0

我正在尝试使用 Spyder 在 Win10 环境中运行带有 pymc3 的分层模型。我有一些全局模型参数（theta、omega、sigma）和一个特定参数（Ci）。

它需要一个 pd Dataframe 作为包含相关数据的输入。第一列称为“队列”，第二列称为“时期”，第三列包含观察结果。

观察的数量在队列之间有所不同。

代码如下所示：

import pymc3 as pm
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import theano.tensor as tt 

cohorts_idx, cohorts = pd.factorize(inputs['Cohort'], sort = True)
periods_idx, periods = pd.factorize(inputs['Period'], sort = True)

coords = {
    "cohort": cohorts,
    "period": periods,
    "collections": np.arange(len(cohorts_idx))
    }

with pm.Model(coords = coords) as model:
    
    # global model parameters
    omega = pm.HalfNormal("omega", sigma = 3)
    theta = pm.HalfNormal("theta", sigma = 5)
    sigma = pm.HalfNormal("sigma", sigma = 20)
            
    # cohort specific parameter
    Ci = pm.TruncatedNormal("Ci", mu = 60, sigma = 10, lower = 10, upper = 110, dims = "cohort")
    

    mu_i_t = Ci[cohorts_idx] * (1 - tt.exp(- (periods[periods_idx] / theta) ** omega))
    sigma_i_t = sigma * mu_i_t ** 0.5
    
    _ = pm.Normal("Collections_i_t",
                                mu = mu_i_t,
                                sigma = sigma_i_t,            
                                observed = inputs['Collections'],
                                dims = "collections")
         
    results = pm.sample(draws = 1000, tune = 1000, cores = 8)
 
    return pm.summary(results)

产生的错误信息是：

mu_i_t = Ci[cohorts_idx] * (1 - tt.exp(- (periods[periods_idx] / theta) ** omega))

File "C:\Users\alexi\Anaconda3\lib\site-packages\pandas\core\indexes\base.py", line 139, in 
index_arithmetic_method
result = op(Series(self), other)

File "C:\Users\alexi\Anaconda3\lib\site-packages\pandas\core\ops\common.py", line 64, in new_method
return method(self, other)

File "C:\Users\alexi\Anaconda3\lib\site-packages\pandas\core\ops\__init__.py", line 505, in wrapper
return _construct_result(left, result, index=left.index, name=res_name)

File "C:\Users\alexi\Anaconda3\lib\site-packages\pandas\core\ops\__init__.py", line 478, in _ 
construct_result
out = left._constructor(result, index=index)

File "C:\Users\alexi\Anaconda3\lib\site-packages\pandas\core\series.py", line 279, in __init__
data = com.maybe_iterable_to_list(data)

File "C:\Users\alexi\Anaconda3\lib\site-packages\pandas\core\common.py", line 280, in 
maybe_iterable_to_list
return list(obj)

File "C:\Users\alexi\Anaconda3\lib\site-packages\theano\tensor\var.py", line 640, in __iter__
for i in xrange(theano.tensor.basic.get_vector_length(self)):

File "C:\Users\alexi\Anaconda3\lib\site-packages\theano\tensor\basic.py", line 4828, in 
get_vector_length
raise ValueError("length not known: %s" % msg)

ValueError: length not known: Elemwise{true_div,no_inplace} [id A] ''   
|TensorConstant{[ 1  2  3 ..  1  2  1]} [id B]
|InplaceDimShuffle{x} [id C] ''   
|ViewOp [id D] 'theta'   
 |Elemwise{exp,no_inplace} [id E] ''   
   |theta_log__ [id F]

我不知道为什么。请注意，如果不是在导致错误的行中除以 Theta，而是进行加法，那么它可以工作（但显然这不是我想要的）。

我该如何解决这个问题并让这个部门工作？

4

1 回答 1

0

好的，我发现了。我不足以解释原因，但以下工作。需要更换线路：

mu_i_t = Ci[cohorts_idx] * (1 - tt.exp(- (periods[periods_idx] / theta) ** omega))

经过：

mu_i_t = Ci[cohorts_idx] * (1 - tt.exp(- (periods[periods_idx] .to_numpy() / theta) ** omega))

于 2020-10-31T15:01:09.877 回答