使用NMF分解混合光谱时fit_transform报ValueError的解决请求

问题场景 #

用非负矩阵分解（NMF）拆分混合化合物光谱，mixed_spectrum为混合光谱的numpy数组，W_init包含已知纯化合物光谱列向量及未知成分的初始值，运行代码时触发维度错误。

原代码 #

import numpy as np
from sklearn.decomposition import NMF

pure_compounds /= np.max(pure_compounds)
n_components = len(pure_compounds) + 1
model = NMF(n_components=n_components, init='custom')
W_init = np.concatenate((pure_compounds, np.ones((pure_compounds.shape[0], 1))), axis=1)
print(W_init)
print(np.shape(W_init))
mix = mixed_spectrum.reshape(-1, 1)
print(mix)
print(np.shape(mix))
W = model.fit_transform(mix, W=W_init)
H = model.components_
for i in range(n_components):
    coeff = W[:, i] / np.sum(W, axis=1)
    print('Concentration coefficients for compound', i, ':', coeff)

错误输出 #

[[0.00000000e+00 2.76922293e-04 1.44273203e-03 0.00000000e+00
  1.00000000e+00]
 [0.00000000e+00 2.85985939e-04 1.40933148e-03 0.00000000e+00
  1.00000000e+00]
 [0.00000000e+00 2.54579755e-04 1.52763958e-03 0.00000000e+00
  1.00000000e+00]
 ...
 [1.00000000e-06 1.00000000e-06 1.00000000e-06 1.00000000e-06
  1.00000000e+00]
 [1.00000000e-06 1.00000000e-06 1.00000000e-06 1.00000000e-06
  1.00000000e+00]
 [1.00000000e-06 1.00000000e-06 1.00000000e-06 1.00000000e+00]]
(8954, 5)
[[2.31330983e-02]
 [1.00000000e-06]
 [1.61938395e-02]
 ...
 [2.42525353e-02]
 [1.95440454e-02]
 [2.98149646e-03]]
(8954, 1)

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
~\AppData\Local\Temp\ipykernel_6528\919808671.py in <module>
      8 print(mix)
      9 print(np.shape(mix))
---> 10 W = model.fit_transform(mix, W=W_init)
     11 H = model.components_
     12 for i in range(n_components):

~\Anaconda3\lib\site-packages\sklearn\decomposition\_nmf.py in fit_transform(self, X, y, W, H)
   1536 
   1537         with config_context(assume_finite=True):
-> 1538             W, H, n_iter = self._fit_transform(X, W=W, H=H)
   1539 
   1540         self.reconstruction_err_ = _beta_divergence(

~\Anaconda3\lib\site-packages\sklearn\decomposition\_nmf.py in _fit_transform(self, X, y, W, H, update_H)
   1595 
   1596         # initialize or check W and H
-> 1597         W, H = self._check_w_h(X, W, H, update_H)
   1598 
   1599         # scale the regularization terms

~\Anaconda3\lib\site-packages\sklearn\decomposition\_nmf.py in _check_w_h(self, X, W, H, update_H)
   1460         n_samples, n_features = X.shape
   1461         if self.init == "custom" and update_H:
-> 1462             _check_init(H, (self._n_components, n_features), "NMF (input H)")
   1463             _check_init(W, (n_samples, self._n_components), "NMF (input W)")
   1464             if H.dtype != X.dtype or W.dtype != X.dtype:

~\Anaconda3\lib\site-packages\sklearn\decomposition\_nmf.py in _check_init(A, shape, whom)
     52 
     53 def _check_init(A, shape, whom):
-> 54     A = check_array(A)
     55     if np.shape(A) != shape:
     56         raise ValueError(

~\Anaconda3\lib\site-packages\sklearn\utils\validation.py in check_array(array, accept_sparse, accept_large_sparse, dtype, order, copy, force_all_finite, ensure_2d, allow_nd, ensure_min_samples, ensure_min_features, estimator)
    759             # If input is scalar raise error
    760             if array.ndim == 0:
-> 761                 raise ValueError(
    762                     "Expected 2D array, got scalar array instead:\narray={}.\n"
    763                     "Reshape your data either using array.reshape(-1, 1) if "

ValueError: Expected 2D array, got scalar array instead:
array=None.
Reshape your data either using array.reshape(-1, 1) if your data has a single feature or array.reshape(1, -1) if it contains a single sample.

修复方案 #

报错核心原因：当NMF设置init='custom'时，scikit-learn要求同时传入W和H的初始值，仅传W_init会导致H默认值为None，触发维度校验错误。

修改步骤：

• 初始化H的初始矩阵，维度需匹配(n_components, n_features)，其中n_features是输入X的特征数（此处mix的shape为(8954,1)，故n_features=1），用极小非负值初始化避免零值问题。
• 在fit_transform中同时传入W和H的初始值。

修改后的代码：

import numpy as np
from sklearn.decomposition import NMF

pure_compounds /= np.max(pure_compounds)
n_components = len(pure_compounds) + 1
model = NMF(n_components=n_components, init='custom')
W_init = np.concatenate((pure_compounds, np.ones((pure_compounds.shape[0], 1))), axis=1)
# 初始化H的初始矩阵，维度与模型要求一致
H_init = np.ones((n_components, 1)) * 1e-6
print(W_init.shape)
mix = mixed_spectrum.reshape(-1, 1)
print(mix.shape)
# 同时传入W和H的初始值
W = model.fit_transform(mix, W=W_init, H=H_init)
H = model.components_
for i in range(n_components):
    coeff = W[:, i] / np.sum(W, axis=1)
    print('Concentration coefficients for compound', i, ':', coeff)

额外注意：

• 你的W_init维度为(8954,5)，与n_components=5匹配，这部分无需调整。
• 若无需更新H，可设置update_H=False，但该场景仅适用于已知H真实值的情况，不符合当前需求，因此初始化H并传入是更合理的选择。

内容的提问来源于stack exchange，提问作者gionti