(N-1)) {
K[i, i] = sq_alpha + delta;
for (j in (i + 1):N) {
real v = sq_alpha * exp(neg_half *
dot_self((x[i] - x[j]) ./ rho));
K[i, j] = v;
K[j, i] = v;
}
}
K[N, N] = sq_alpha + delta;
return K;
}
}
data {
int<lower=1> N;
int<lower=1> D;
vector[D] x[N];
vector[N] y;
}
transformed data {
real delta = 1e-9;
}
parameters {
vector<lower=0>[D] rho;
real<lower=0> alpha;
real<lower=0> sigma;
}
model {
matrix[N, N] cov = cov_exp_quad_ARD(x, alpha, rho, delta)
+ diag_matrix(rep_vector(square(sigma), N));
matrix[N, N] L_cov = cholesky_decompose(cov);
rho ~ inv_gamma(5, 5);
alpha ~ normal(0, 1);
sigma ~ normal(0, 1);
y ~ multi_normal_cholesky(rep_vector(0, N), L_cov);
}
"""
data = {
"N": x.shape[0],
"D": x.shape[1],
"x": x,
"y": y,
}
filename = "/tmp/stan_model_%s" % hashlib.md5(
code.encode("ascii")).hexdigest()
print(filename)
try:
sm = pickle.load(open(filename, "rb"))
except FileNotFoundError:
sm = pystan.StanModel(model_code=code)
with open(filename, "wb") as f:
pickle.dump(sm, f)
fit = sm.sampling(data=data, iter=100000, chains=12)
print(fit)
params = fit.extract(["rho", "alpha", "sigma"])
params = np.concatenate([
params["rho"],
params["alpha"][Ellipsis, np.newaxis],
params["sigma"][Ellipsis, np.newaxis],
], -1)
mean = params.mean(0)
square = (params**2.).mean(0)
print(params.shape)
print(mean.shape)
SaveJSON({"mean": mean, "square": square}, "/tmp/gp_reg_0")
if _model == 'GTR':
data['frequencies_alpha'] = [1, 1, 1, 1]
data['rates_alpha'] = [1, 1, 1, 1, 1, 1]
elif _model == 'HKY':
data['frequencies_alpha'] = [1, 1, 1, 1]
# Samples output file
sample_path = f'{stem}{epi}'
tree_path = f'{sample_path}.trees'
binary = _script.replace('.stan', '.pkl')
if binary == _script:
binary = _script + '.pkl'
if not os.path.lexists(binary) or _compile:
sm = pystan.StanModel(file=_script)
with open(binary, 'wb') as f:
pickle.dump(sm, f)
else:
sm = pickle.load(open(binary, 'rb'))
if _algorithm == 'vb':
stan_args = {}
stan_args['output_samples'] = _samples
if _eta:
stan_args['eta'] = _eta
stan_args['adapt_engaged'] = False
if _seed:
stan_args['seed'] = _seed
fit = sm.vb(data=data,
import os
import pickle
import pystan
"""
Script that compiles models in advance in order to save time during execution
"""
dir_name = os.getcwd() + '/stan_models/'
sm = pystan.StanModel(file=dir_name + 'usage_exp.stan', verbose=True)
with open(dir_name + 'usage_exp_model.pkl', 'wb') as f:
pickle.dump(sm, f)
# sm = pystan.StanModel(file=dir_name + 'usage_exp_dummy.stan')
# with open(dir_name + 'usage_exp_dummy_model.pkl', 'wb') as f:
# pickle.dump(sm, f)
# sm = pystan.StanModel(file=dir_name + 'usage_exp_dummy_markov.stan')
# with open(dir_name + 'usage_exp_dummy_markov_model.pkl', 'wb') as f:
# pickle.dump(sm, f)
model {
// Priors.
alpha ~ lognormal(m_alpha, s_alpha);
rho ~ lognormal(m_rho, s_rho);
eta ~ std_normal();
beta ~ std_normal();
// Model.
