如何将两个特征/分类器组合成一个统一且更好的分类器？
Posted
技术标签:
【中文标题】如何将两个特征/分类器组合成一个统一且更好的分类器？【英文标题】：How do I combine two features / classifiers into a single united and better classifier? 【发布时间】：2015-09-18 21:37:12 【问题描述】：我正在创建一个情绪识别程序，并设法生成两种不同的算法/特征来输入 sklearn 的 SVM。我得到了密集光流数据，将其压缩成一个矩阵并将其输入到 SVM 函数中，同时我对来自面部地标的跟踪数据做了同样的事情。
现在，我有两个不同的程序，每个程序产生不同的准确性，但做的事情相同：根据面部动作识别情绪。
我现在的目标是将密集光流和面部地标特征/分类器结合在一起，并将它们结合起来以获得更好的分类器，该分类器将使用这两者来实现更高的分类准确度。
基本上，我正在尝试从中重新创建一个分类器： http://cgit.nutn.edu.tw:8080/cgit/PaperDL/LZJ_120826151743.PDF
Confusion Matrix for Dense Optical Flow:
[[27 22 0 0]
[ 0 57 1 0]
[ 0 12 60 0]
[ 0 9 3 68]]
Accuracy: 80-90% range
Confusion Matrix for Facial Landmarks:
[[27 10 5 2]
[ 7 44 5 3]
[ 6 14 33 1]
[ 1 13 1 60]]
Accuracy: 60-72% range
包含密集光流数据的矩阵的矩阵结构：
>>> main.shape
(646, 403680)
>>> main
array([[ -1.18353125e-03, -2.41295085e-04, -1.88367767e-03, ...,
-5.19892928e-05, 8.53588153e-06, -3.90818786e-05],
[ 6.32877424e-02, -7.24349543e-02, 8.19472596e-02, ...,
-4.71765925e-05, 5.41217596e-05, -3.12083102e-05],
[ -1.66368652e-02, 2.50510368e-02, -6.03965335e-02, ...,
-9.85100851e-05, -7.69595645e-05, -7.09727174e-05],
...,
[ -3.44874617e-03, 5.31123485e-03, -8.47499538e-03, ...,
-2.77953018e-06, -2.96417579e-06, -1.51305017e-06],
[ 3.24894954e-03, 5.05338283e-03, 3.91049543e-03, ...,
-3.23493354e-04, 1.30995919e-04, -3.06804082e-04],
[ 7.82454386e-03, 1.69946514e-02, 8.11014231e-03, ...,
-1.02751539e-03, 7.68289610e-05, -7.82517891e-04]], dtype=float32)
包含人脸地标跟踪信息的矩阵结构：
>>> main.shape
(646, 17, 68, 2)
>>> main
array([[[[ 0. , 0. ],
[ 0. , 0. ],
[ 0. , 0. ],
...,
[ 0. , 0. ],
[ 0. , 0. ],
[ 0. , 0. ]],
[[ -2.23606798, -1.10714872],
[ -2.23606798, -1.10714872],
[ 3. , 1. ],
...,
[ 1.41421356, 0.78539816],
[ 1.41421356, 0.78539816],
[ 1. , 0. ]],
[[ 2.82842712, -0.78539816],
[ 2.23606798, -1.10714872],
[ 2.23606798, -1.10714872],
...,
[ -1. , -0. ],
[ -1. , -0. ],
[ -1. , -0. ]],
...,
[[ 2. , 1. ],
[ -2.23606798, 1.10714872],
[ -3.16227766, 1.24904577],
...,
[ -1. , -0. ],
[ -1.41421356, 0.78539816],
[ -1. , -0. ]],
[[ -1.41421356, -0.78539816],
[ 1. , 1. ],
[ -1.41421356, -0.78539816],
...,
[ 0. , 0. ],
[ 0. , 0. ],
[ 0. , 0. ]],
[[ 3. , 1. ],
[ 4. , 1. ],
[ 4. , 1. ],
...,
[ 1.41421356, -0.78539816],
[ 1. , 0. ],
[ 1. , 0. ]]],
[[[ 0. , 0. ],
[ 0. , 0. ],
[ 0. , 0. ],
...,
[ 0. , 0. ],
[ 0. , 0. ],
[ 0. , 0. ]],
[[ 1. , 1. ],
[ -1.41421356, -0.78539816],
