code in BCI
BOSCH Python Repositories: https://anu9rng:AP6eY5xuhS1MqAdy5jedftw3ndQq7MHjXL8Rpb@rb-artifactory.bosch.com/artifactory/api/pypi/python-virtual/simple/
BOSCH Angular NPM Repositories:https://rb-artifactory.bosch.com/artifactory/webapp/#/artifacts/browse/tree/General/bci-npm-virtual
Naive presents factors are independence for each other
where Y represents labels, X represents factor.
where X1, X2, ... , Xn represent varies of factors values.
the probability of prediction:
or
we can also set the max(p(Y1|X1,X2,.......,Xn), p(Y2X1,X2,.......,Xn)) value as the probability.
Hyperparams: num_neurons:the number of neuons Dropout:A Simple Way to Prevent Neural Networks from Overfitting. Activations: Relu was choosen as activaions. Softmoid: It can transIform continuous-valued to between 0 and 1,but the gradient may be missed with iteration increase. Hence, it is seldom used. Hyperbolic Tangent (tanh): Its put value is zero-centered, but the problem of gradient missing is still exist. Relu: It is a Maximum function and the most popular activation. Elu: It is similar with the Relu, but speend more computational cost. loss: mean_squared_error：
mean_absolute_error:
mean_absolute_percentage_error
binary_crossentropy: binary cross entropy loss function
categorical_crossentropy
User-defined: binary_crossentropy (unbalanced data)
the loss function increase weight of less sample to the gradient of loss function incline less sample under the situation of unbalanced samples between positive and negative.
def loss(y_true,y_pred):
return K.mean(((y_true-1)*K.log(1-y_pred+K.epsilon())-y_true*500*K.log(y_pred+K.epsilon())),axis=-1)
optimizer: adam
metric:
Positive Negative
True TP FN Flase FP TN
Positive_Score=TP/（FP+TN）Negtive_Score=TN/(TN+FP)
Metric_Score=2Negtive_ScorePositive_Score/（Negtive_Score+Positive_Score）