Styjun

I am working on a deep learning project in Keras, and have implemented a sensitivity function using TensorFlow backend, since that is needed if I want to evaluate a model using it.
However, I cannot extract the value from the tensor. I want to return it, so that I can use the values in other functions. Ideally, the return value should be an int. Whenever I evaluate the function, I just get the tensor object itself, not its real value.

I have tried creating a session and evaluating, but to no avail. I am able to print the value just fine in this way, but I cannot assign the value to another variable.

def calculate_tp(y, y_pred):
 TP = 0
 FP = 0
 TN = 0
 FN = 0
 for i in range(5):
 true = K.equal(y, i)
 preds = K.equal(y_pred, i)
 TP += K.sum(K.cast(tf.boolean_mask(preds, tf.math.equal(true, True)), 'int32'))
 FP += K.sum(K.cast(tf.boolean_mask(true, tf.math.equal(~preds, True)), 'int32'))
 TN += K.sum(K.cast(tf.boolean_mask(~preds, tf.math.equal(true, True)), 'int32'))
 FN += K.sum(K.cast(tf.boolean_mask(true, tf.math.equal(preds, False)), 'int32'))

 """with tf.Session() as sess:
 TP = TP.eval()
 FP = FP.eval()
 FN = FN.eval()
 FP = FP.eval()
 print(TP, FP, TN, FN)
 #sess.run(FP)"""
 return TP / (TP + FN)

If I understand well your problem, you can simply create a new tensor with the values
obtened.

For example :

tensor = tf.constant([5, 5, 8, 6, 10, 1, 2])
tensor_value = tensor.eval(session=tf.Session())
print(tensor_value) #get [ 5 5 8 6 10 1 2]
new_tensor = tf.constant(tensor_value)
print(new_tensor) #get Tensor("Const_25:0", shape=(7,), dtype=int32)

Hope I helped !

Ok so could it be because in your tries TP is always 0 ?

If i try that :

y = np.array([0, 0, 0, 0, 0, 1, 1, 1 ,1 ,1])
y_pred = np.array([0.01, 0.005, 0.5, 0.09, 0.56, 0.999, 0.89, 0.987 ,0.899 ,1])

def calculate_tp(y, y_pred):
 TP = 0
 FP = 0
 TN = 0
 FN = 0
 for i in range(5):
 true = K.equal(y, i)
 preds = K.equal(y_pred, i)
 TP += K.sum(K.cast(tf.boolean_mask(preds, tf.math.equal(true, True)), 'int32'))
 FP += K.sum(K.cast(tf.boolean_mask(true, tf.math.equal(~preds, True)), 'int32'))
 TN += K.sum(K.cast(tf.boolean_mask(~preds, tf.math.equal(true, True)), 'int32'))
 FN += K.sum(K.cast(tf.boolean_mask(true, tf.math.equal(preds, False)), 'int32'))

 TP = TP.eval(session=tf.Session())
 FP = FP.eval(session=tf.Session())
 TN = TN.eval(session=tf.Session())
 FN = FN.eval(session=tf.Session())
 print(TP, FP, TN, FN)

 results = TP / (TP + FN)

 return results

res = calculate_tp(y, y_pred)
print(res) 

#Outputs : 
#0 5 5 5
#1 9 9 9
#1 9 9 9
#1 9 9 9
#1 9 9 9
#0.1

It give me a float number, like you want.

Is it helping ?