Conditional Replacement in Keras Backend - tensorflow

Question

Target:
Hey i want to build a custom metrics method, where different specific classes are allowed do be confused with specific (because their idea in some conditions in nearly the same).
How it could be reached:
I got 25 classes, where the following confusions should be allowed
(1 can be 2,3,4) ; (2 can be 3,4) ; (3 can be 4) ;(15 can be 16,17,18);(16 can be 17) ; (17 can be 18) ; (6 can be 7,8 ) ; (7 can be 8 ); (19 can be 20,21 ) ; (20 can be 21 );(10 can be 11 );(13 can be 14 );(22 can be 23 ); (24 can be 14 );
quite much huh... therefore i put it into a mask:
mask = mask =(y_true!=y_pred)&(((y_true==1)&((y_pred == 2)| (y_pred == 3)| (y_pred == 4))) | ((y_true==2) & ((y_pred == 3)|(y_pred == 4))) | ((y_true==3)&((y_pred == 4)))| ((y_true==15)&((y_pred == 16) | (y_pred == 17)| (y_pred == 18))) | ((y_true==16) & ((y_pred == 17)|(y_pred == 18))) | ((y_true==17)&((y_pred == 18))) | ((y_true==6) & ((y_pred == 7)|(y_pred == 8)))|((y_true==7)&((y_pred == 8)))| ((y_true==19) & ((y_pred == 20)|(y_pred == 21)))|((y_true==20)&((y_pred == 21)))|((y_true==10)&((y_pred == 11)))|((y_true==13)&((y_pred == 14)))|((y_true==22)&((y_pred == 23)))|((y_true==24)&((y_pred == 14))) )
I was hoping to find a function (equivalent to np.where) to conditional replace the values in the "y_pred" which fit the condition, by the "y_true" tensors values at the same index.
I thought tf.where_v2() could work (sadly, i wasnt able to check if it works because tf.equal() and tf.reduce_all() only have limited abilities to work in a metrics function (or am I wrong?)
the following creates no debug errors
def profit_metrics(y_true, y_pred):
mask =(y_true!=y_pred)&(((y_true==1)&((y_pred == 2)| (y_pred == 3)| (y_pred == 4))) | ((y_true==2) & ((y_pred == 3)|(y_pred == 4))) | ((y_true==3)&((y_pred == 4)))| ((y_true==15)&((y_pred == 16) | (y_pred == 17)| (y_pred == 18))) | ((y_true==16) & ((y_pred == 17)|(y_pred == 18))) | ((y_true==17)&((y_pred == 18))) | ((y_true==6) & ((y_pred == 7)|(y_pred == 8)))|((y_true==7)&((y_pred == 8)))| ((y_true==19) & ((y_pred == 20)|(y_pred == 21)))|((y_true==20)&((y_pred == 21)))|((y_true==10)&((y_pred == 11)))|((y_true==13)&((y_pred == 14)))|((y_true==22)&((y_pred == 23)))|((y_true==24)&((y_pred == 14))) )
new_pred = tf.where_v2(mask,y_true,y_pred)
return sparse_categorical_accuracy(y_true, new_pred)
(...)
model.compile(...,metrics=['accuracy',profit_metrics,])
but the thing is... it returns exactly the same accuracy value as the standard accuracy method!
I tried to check if both tensors are still the same, but sadly all i tried didnt work out, maybe because i need to use only Keras backend functions?
def profit_metrics(y_true, y_pred):
mask =(y_true!=y_pred)&(((y_true==1)&((y_pred == 2)| (y_pred == 3)| (y_pred == 4))) | ((y_true==2) & ((y_pred == 3)|(y_pred == 4))) | ((y_true==3)&((y_pred == 4)))| ((y_true==15)&((y_pred == 16) | (y_pred == 17)| (y_pred == 18))) | ((y_true==16) & ((y_pred == 17)|(y_pred == 18))) | ((y_true==17)&((y_pred == 18))) | ((y_true==6) & ((y_pred == 7)|(y_pred == 8)))|((y_true==7)&((y_pred == 8)))| ((y_true==19) & ((y_pred == 20)|(y_pred == 21)))|((y_true==20)&((y_pred == 21)))|((y_true==10)&((y_pred == 11)))|((y_true==13)&((y_pred == 14)))|((y_true==22)&((y_pred == 23)))|((y_true==24)&((y_pred == 14))) )
new_pred = tf.where_v2(mask,y_true,y_pred)
bool_tensor = tf.equal(new_pred,y_pred)
same = tf.math.reduce_all(bool_tensor)
session = tf.keras.backend.get_session(op_input_list=())
K.set_session(same)
if session.run(same)==True:
return sparse_categorical_accuracy(y_true, new_pred)#K.mean(K.equal(K.argmax(y_true, axis=-1), K.argmax(new_pred, axis=-1)))
else:
return tf.zeros([16, 25], tf.int32)
(both tensors have shape [16, 25])
something similar to where_v2 from keras.backend
or at least a method to show if my code works somehow(?) would be great
But in the end:
All i want is to make the conditional confusion for the mentioned conditions possible in my custom accuracy metrics and i would be extremely glad for anyones helpful idea (thank you) :'/