我正在尝试在keras中实现一个简单的神经网络。代码是:
model = Sequential()
model.add(Dense(512, input_dim = 55 , kernel_regularizer=l2(0.00001),
activation = 'relu'))
model.add(Dense(8, activation = 'softmax'))
model.compile(loss = 'categorical_crossentropy' , optimizer = 'adam' , metrics = ['accuracy'] )
model.fit(X_train, dummy_y, epochs = 20, batch_size = 30, class_weight=class_weights)
我有55个功能,我想预测8个类之一(0,1,2,3,4,5,6,7)。我也喜欢y_train
:
encoder = LabelEncoder()
encoder.fit(y_train)
encoded_Y = encoder.transform(y_train)
# convert integers to dummy variables (i.e. one hot encoded)
dummy_y = np_utils.to_categorical(encoded_Y)
但是,当我使用predict()
时,输出是每个类的概率的数组:
array([[3.3881092e-01, 2.6201099e-06, 1.9504215e-03, ..., 7.0641324e-02,
4.4026113e-01, 1.2641836e-02],
[2.3457911e-02, 5.5409328e-04, 2.8759112e-05, ..., 2.1585675e-03,
5.5625242e-01, 1.0208529e-01],
[4.6981460e-01, 2.0882198e-05, 1.4895502e-01, ..., 1.3179567e-01,
2.2908358e-01, 1.4160757e-03],
...
我应该如何修改网络以输出最高概率的类?这样:
[[0,5,7,3,2,0,0,.....]]
您可以简单地使用predict_classes
方法:
preds_classes = model.predict_classes(X_test)
您看到的那些数字是predict
方法的输出是每个类的概率或置信度得分。因此,作为替代解决方案,您可以获取与预测类相对应的最大分数的索引:
import numpy as np
probs = model.predict(X_test)
classes = np.argmax(probs, axis=-1)