构建自定义指标与Keras损失函数，有错误

Question

我想编写自定义指标函数编译步骤设置这样写道：

self.model.compile(optimizer=sgd,loss='categorical_crossentropy',metrics=[self.dice_similarity_coefficient_metric,self.positive_predictive_value_metric,self.sensitivity_metric]) 

我写骰子相似系数，阳性预测值和相似性是这样的：

• FP =假阳性
• TP =真阳性
• FN =假阴性

---

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()))) 

当我运行的代码，我有以下错误：

InvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'dense_3_target' with dtype float [[Node: dense_3_target = Placeholderdtype=DT_FLOAT, shape=[], _device="/job:localhost/replica:0/task:0/cpu:0"]]

能有人照顾解释问题出在哪里？ 我在哪里错了？

谢谢

Answer 1

也许，这是更好地定义使用后端功能指标。例如：

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

：

def false_negatives(Y_true, Y_pred): 
    return K.sum(K.round(K.clip(Y_true - Y_pred, 0, 1))) 

它可以在用5 FN的示例性数据来检查