我已经建立了一个小程序，用scikit learn为给定的数据集创建一个分类器。现在我想试试这个例子，看看分类器在工作。例如，clf必须检测“猫”。
我是这样说的：
我有50张猫的照片和50张“没有猫”的照片。
>
data_set的描述符training_settraining_setantest_set的histogramm数据尝试以下代码：
tuned_parameters = [{'kernel': ['rbf'], 'gamma': [1e-3, 1e-4],
'C': [1, 10, 100, 1000]},
{'kernel': ['linear'], 'C': [1, 10, 100, 1000]}]
scores = ['precision', 'recall']
for score in scores:
print("# Tuning hyper-parameters for %s" % score)
print()
clf = GridSearchCV(SVC(C=1), tuned_parameters, cv=5, scoring=score)
clf.fit(X_train, y_train)
print("Best parameters set found on development set:")
print()
print(clf.best_estimator_)
print()
print("Grid scores on development set:")
print()
for params, mean_score, scores in clf.grid_scores_:
print("%0.3f (+/-%0.03f) for %r"
% (mean_score, scores.std() / 2, params))
print()
print("Detailed classification report:")
print()
print("The model is trained on the full development set.")
print("The scores are computed on the full evaluation set.")
print()
y_true, y_pred = y_test, clf.predict(X_test)
print y_true
print y_pred
print(classification_report(y_true, y_pred))
print()
print clf.score(X_train, y_train)
print "score"
print clf.best_params_
print "best_params"
pred = clf.predict(X_test)
print accuracy_score(y_test, pred)
print "accuracy_score"
我得到的结果是：
# Tuning hyper-parameters for recall
()
/usr/local/lib/python2.7/dist-packages/sklearn/metrics/metrics.py:1760: UserWarning: The sum of true positives and false positives are equal to zero for some labels. Precision is ill defined for those labels [ 0.]. The precision and recall are equal to zero for some labels. fbeta_score is ill defined for those labels [ 0.].
average=average)
/usr/local/lib/python2.7/dist-packages/sklearn/metrics/metrics.py:1760: UserWarning: The sum of true positives and false positives are equal to zero for some labels. Precision is ill defined for those labels [ 1.]. The precision and recall are equal to zero for some labels. fbeta_score is ill defined for those labels [ 1.].
average=average)
Best parameters set found on development set:
()
SVC(C=0.001, cache_size=200, class_weight=None, coef0=0.0, degree=3,
gamma=0.001, kernel=rbf, max_iter=-1, probability=False,
random_state=None, shrinking=True, tol=0.001, verbose=False)
()
Grid scores on development set:
()
0.800 (+/-0.200) for {'kernel': 'rbf', 'C': 0.001, 'gamma': 0.001}
0.800 (+/-0.200) for {'kernel': 'rbf', 'C': 0.001, 'gamma': 0.0001}
0.800 (+/-0.200) for {'kernel': 'rbf', 'C': 0.01, 'gamma': 0.001}
0.800 (+/-0.200) for {'kernel': 'rbf', 'C': 0.01, 'gamma': 0.0001}
0.800 (+/-0.200) for {'kernel': 'rbf', 'C': 0.10000000000000001, 'gamma': 0.001}
0.800 (+/-0.200) for {'kernel': 'rbf', 'C': 0.10000000000000001, 'gamma': 0.0001}
0.800 (+/-0.200) for {'kernel': 'rbf', 'C': 1.0, 'gamma': 0.001}
0.800 (+/-0.200) for {'kernel': 'rbf', 'C': 1.0, 'gamma': 0.0001}
0.800 (+/-0.200) for {'kernel': 'rbf', 'C': 10.0, 'gamma': 0.001}
0.800 (+/-0.200) for {'kernel': 'rbf', 'C': 10.0, 'gamma': 0.0001}
0.800 (+/-0.200) for {'kernel': 'rbf', 'C': 100.0, 'gamma': 0.001}
0.800 (+/-0.200) for {'kernel': 'rbf', 'C': 100.0, 'gamma': 0.0001}
0.800 (+/-0.200) for {'kernel': 'rbf', 'C': 1000.0, 'gamma': 0.001}
0.800 (+/-0.200) for {'kernel': 'rbf', 'C': 1000.0, 'gamma': 0.0001}
()
Detailed classification report:
()
The model is trained on the full development set.