[ -1. , -0. ],
...,
[ 2. , 0. ],
[ 1. , 0. ],
[ -1. , -0. ]],
[[ 0. , 0. ],
[ 1. , 1. ],
[ 0. , 0. ],
...,
[ -4. , -0. ],
[ -3. , -0. ],
[ -2. , -0. ]],
...,
[[ -2.23606798, -1.10714872],
[ -2.23606798, -1.10714872],
[ 2. , 1. ],
...,
[ 0. , 0. ],
[ 1.41421356, 0.78539816],
[ 1.41421356, 0.78539816]],
[[ 0. , 0. ],
[ 0. , 0. ],
[ -1. , -0. ],
...,
[ 1. , 1. ],
[ 0. , 0. ],
[ -1.41421356, 0.78539816]],
[[ 1. , 1. ],
[ 1. , 1. ],
[ 1. , 1. ],
...,
[ 1. , 1. ],
[ 0. , 0. ],
[ 1. , 0. ]]],
[[[ 0. , 0. ],
[ 0. , 0. ],
[ 0. , 0. ],
...,
[ 0. , 0. ],
[ 0. , 0. ],
[ 0. , 0. ]],
[[ 3.16227766, 1.24904577],
[ 2.23606798, 1.10714872],
[ 2.23606798, 1.10714872],
...,
[ -1.41421356, -0.78539816],
[ -1. , -0. ],
[ -1.41421356, 0.78539816]],
[[ -1.41421356, 0.78539816],
[ 0. , 0. ],
[ 1.41421356, 0.78539816],
...,
[ -1.41421356, 0.78539816],
[ -1. , -0. ],
[ 0. , 0. ]],
...,
[[ 1. , 1. ],
[ 1. , 1. ],
[ 0. , 0. ],
...,
[ 1. , 1. ],
[ 1. , 1. ],
[ -1.41421356, 0.78539816]],
[[ 1. , 1. ],
[ 2. , 1. ],
[ 2.23606798, 1.10714872],
...,
[ 1. , 1. ],
[ 1. , 1. ],
[ -1.41421356, -0.78539816]],
[[ 1. , 1. ],
[ 1. , 1. ],
[ 1. , 1. ],
...,
[ -2. , -0. ],
[ -2. , -0. ],
[ -1. , -0. ]]],
...,
[[[ 0. , 0. ],
[ 0. , 0. ],
[ 0. , 0. ],
...,
[ 0. , 0. ],
[ 0. , 0. ],
[ 0. , 0. ]],
[[ 1.41421356, 0.78539816],
[ 1.41421356, 0.78539816],
[ 1.41421356, 0.78539816],
...,
[ 1. , 1. ],
[ 0. , 0. ],
[ 1. , 1. ]],
[[ 5. , 1. ],
[ -4.12310563, 1.32581766],
[ -4.12310563, 1.32581766],
...,
[ 1. , 1. ],
[ 0. , 0. ],
[ 1. , 1. ]],
...,
[[ 3.16227766, 1.24904577],
[ 2. , 1. ],
[ 2. , 1. ],
...,
[ 0. , 0. ],
[ 0. , 0. ],
[ 0. , 0. ]],
[[ -3.16227766, 1.24904577],
[ 2. , 1. ],
[ -2.23606798, 1.10714872],
...,
[ 0. , 0. ],
[ 1. , 0. ],
[ 1. , 0. ]],
[[ 1. , 1. ],
[ 1. , 1. ],
[ 1.41421356, 0.78539816],
...,
[ 0. , 0. ],
[ 0. , 0. ],
[ 0. , 0. ]]],
[[[ 0. , 0. ],
[ 0. , 0. ],
[ 0. , 0. ],
...,
[ 0. , 0. ],
[ 0. , 0. ],
[ 0. , 0. ]],
[[ -2.23606798, 0.46364761],
[ -1.41421356, 0.78539816],
[ -2.23606798, 0.46364761],
...,
[ 1. , 0. ],
[ 1. , 0. ],
[ 1. , 1. ]],
[[ -2.23606798, -0.46364761],
[ -1.41421356, -0.78539816],
[ 2. , 1. ],
...,
[ 0. , 0. ],
[ 1. , 0. ],
[ 1. , 0. ]],
...,
[[ 1. , 0. ],
[ 1. , 1. ],
[ -2.23606798, -1.10714872],
...,
[ 19.02629759, 1.51821327],
[ 19. , 1. ],
[-19.10497317, -1.46591939]],
[[ 3.60555128, 0.98279372],
[ 3.60555128, 0.5880026 ],
[ 5. , 0.64350111],
...,
