2

我在 Python ( mkl_regressor) 中有一个自定义的估算器对象。这种对象的学习参数之一是numpy.array浮点数。通常 sklearn 估计器对象由单个参数调整，例如CSVM 的。因此，randomizedSearchCV搜索对象采用一个分布或一个值列表，用于从给定分布（在我的示例中为scipy.stats.expon）中获取所需参数的某个值。我试图传递分布列表，但没有成功，因为randomizedSearchCV不执行分布数组中的元素。这是我尝试过的：

from modshogun import *
import Gnuplot, Gnuplot.funcutils
from numpy import *
from sklearn.metrics import r2_score

class mkl_regressor():

    def __init__(self, widths = [0.01, 0.1, 1.0, 10.0, 50.0, 100.0], kernel_weights = [0.01, 0.1, 1.0,], svm_c = 0.01, mkl_c = 1.0, svm_norm = 1, mkl_norm = 1, degree = 2):
        self.svm_c = svm_c
        self.mkl_c = mkl_c
        self.svm_norm = svm_norm
        self.mkl_norm = mkl_norm
        self.degree = degree
        self.widths = widths
        self.kernel_weights = kernel_weights


    def fit(self, X, y, **params):
        for parameter, value in params.items():
            setattr(self, parameter, value)        

        self.feats_train = RealFeatures(X.T)
        labels_train = RegressionLabels(y.reshape((len(y), )))
        self._kernels_  = CombinedKernel()
        for width in self.widths:
            kernel = GaussianKernel()
            kernel.set_width(width)
            kernel.init(self.feats_train,self.feats_train)
            self._kernels_.append_kernel(kernel)
            del kernel

        kernel = PolyKernel(10, self.degree)            
        self._kernels_.append_kernel(kernel)
        del kernel

        self._kernels_.init(self.feats_train, self.feats_train)

        binary_svm_solver = SVRLight()
        self.mkl = MKLRegression(binary_svm_solver)

        self.mkl.set_C(self.svm_c, self.svm_c)
        self.mkl.set_C_mkl(self.mkl_c)
        self.mkl.set_mkl_norm(self.mkl_norm)
        self.mkl.set_mkl_block_norm(self.svm_norm)

        self.mkl.set_kernel(self._kernels_)
        self.mkl.set_labels(labels_train)
        self.mkl.train()
        self.kernel_weights = self._kernels_.get_subkernel_weights()

    def predict(self, X):
        self.feats_test = RealFeatures(X.T)
        self._kernels_.init(self.feats_train, self.feats_test) 
        self.mkl.set_kernel(self._kernels_)
        return self.mkl.apply_regression().get_labels()

    def set_params(self, **params):
        for parameter, value in params.items():
            setattr(self, parameter, value)

        return self

    def get_params(self, deep=False):

        return {param: getattr(self, param) for param in dir(self) if not param.startswith('__') and not callable(getattr(self,param))}    

    def score(self,  X_t, y_t):

        predicted = self.predict(X_t)
        return r2_score(predicted, y_t)    

if __name__ == "__main__":

    from sklearn.grid_search import RandomizedSearchCV as RS
    from scipy.stats import randint as sp_randint
    from scipy.stats import expon

    labels = array([2.0,0.0,2.0,1.0,3.0,2.0])
    labels = labels.reshape((len(labels), 1))
    data = array([[1.0,2.0,3.0],[1.0,2.0,9.0],[1.0,2.0,3.0],[1.0,2.0,0.0],[0.0,2.0,3.0],[1.0,2.0,3.0]])
    labels_t = array([1.,3.,4])
    labels_t = labels_t.reshape((len(labels_t), 1))
    data_t = array([[20.0,30.0,40.0],[10.0,20.0,30.0],[10.0,20.0,40.0]])
    k = 3

    param_grid = [ {'svm_c': expon(scale=100, loc=5),
                'mkl_c': expon(scale=100, loc=5),
                'degree': sp_randint(0, 32),
                #'widths': [array([4.0,6.0,8.9,3.0]), array([4.0,6.0,8.9,3.0,2.0, 3.0, 4.0]), array( [100.0, 200.0, 300.0, 400.0]) 
                'widths': [[expon, expon]] 
              }]

    mkl = mkl_regressor()
    rs = RS(mkl, param_distributions = param_grid[0], n_iter = 10, n_jobs = 24, cv = k)#, scoring="r2", verbose=True)
    rs.fit(data, labels)
    preds = rs.predict(data_t)

    print "R^2: ", rs.score(data_t, labels_t)
    print "Parameters: ", rs.best_params_

上面的代码通过将 numpy 数组作为'widths'参数字典列表的元素传递，效果很好。但是，当我尝试传递分布列表时，随机搜索CV 对象不会按预期响应：

/home/ignacio/distributionalSemanticStabilityThesis/mkl_test.py in fit(self=<__main__.mkl_regressor instance>, X=array([[ 1.,  2.,  3.],
       [ 1.,  2.,  0.],
       [ 0.,  2.,  3.],
       [ 1.,  2.,  3.]]), y=array([[ 2.],
       [ 1.],
       [ 3.],
       [ 2.]]), **params={})
     24         self.feats_train = RealFeatures(X.T)
     25         labels_train = RegressionLabels(y.reshape((len(y), )))
     26         self._kernels_  = CombinedKernel()
     27         for width in self.widths:
     28             kernel = GaussianKernel()
---> 29             kernel.set_width(width)
        kernel.set_width = <built-in method set_width of GaussianKernel object>
        width = <scipy.stats._continuous_distns.expon_gen object>
     30             kernel.init(self.feats_train,self.feats_train)
     31             self._kernels_.append_kernel(kernel)
     32             del kernel
     33 

TypeError: in method 'GaussianKernel_set_width', argument 2 of type 'float64_t'

我不想强制估计器执行每个分布生成器，因为在这种情况下，randomizedSearchCV将无法控制使用的值。

一些建议？谢谢你。

4

2 回答 2

1

RandomizedSearchCV 可以采用参数值列表进行尝试，也可以采用具有 rvs 方法的分布对象进行采样。如果您向它传递一个列表，它将假定您传递了一组离散的参数值以从中进行采样。它不支持单个参数的分布列表。如果现有发行版不适合您的需求，请制作自定义发行版。

如果您需要一个返回数组的分布，只需创建一个具有 rvs() 方法的类来返回一个随机样本并传递一个实例而不是单变量分布列表。

于 2016-07-30T01:44:02.810 回答
1

@bachev 建议的解决方案对我有用。分布类：

class expon_vector(stats.rv_continuous):

    def __init__(self, loc = 1.0, scale = 50.0, min_size=2, max_size = 10):
        self.loc = loc
        self.scale = scale
        self.min_size = min_size
        self.max_size = max_size
        self.size = max_size - min_size # Only for initialization

    def rvs(self):

        self.size = randint.rvs(low = self.min_size, 
                                high = self.max_size, size = 1)
        return expon.rvs(loc  = self.loc, scale = self.scale, size = self.size)

这包含在我正在使用的自定义估计器的参数字典中：

param_grid = [ {'svm_c': expon(scale=100, loc=5),
                    'mkl_c': expon(scale=100, loc=5),
                    'degree': sp_randint(0, 24),
                    'widths': expon_vector(loc = 0.1, scale = 100.0, 
                              min_size = 2, max_size = 10) } ]
于 2016-07-30T17:53:34.833 回答