The scores are computed on the full evaluation set.
()
[ 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1.
1. 1. 1. 1. 1. 1. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[ 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1.
1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1.
1. 1. 1. 1. 1. 1. 1. 1. 0. 1. 1. 1. 1. 1.]
precision recall f1-score support
0.0 1.00 0.04 0.08 25
1.0 0.51 1.00 0.68 25
avg / total 0.76 0.52 0.38 50
()
0.52
score
{'kernel': 'rbf', 'C': 0.001, 'gamma': 0.001}
best_params
0.52
accuracy_score
似乎CLF对所有人都说它是一只猫......但是为什么呢？
是否将数据设置为较小以获得良好结果？
编辑：我正在使用VLFeat来检测筛选描述符
功能：
def create_descriptor_data(data, ID):
descriptor_list = []
datas = numpy.genfromtxt(data,dtype='str')
for p in datas:
locs, desc = vlfeat_module.vlf_create_descriptors(p,str(ID)+'.key',ID) # create descriptors and save descs in file
if len(desc) > 500:
desc = desc[::round((len(desc))/400, 1)] # take between 400 - 800 descriptors
descriptor_list.append(desc)
ID += 1 # ID for filename
return descriptor_list
# create k-mean centers from all *.txt files in directory (data)
def create_center_data(data):
#data = numpy.vstack(data)
n_clusters = len(numpy.unique(data))
kmeans = KMeans(init='k-means++', n_clusters=n_clusters, n_init=1)
kmeans.fit(data)
return kmeans, n_clusters
def create_histogram_data(kmeans, descs, n_clusters):
histogram_list = []
# load from each file data
for desc in descs:
length = len(desc)
# create histogram from descriptors
histogram = kmeans.predict(desc)
histogram = numpy.bincount(histogram, minlength=n_clusters) #minlength = k in k-means
histogram = numpy.divide(histogram, length, dtype='float')
histogram_list.append(histogram)
histogram = numpy.vstack(histogram_list)
return histogram
电话：
X_desc_pos = lib.dataset_module.create_descriptor_data("./static/picture_set/dataset_pos.txt",0) # create desc from dataset_pos, 25 pics
X_desc_neg = lib.dataset_module.create_descriptor_data("./static/picture_set/dataset_neg.txt",51) # create desc from dataset_neg, 25 pics
X_train_pos, X_test_pos = train_test_split(X_desc_pos, test_size=0.5)
X_train_neg, X_test_neg = train_test_split(X_desc_neg, test_size=0.5)
x1 = numpy.vstack(X_train_pos)
x2 = numpy.vstack(X_train_neg)
kmeans, n_clusters = lib.dataset_module.create_center_data(numpy.vstack((x1,x2)))
X_train_pos = lib.dataset_module.create_histogram_data(kmeans, X_train_pos, n_clusters)
X_train_neg = lib.dataset_module.create_histogram_data(kmeans, X_train_neg, n_clusters)
X_train = numpy.vstack([X_train_pos, X_train_neg])
y_train = numpy.hstack([numpy.ones(len(X_train_pos)), numpy.zeros(len(X_train_neg))])
X_test_pos = lib.dataset_module.create_histogram_data(kmeans, X_test_pos, n_clusters)
X_test_neg = lib.dataset_module.create_histogram_data(kmeans, X_test_neg, n_clusters)
X_test = numpy.vstack([X_test_pos, X_test_neg])
y_test = numpy.hstack([numpy.ones(len(X_test_pos)), numpy.zeros(len(X_test_neg))])
tuned_parameters = [{'kernel': ['rbf'], 'gamma': [1e-3, 1e-4],
'C': [1, 10, 100, 1000]},
{'kernel': ['linear'], 'C': [1, 10, 100, 1000]}]
scores = ['precision', 'recall']
for score in scores:
print("# Tuning hyper-parameters for %s" % score)
print()
clf = GridSearchCV(SVC(C=1), tuned_parameters, cv=5, scoring=score)
clf.fit(X_train, y_train)
print("Best parameters set found on development set:")
print()
print(clf.best_estimator_)
print()
print("Grid scores on development set:")
print()
for params, mean_score, scores in clf.grid_scores_:
print("%0.3f (+/-%0.03f) for %r"
% (mean_score, scores.std() / 2, params))
print()
print("Detailed classification report:")
print()
print("The model is trained on the full development set.")
