Build custom Metric for Loss Function with Keras, with errors - customization

Question

I'm trying to write custom metric function to set in the compile step wrote in this way:
self.model.compile(optimizer=sgd,loss='categorical_crossentropy',metrics=[self.dice_similarity_coefficient_metric,self.positive_predictive_value_metric,self.sensitivity_metric])
I wrote Dice Similarity Coefficient, Positive Predictive Value and Similarity in this way:
FP = false positive
TP = true positive
FN = false negative
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())))
when i run the code i have the following error:
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"]]
Can someone care to explain where is the problem?
Where i'm wrong?
Thank you

Answer 1

Probably, it's better to define metrics using backend functions. For example:
def false_negatives(Y_true, Y_pred):
return K.sum(K.round(K.clip(Y_true - Y_pred, 0, 1)))
It can be checked on an example data with 5 FN:
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