y ~ bernoulli_logit(beta + f);
}
"""
# Compile model. This takes about a minute.
sm = pystan.StanModel(model_code=model_code)
# Make data.
X, y = make_moons(n_samples=50, shuffle=True, noise=0.1, random_state=1)
# Generate stan data dictionary.
stan_data = create_stan_data(X, y)
# Fit via ADVI.
vb_fit = sm.vb(data=stan_data, iter=1000, seed=1,
grad_samples=1, elbo_samples=1)
# Fit via HMC.
# - stepsize = 0.05
# - num leapfrog steps = 20
def fit(self, graph):
if not nx.is_directed_acyclic_graph(graph):
raise ValueError(
"Input graph must be a directed acyclic graph (DAG)")
all_dat = {}
all_init = {}
for i in graph.nodes:
node = graph.nodes[i]
parents = list(graph.predecessors(i))
if node.get("prophet") is None:
raise ValueError(
"All nodes must have an attribute called 'prophet' contained a prophet object"
)
m = graph.nodes[i]["prophet"]
if node.get("df") is None:
raise ValueError(
"All nodes must have an attribute called 'df' containing a pandas DataFrame"
)
if node.get("future") is None:
raise ValueError(
"All nodes must have an attribute called 'future' containing a pandas DataFrame"
)
df = graph.nodes[i]["df"]
future = graph.nodes[i]["future"]
future = m.setup_dataframe(future.copy())
only_future = future[df.shape[0]:]
history = df[df['y'].notnull()].copy()
m.history_dates = pd.to_datetime(
pd.Series(df['ds'].unique(), name='ds')).sort_values()
history = m.setup_dataframe(history, initialize_scales=True)
m.history = history
m.set_auto_seasonalities()
seasonal_features, prior_scales, component_cols, modes = (
m.make_all_seasonality_features(history))
m.train_component_cols = component_cols
m.component_modes = modes
m.set_changepoints()
trend_indicator = {'linear': 0, 'logistic': 1, 'flat': 2}
seasonal_features_future, _, _, _ = (
m.make_all_seasonality_features(future))
dat = {
'T': m.history.shape[0],
'T_pred': only_future.shape[0],
't_pred': np.array(only_future.t),
f'K_{i}': seasonal_features.shape[1],
f'S_{i}': len(m.changepoints_t),
f'y_{i}': m.history['y_scaled'],
't': m.history['t'],
f't_change_{i}': m.changepoints_t,
f'X_{i}': seasonal_features,
f'sigmas_{i}': prior_scales,
f'tau_{i}': m.changepoint_prior_scale,
f'trend_indicator_{i}': trend_indicator[m.growth],
f's_a_{i}': component_cols['additive_terms'],
f's_m_{i}': component_cols['multiplicative_terms'],
f'a_{i}': np.array([1, 1, 1, 1]),
f'm_{i}': np.array([0, 0, 0, 0]),
f'X_pred_{i}': seasonal_features_future[df.shape[0]:],
f'S_pred_{i}': 3
}
if m.growth == 'linear':
dat[f'cap_{i}'] = np.zeros(m.history.shape[0])
dat[f'cap_pred_{i}'] = np.zeros(only_future.shape[0])
kinit = m.linear_growth_init(history)
elif m.growth == 'flat':
dat[f'cap_{i}'] = np.zeros(m.history.shape[0])
dat[f'cap_pred_{i}'] = np.zeros(only_future.shape[0])
kinit = m.flat_growth_init(history)
else:
dat[f'cap_{i}'] = history['cap_scaled']
dat[f'cap_pred_{i}'] = only_future['cap_scaled']
kinit = m.logistic_growth_init(history)
stan_init = {
f'k_{i}': kinit[0],
f'offset_{i}': kinit[1],
f'delta_{i}': np.zeros(len(m.changepoints_t)),
f'beta_{i}':
np.zeros(seasonal_features.shape[1] + len(parents)),
f'sigma_obs_{i}': 1,
}
all_dat.update(dat)
all_init.update(stan_init)
all_dat['n_samp'] = self.n_samp
self.dat = all_dat
model_code = self.generate_stan_code(graph)
model = pystan.StanModel(model_code=model_code)
fit = model.optimizing(data=all_dat, init=lambda: all_init, iter=1e4)
# Loop through nodes again to put forcast results back
for i in graph.nodes:
m = graph.nodes[i]["prophet"]
scale = m.y_scale