print("The scores are computed on the full evaluation set.")
print()
y_true, y_pred = y_test, clf.predict(X_test)
print y_true
print y_pred
print(classification_report(y_true, y_pred))
print()
print clf.score(X_train, y_train)
print "score"
print clf.best_params_
print "best_params"
pred = clf.predict(X_test)
print accuracy_score(y_test, pred)
print "accuracy_score"
编辑：通过更新范围和再次保存“精度”进行了一些更改
# Tuning hyper-parameters for accuracy
()
Best parameters set found on development set:
()
SVC(C=1000.0, cache_size=200, class_weight=None, coef0=0.0, degree=3,
gamma=1.0, kernel=rbf, max_iter=-1, probability=False, random_state=None,
shrinking=True, tol=0.001, verbose=False)
()
Grid scores on development set:
()
...
()
Detailed classification report:
()
The model is trained on the full development set.
The scores are computed on the full evaluation set.
()
[ 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1.
1. 1. 1. 1. 1. 1. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
[ 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 1. 0. 1. 0. 1. 1. 1.
1. 1. 1. 0. 1. 1. 1. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0.
0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0.]
precision recall f1-score support
0.0 0.88 0.92 0.90 25
1.0 0.92 0.88 0.90 25
avg / total 0.90 0.90 0.90 50
()
1.0
score
{'kernel': 'rbf', 'C': 1000.0, 'gamma': 1.0}
best_params
0.9
accuracy_score
但是通过在一张照片上用
rslt = clf.predict(test_histogram)
他仍然对沙发说：“你是一只猫”： D
似乎clf对所有人说它是一只猫。。。。但是为什么呢？
从粘贴的输出中很难分辨出来，但这似乎是分数=['precision'，'recall']上循环的第二次迭代，因此您正在优化召回。这与分类报告一致，该报告指出，对于阳性类别，召回率为1.00（完美）。
回忆什么时候是完美的？嗯，当没有假阴性时，就没有猫不被发现。因此，获得完美回忆的简单方法是预测每个输入图片的“cat”，而不管它是否是cat，并且GridSearchCV找到了一个能够准确做到这一点的分类器。
当您优化精度时，也可能会发生类似的情况：由于不会出现误报，所以永远不要预测“cat”可以实现完美的精度。
要避免这种情况，请针对准确性而不是精确性或召回率进行优化，或者针对F进行优化ᵦ 如果你有一个班级不平衡的情况。
这种行为有许多可能性：
C和gamma参数的范围可能太窄-此变量高度依赖于数据，您的表示法的值可能需要与当前使用的[欠拟合/过拟合]完全不同的C和gamma我个人的猜测（因为没有数据很难重现问题）这里是第三个选择-坏C和gamma参数，以找到一个好的模型。
编辑
你应该尝试更大范围的数值，例如。
>
C=[]
gamma=[]
for i in range(21): C.append(10.0**(i-5))
for i in range(17): gamma.append(10**(i-14))
编辑2
一旦参数的范围被纠正，现在你应该执行实际的“案例研究”。收集更多的图像，分析你的数据表示（直方图真的足够完成这项任务吗？），处理你的数据（它已经标准化了吗？也许尝试一些去相关？），考虑使用更简单的内核——rbf可能非常具有欺骗性——一方面，它可以在训练中获得高分，但另一方面，在测试中完全失败。这是其过拟合能力的结果（对于任何一致的数据集，RBF-SVM都可以在训练期间获得100%的分数），因此在模型的功率和泛化能力之间找到平衡是一个难题。这就是真正的“机器学习之旅”开始的时候，玩得开心！
