我正在尝试编写自定义度量函数,以便以这种方式编写的编译步骤中设置:
self.model.compile(optimizer=sgd,loss='categorical_crossentropy',metrics=[self.dice_similarity_coefficient_metric,self.positive_predictive_value_metric,self.sensitivity_metric])
我用这种方式写了骰子相似系数,正预测值和相似度:
def dice_similarity_coefficient_metric(self, y_true, y_pred):
y_true = np.array(K.eval(y_true))
y_pred = np.array(K.eval(y_pred))
FP = np.sum(y_pred & np.logical_not(y_true)).astype(float)
TP = np.sum(y_true & y_pred).astype(float)
FN = np.sum(np.logical_not(y_pred) &
np.logical_not(y_true)).astype(float)
return K.variable(np.array((2 * TP) / (FP + (2 * TP) + FN +
K.epsilon())))
def positive_predictive_value_metric(self, y_true, y_pred):
y_true = np.array(K.eval(y_true))
y_pred = np.array(K.eval(y_pred))
FP = np.sum(y_pred & np.logical_not(y_true)).astype(float)
TP = np.sum(y_true & y_pred).astype(float)
return K.variable(np.array(TP / (FP + TP + K.epsilon())))
def sensitivity_metric(self, y_true, y_pred):
y_true = np.array(K.eval(y_true))
y_pred = np.array(K.eval(y_pred))
TP = np.sum(y_true & y_pred).astype(float)
FN = np.sum(np.logical_not(y_pred) &
np.logical_not(y_true)).astype(float)
return K.variable(np.array(TP / (TP + FN + K.epsilon())))
当我运行代码时,我有以下错误:
InvalidArgumentError(请参阅上面的回溯):您必须使用dtype float为占位符张量'dense_3_target'提供值 [[节点:dense_3_target = Placeholderdtype = DT_FLOAT,shape = [],_ device =“/ job:localhost / replica:0 / task:0 / cpu:0”]]
有人可以解释问题的位置吗? 哪里我错了?
谢谢
答案 0 :(得分:0)
使用后端功能定义指标可能更好。例如:
def false_negatives(Y_true, Y_pred):
return K.sum(K.round(K.clip(Y_true - Y_pred, 0, 1)))
可以使用5 FN检查示例数据:
y_true = np.array([[1.0, 1.0, 0.0, 1.0], [1.0, 1.0, 0.0, 1.0], [1.0, 1.0, 0.0, 1.0]], dtype=np.float32)
y_pred = np.array([[0.3, 0.99, 0.99, 0.1], [0.6, 0.99, 0.99, 0.1], [0.1, 0.99, 0.99, 0.1]], dtype=np.float32)
n_fn = np.sum((y_true - y_pred) > 0.5)
Y_true = K.placeholder((None, 4), dtype=K.floatx())
Y_pred = K.placeholder((None, 4), dtype=K.floatx())
n_fn = false_negatives(Y_true, Y_pred).eval(inputs_to_values={Y_true: y_true, Y_pred: y_pred})
HTH