graph.nodes[i]["y_samples"] = fit[f"y_pred_{i}"] * scale
graph.nodes[i]["y_hat"] = fit[f"y_hat_{i}"] * scale
return (graph)
'gamma02': 4./5,\
'gamma03': 3./5,\
'gamma04': 2./5,\
'gamma05': 1./5,\
'mag_int_raw': mag_renorm, \
'L_Omega': numpy.identity(5), \
'Delta_unit':R_simplex, \
'Delta_scale': 15./4, \
'k_unit': R_simplex, \
'R_unit': R_simplex, \
'rho11': -17./5,\
'rho12': 0.*3./5,\
'rho13': 0./5,\
'rho14': 0.*3./5,\
'rho15': 0.*3./5,\
} \
for _ in range(8)]
sm = pystan.StanModel(file='gerard11.stan')
control = {'stepsize': 1}
fit = sm.sampling(data=data,
iter=5000,
chains=8,
control=control,
init=init,
thin=1)
output = open('temp11.pkl', 'wb')
pickle.dump((fit.extract(), fit.get_sampler_params()), output, protocol=2)
output.close()
print fit
def run_fir_hmc(data_path, input_order, prior='tc', hot_start=False):
""" Input order gives the terms # gives the terms b_0 * u_k + b_1 * u_{k-1} + .. + b_{input_order-1} * u_{k-input_order+1}"""
""" prior can be 'tc' for tuned correlated kernel or 'hs' for the horseshoe sparesness prior """
""" hot start will use least squares values as starting point """
data = loadmat(data_path)
y_est = data['y_estimation'].flatten()
u_est = data['u_estimation'].flatten()
y_val = data['y_validation'].flatten()
u_val = data['u_validation'].flatten()
# build regression matrix
est_input_matrix = build_input_matrix(u_est, input_order)
val_input_matrix = build_input_matrix(u_val, input_order)
# trim measurement vectors to suit regression matrix
max_delay = (input_order-1)
y_est = y_est[int(max_delay):]
y_val = y_val[int(max_delay):]
# calcualte an intial guess using least squares (ML)
if hot_start:
Ainv = np.linalg.pinv(est_input_matrix)
b_init = np.matmul(Ainv, y_est)
else:
b_init = np.zeros((input_order))
# Run Stan
def init_function():
sig_e = data['sig_e'].flatten()
output = dict(b_coefs=b_init * np.random.uniform(0.8, 1.2, len(b_init)),
sig_e=(sig_e * np.random.uniform(0.8, 1.2))[0],
b_coefs_hyperprior=np.abs(np.random.standard_cauchy(len(b_init))),
shrinkage_param=np.abs(np.random.standard_cauchy(1))[0]
)
return output
stan_data = {'input_order': int(input_order),
'no_obs_est': len(y_est),
'no_obs_val': len(y_val),
'y_est': y_est,
'est_input_matrix': est_input_matrix,
'val_input_matrix': val_input_matrix
}
# specify model file
if prior == 'hs':
model_path = 'stan/fir_hs.pkl'
if Path(model_path).is_file():
model = pickle.load(open(model_path, 'rb'))
else:
model = pystan.StanModel(file='stan/fir.stan')
with open(model_path, 'wb') as file:
pickle.dump(model, file)
elif prior == 'tc':
model_path = 'stan/fir_tc.pkl'
if Path(model_path).is_file():
model = pickle.load(open(model_path, 'rb'))
else:
model = pystan.StanModel(file='stan/fir_tc.stan')
with open(model_path, 'wb') as file:
pickle.dump(model, file)
else:
print("invalid prior, options are 'hs' or 'tc' ")
fit = model.sampling(data=stan_data, init=init_function, iter=6000, chains=4)
traces = fit.extract()
return (fit, traces)
real b;
}
model {
for (n in 1:N){
Y[n] ~ bernoulli_logit(a * X[n] + b);
}
}
"""
# +
# PyStanはコンパイルが必要（C++でコンパイルされる）
# 結構時間かかる
sm = pystan.StanModel(model_code= stan_model)
# +
# dataブロックに入れるデータを辞書型で渡す
stan_data = {"N":df.shape[0], "X":df["log10 C"], "Y":df["death"]}
# +
# MCMCでサンプリング
fit = sm.sampling(data = stan_data, iter = 2000, warmup=500, chains=3, seed=123)
# +
# 結果の抽出
# 事後分布の平均値や誤差が表示される
# Rhat<=1ならうまく収束している
shock_data['replicate_number'] = 0
data = pd.concat([mlg910, shock_data], ignore_index=True)
data.dropna(axis=1, inplace=True)