[ 7.28010989, -1.29249667],
[ 7.61577311, -1.16590454],
[ 8.06225775, -1.05165021]],
[[ -7.28010989, 1.29249667],
[ -5. , 0.92729522],
[ -5.83095189, 0.5404195 ],
...,
[ 20.09975124, 1.47112767],
[ 21.02379604, 1.52321322],
[-20.22374842, -1.42190638]]],
[[[ 0. , 0. ],
[ 0. , 0. ],
[ 0. , 0. ],
...,
[ 0. , 0. ],
[ 0. , 0. ],
[ 0. , 0. ]],
[[ -1.41421356, 0.78539816],
[ -2.23606798, 1.10714872],
[ 2. , 1. ],
...,
[ 1. , 1. ],
[ 1. , 0. ],
[ 2.23606798, -0.46364761]],
[[ 1. , 0. ],
[ 1.41421356, 0.78539816],
[ 1. , 1. ],
...,
[ 0. , 0. ],
[ 1. , 1. ],
[ 0. , 0. ]],
...,
[[ -1.41421356, -0.78539816],
[ 0. , 0. ],
[ 1. , 1. ],
...,
[ 1. , 0. ],
[ 1. , 0. ],
[ 1. , 0. ]],
[[ 1. , 1. ],
[ -1. , -0. ],
[ 1. , 1. ],
...,
[ -1. , -0. ],
[ 0. , 0. ],
[ -1. , -0. ]],
[[ 0. , 0. ],
[ 1.41421356, -0.78539816],
[ -1. , -0. ],
...,
[ 1. , 0. ],
[ 0. , 0. ],
[ 1. , 0. ]]]])
我的密集光流分类器代码：
features_train, features_test, labels_train, labels_test = cross_validation.train_test_split(main, target, test_size = 0.4)
# Determine amount of time to train
t0 = time()
model = SVC(probability=True)
#model = SVC(kernel='poly')
#model = GaussianNB()
model.fit(features_train, labels_train)
print 'training time: ', round(time()-t0, 3), 's'
# Determine amount of time to predict
t1 = time()
pred = model.predict(features_test)
我的人脸地标跟踪分类器代码：
features_train, features_test, labels_train, labels_test = cross_validation.train_test_split(main.reshape(len(main), -1), target, test_size = 0.4)
# Determine amount of time to train
t0 = time()
#model = SVC()
model = SVC(kernel='linear')
#model = GaussianNB()
model.fit(features_train, labels_train)
# Determine amount of time to predict
t1 = time()
pred = model.predict(features_test)
在 sklearn（或一般的机器学习）中，我如何将这两个特征结合在一起，以创建一个统一且更好的分类器，在训练和预测时将这两个信息考虑在内？
【问题讨论】：
【参考方案1】：看看VotingClassifier。它允许您组合多个分类器并根据每个分类器的单独预测选择最终预测。
以上假设您可以使用 sklearn 的 dev (0.17) 版本。如果没有，您可以将 VotingClassifier 源复制到您的代码：https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/ensemble/voting_classifier.py。代码很简单。
【讨论】：
【参考方案2】：您可以按照 Daniel 的建议构建单独的分类器。但是，您可以考虑连接两个数据集：
main_dense_optical
main_face_landmark = main_face_landmark.reshape( len(main_face_landmark), -1 )
main = np.concatenate( [main_dense_optical, main_face_landmark], axis=1 )
# Code for train/test, training, evaluating here
【讨论】：
我考虑过，通常会这样做，但我担心由于某些功能彼此不同，以上是关于如何将两个特征/分类器组合成一个统一且更好的分类器？的主要内容，如果未能解决你的问题，请参考以下文章