我在这里发布了一个 python 笔记本:http: //nbviewer.ipython.org/gist/awellis/9067358
我正在尝试使用 PyMC 3 创建一个概率回归模型,使用生成的数据来恢复已知参数(参见笔记本)。截距的估计值差不多,但斜率估计值太离谱了。
我的模型如下所示:
with pm.Model() as model:
# priors
alpha = pm.Normal('alpha', mu=0, tau=0.001)
beta = pm.Normal('beta', mu=0, tau=0.001)
# linear predictor
theta_p = (alpha + beta * x)
# logic transform (just for comparison - this seems to work ok)
# def invlogit(x):
# import theano.tensor as t
# return t.exp(x) / (1 + t.exp(x))
# theta = invlogit(theta_p)
# Probit transform: this doesn't work
def phi(x):
import theano.tensor as t
return 0.5 * (1 + t.erf(x / t.sqr(2)))
theta = phi(theta_p)
# likelihood
y = pm.Bernoulli('y', p=theta, observed=y)
with model:
# Inference
start = pm.find_MAP() # Find starting value by optimization
print("MAP found:")
print("alpha:", start['alpha'])
print("beta:", start['beta'])
print("Compare with true values:")
print("true_alpha", true_alpha)
print("true_beta", true_beta)
with model:
step = pm.NUTS()
trace = pm.sample(2000,
step,
start=start,
progressbar=True) # draw posterior samples
它似乎工作的唯一方法是使用 Theano 定义 phi(x),使用错误函数,类似于 PyMC 存储库中的逻辑回归示例。
谁能指出我正确的方向?有没有更好/更简单的方法来做到这一点?