# Insert the shock rate identifier.
data.loc[:, 'idx'] = 0
# Rescale the intensities to meaningful values.
max_exp = data['exposure_ms'].max()
data['scaled_intensity'] = (data['intensity'] -
data['mean_bg']) * max_exp / data['exposure_ms']
data = data[data['scaled_intensity'] >= 0].copy()
# Load the stan model.
model = pystan.StanModel('../stan/complete_analysis.stan')
# Assemble the data dictionary.
data_dict = dict(
J1=6,
J2=1,
N1=len(data[data['rbs'] == 'mlg910']),
N2=len(data[data['rbs'] != 'mlg910']),
repl=data[data['replicate_number'] > 0]['replicate_number'].values.astype(
int),
shock=np.ones(len(data[data['rbs'] != 'mlg910'])).astype(int),
A=data[data['rbs'] == 'mlg910']['area'],
I_A_sc=data[data['rbs'] == 'mlg910']['scaled_intensity'],
I_A_sr=data[data['rbs'] != 'mlg910']['scaled_intensity'],
ntot=CHANNEL_NUM,
sigma_ntot=CHANNEL_SEM,
sales_df_5['date'] = sales_df_5['date'].view('int64') // 10**9
print(sales_df_5.head())
sales = sales_df_5['sales']
T = len(sales)
stan_data = {'T': T, 'y': sales}
filename = '5-7-2-autoregressive'
if os.path.exists('%s.pkl' % filename):
sm = pickle.load(open('%s.pkl' % filename, 'rb'))
# sm = pystan.StanModel(file='5-4-1-simple-reg.stan')
else:
# a model using prior for mu and sigma.
sm = pystan.StanModel(file='%s.stan' % filename)
control = {'adapt_delta': 0.8, 'max_treedepth': 16}
mcmc_result = sm.sampling(data=stan_data,
seed=1,
chains=4,
iter=8000,
warmup=2000,
control=control,
thin=6)
print(mcmc_result)
mcmc_result.plot()
plt.show()
import numpy as np
import os
import pickle
import matplotlib.pyplot as plt
import torch
import torch.distributions as dist
model_name = 'mix'
# Compile stan model, if needed. Otherwise, load model.
if os.path.exists('{}.pickle'.format(model_name)):
# Load model if it is cached.
sm = pickle.load(open('{}.pickle'.format(model_name), 'rb'))
else:
# compile model
sm = pystan.StanModel(file='{}.stan'.format(model_name))
# save model for later use.
with open('{}.pickle'.format(model_name), 'wb') as f:
pickle.dump(sm, f)
# Set random seed
np.random.seed(0)
# Simulate data
N = 500
mu_true = 2
sig_true = .3
y = np.random.randn(N) * sig_true + mu_true
# Fit STAN model (NUTS)
Kmcmc = 3
# # 'z0':mu0.flatten(),
# }
################ Coupled version
# state initialisation point
z_init = np.zeros((4, no_obs + 1))
z_init[0, :-1] = y[0, :]
z_init[1, :-1] = y[1, :]
z_init[0, -1] = y[0, -1] # repeat last entry
z_init[1, -1] = y[1, -1] # repeat last entry
z_init[2, :-2] = (y[0, 1:] - y[0, 0:-1]) / Ts
z_init[2, -1:] = z_init[2, -3]
z_init[3, :-2] = (y[1, 1:] - y[1, 0:-1]) / Ts
z_init[3, -1:] = z_init[3, -3]
model = pystan.StanModel(file='stan/pendulum_coupled_noprior.stan')
mu0 = np.zeros((4, ))
cP0 = np.array(
[np.deg2rad(10),
np.deg2rad(10),
np.deg2rad(100),
np.deg2rad(100)])
stan_data = {
'no_obs': no_obs,
'Ts': Ts[0, 0],
'y': y,
'u': u.flatten(),
'Lr': Lr,
'Mp': Mp,
'Lp': Lp,
'g': g,
ind = np.argmax(mag)
omega_init = omegas[ind]
A_init = np.round(A[ind] / N * 100) / 100
