我正在尝试使用SciPy的scipy.optimize.minimize函数来最小化我创建的函数。但是，我尝试对其进行优化的功能本身是由其他基于熊猫DataFrame执行计算的功能构成的。

我了解SciPy的最小化函数可以通过元组输入多个参数（例如，scipy最小化函数的输入结构）。但是，我不知道如何传递依赖于熊猫DataFrame的函数。

我在下面创建了一个可复制的示例。

import pandas as pd
import numpy as np
from scipy.stats import norm
from scipy.optimize import minimize


####################     Data     ####################
# Initialize dataframe. 
data = pd.DataFrame({'id_i': ['AA','BB','CC','XX','DD'],'id_j': ['ZZ','YY','AA'],'y': [0.30,0.60,0.70,0.45,0.65],'num': [1000,2000,1500,1200,1700],'bar': [-4.0,-6.5,1.0,-3.0,-5.5],'mu': [-4.261140,-5.929608,1.546283,-1.810941,-3.186412]})

data['foo_1'] = data['bar'] - 11 * norm.ppf(1/1.9)
data['foo_2'] = data['bar'] - 11 * norm.ppf(1 - (1/1.9))

# Store list of ids.
id_list = sorted(pd.unique(pd.concat([data['id_i'],data['id_j']],axis=0)))


####################     Functions     ####################
# Function 1: Intermediate calculation to calculate predicted values.
def calculate_y_pred(row,delta_params,sigma_param,id_list):

    # Extract the relevant values from delta_params.
    delta_i = delta_params[id_list.index(row['id_i'])]
    delta_j = delta_params[id_list.index(row['id_j'])]

    # Calculate adjusted version of mu. 
    mu_adj = row['mu'] - delta_i + delta_j

    # Calculate predicted value of y.
    y_pred = norm.cdf(row['foo_1'],loc=mu_adj,scale=sigma_param) / \
                (norm.cdf(row['foo_1'],scale=sigma_param) + 
                    (1 - norm.cdf(row['foo_2'],scale=sigma_param)))

    return y_pred

# Function to calculate the log-likelihood (for a row of DataFrame data).
def loglik_row(row,id_list):

    # Calculate the log-likelihood for this row.
    y_pred = calculate_y_pred(row,id_list)
    y_obs = row['y']
    n = row['num']
    loglik_row = np.log(norm.pdf(((y_obs - y_pred) * np.sqrt(n)) / np.sqrt(y_pred * (1-y_pred))) / 
                            np.sqrt(y_pred * (1-y_pred) / n))

    return loglik_row

# Function to calculate the sum of the negative log-likelihood. 
# This function is called via SciPy's minimize function.
def loglik_total(data,id_list,params):

    # Extract parameters.
    delta_params = list(params[0:len(id_list)])
    sigma_param = init_params[-1]

    # Calculate the negative log-likelihood for every row in data and sum the values.
    loglik_total = -np.sum( data.apply(lambda row: loglik_row(row,id_list),axis=1) )

    return loglik_total


####################     Optimize     ####################
# Provide initial parameter guesses. 
delta_params = [0 for id in id_list]
sigma_param = 11
init_params = tuple(delta_params + [sigma_param])

# Maximize the log likelihood (minimize the negative log likelihood). 
minimize(fun=loglik_total,x0=init_params,args=(data,method='nelder-mead')

这将导致以下错误：（AttributeError: 'numpy.ndarray' object has no attribute 'apply'整个错误输出在下面）。我相信这个错误是因为minimize将其X视为numpy数组，而我想将其作为pandas
DataFrame传递。

AttributeError: 'numpy.ndarray' object has no attribute 'apply'
AttributeErrorTraceback (most recent call last)
<ipython-input-93-9a5866bd626e> in <module>()
      1 minimize(fun=loglik_total,----> 2             args=(data,method='nelder-mead')
/Users/adam/anaconda/lib/python2.7/site-packages/scipy/optimize/_minimize.pyc in minimize(fun,x0,args,method,jac,hess,hessp,bounds,constraints,tol,callback,options)
    436                       callback=callback,**options)
    437     elif meth == 'nelder-mead':
--> 438         return _minimize_neldermead(fun,**options)
    439     elif meth == 'powell':
    440         return _minimize_powell(fun,**options)
/Users/adam/anaconda/lib/python2.7/site-packages/scipy/optimize/optimize.pyc in _minimize_neldermead(func,maxiter,maxfev,disp,return_all,initial_simplex,xatol,fatol,**unknown_options)
    515 
    516     for k in range(N + 1):
--> 517         fsim[k] = func(sim[k])
    518 
    519     ind = numpy.argsort(fsim)
/Users/adam/anaconda/lib/python2.7/site-packages/scipy/optimize/optimize.pyc in function_wrapper(*wrapper_args)
    290     def function_wrapper(*wrapper_args):
    291         ncalls[0] += 1
--> 292         return function(*(wrapper_args + args))
    293 
    294     return ncalls,function_wrapper
<ipython-input-69-546e169fc54e> in loglik_total(data,params)
      6 
      7     # Calculate the negative log-likelihood for every row in data and sum the values.
----> 8     loglik_total = -np.sum( data.apply(lambda row: loglik_row(row,axis=1) )
      9 
     10     return loglik_total
AttributeError: 'numpy.ndarray' object has no attribute 'apply'

在SciPy的函数中处理DataFramedata并调用我的函数的正确方法是什么？任何建议都欢迎，我们将不胜感激。loglik_total``minimize

可能的解决方案： 注意，我考虑过可以编辑函数以将其data视为numpy数组，而不是pandas
DataFrame。但是，由于以下两个原因，我想避免这种情况：1）中loglik_total，我使用pandasapply函数将loglik_row函数应用于的每一行data；2）data通过其列名而不是数字索引来引用列很方便。

当在模型.predict()中使用optimizer.minimize()来训练tensorflow.js中具有损失函数的模型时,我遇到了一个问题.仅当我在卷积神经网络中使用maxPooling2D层且代码类似于以下代码时,才会发生这种情况.产生此错误：无法读取未定义的属性“后端”.我不确定是什么原因引起的或如何解决.当使用没有任何池化层的卷积层(tf.layers.conv2d())时,不会发生此错误.我正在使用tensorflow.js版本0.14.2和Google Chrome版本71.0.3578.98.可以使用以下代码来重现此错误：

loss = (pred,label) => pred.sub(label).square().mean();
optimizer = tf.train.sgd(0.001);

const input = tf.input({shape: [100,100,4]});
const conv = tf.layers.conv2d({
    kernelSize: 5,filters: 8,strides: 1,activation: 'relu',kernelInitializer: 'VarianceScaling'
});
const pool = tf.layers.maxPooling2d({
    poolSize: [2,2],strides: [2,2]
});
const flat = tf.layers.flatten();
const dense = tf.layers.dense({units: 10});
const output = dense.apply(flat.apply(pool.apply(conv.apply(input))));
const model = tf.model({inputs: input,outputs: output});

for (var i = 0; i < 10; i++) {
    optimizer.minimize(() =>
        loss(model.predict([tf.ones([1,4])]),tf.ones([1,10]))
    );
}
@H_301_8@
编辑：这已经解决.有关详细信息,请参见scai’s answer.
编辑2：这似乎不是错误,而是对model.predict()的反向传播进行了更改. More information

即改变

    optimizer.minimize(() =>
    loss(model.predict([tf.ones([1,10]))
);
@H_301_8@
至

   optimizer.minimize(() =>
    loss(model.apply([tf.ones([1,4])],{training: true}),10]))
);
@H_301_8@