B_init = np.round(B[ind] / N * 100) / 100
plt.plot(omegas, mag)
plt.title('Peak: omega = ' + str(omega_init) + ' A = ' + str(A_init) +
' B = ' + str(B_init))
plt.show()
model_path = 'model.pkl'
if Path(model_path).is_file():
model = pickle.load(open(model_path, 'rb'))
else:
model = pystan.StanModel(file=model_path[:-4] + '.stan')
with open(model_path, 'wb') as file:
pickle.dump(model, file)
# model = pystan.StanModel(file='model.stan')
def init_function():
output = dict(alpha=omega_init * np.random.uniform(0.8, 1.2))
return output
stan_data = {'N': N, 'x': x, 'y': y, 'lambda1': lambda1}
fit = model.sampling(
data=stan_data, init=init_function, iter=6000, warmup=4000,
chains=4) #, control=dict(adapt_delta=0.9, max_treedepth=13))
range(1, 1 + len(observedLanguages)))
dat["Total2Hidden"] = [1] + list(range(
2, 2 + len(hiddenLanguages))) + [0 for _ in observedLanguages]
dat["ParentDistance"] = [0] + [
distanceToParent[x] for x in hiddenLanguages + observedLanguages
]
dat["CovarianceMatrix"] = covarianceMatrix
dat["prior_only"] = 0
dat["Components"] = 2
dat["FamiliesLists"] = familiesLists
dat["FamiliesNum"] = len(familiesLists)
dat["FamiliesSize"] = len(familiesLists[0])
print(dat)
sm = pystan.StanModel(file=f'{__file__[:-3]}.stan')
#stepping = [0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0]
stepping = [0.0, 0.01, 0.02, 0.05, 0.08, 0.1, 0.2, 0.3, 0.4, 0.7, 1.0]
#stepping = [0.0, 0.005, 0.01, 0.015, 0.02, 0.025, 0.03, 0.035, 0.04, 0.045, 0.05, 0.06, 0.075, 0.08, 0.085, 0.09, 0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.5,0.6, 0.7, 0.8, 1.0]
def mean(x):
return sum(x) / len(x)
import torch
for replicate in range(10):
perStone = []
for idx in range(len(stepping) - 1):
'n2':hidden_dim1,
'm1':z_dim,
'N': 150,
'x':standardized_data.T,
'W1': decoder_dict['linear1.weight'].T.numpy(),
'b1': decoder_dict['linear1.bias'].T.numpy().reshape(1,hidden_dim2),
'W2': decoder_dict['linear2.weight'].T.numpy(),
'b2': decoder_dict['linear2.bias'].T.numpy().reshape(1,hidden_dim1),
'W3': decoder_dict['out1.weight'].T.numpy(),
'b3': decoder_dict['out1.bias'].T.numpy().reshape(1,input_dim),
'cov1': np.identity(input_dim)*9,
'mu1': np.array([0]*input_dim),
'cov2': np.identity(input_dim)*0.25,
'mu2': np.array([0]*input_dim)*0}
#### stan code
sm = pystan.StanModel(file='iris_beta.stan')
fit = sm.sampling(data=stan_data, iter=200, warmup=20, chains=3)
out = fit.extract(permuted=True)
print(fit)
mean_pi = out['pi'].mean(axis=0)
mean_mu = out['theta_mu_star'].mean(axis=0)
mean_sd = out['theta_sd_star'].mean(axis=0)
post_pdf_contribs = scipy.stats.norm.pdf(np.atleast_3d(standardized_data),mean_mu,mean_sd[np.newaxis,:,:])
post_pdf_contribs=post_pdf_contribs.prod(axis=1)*mean_pi[np.newaxis,:]
cluster_2=np.argmax(post_pdf_contribs,axis=1)+1
standardized_iris["Cluster_2"]=cluster_2
sns.set(font_scale=2)
g=sns.pairplot(standardized_iris,vars=['Sepal Length', 'Sepal Width', 'Petal Length','Petal Width'],hue="Cluster_2",height=4)
g._legend.set_title("Cluster")
