Implementing a custom loss function for object detection

Question

I am new to keras and tensorflow . How do I go about implementing a custom loss function while doing object detection , right now I have 5 parameters - 4 for bounding box coordinates and 1 for whether the object is present or not . Loss function should return square of difference between coordinates if object is present else if object is absent it should return a huge value as loss . This is the code I am tring right now:

def loss_func(y_true,y_pred):
  mask = np.array([False, False, False,False,True])   # check column of the class of object
  mask1 = np.array([True, True, True,True,False])     # get the columns of the coordinates of B box
  check_class = K.mean(K.square(tf.subtract(tf.boolean_mask(y_true,mask),tf.boolean_mask(y_pred,mask))))
  mean_square = K.mean(K.square(tf.subtract(tf.boolean_mask(y_true,mask1),tf.boolean_mask(y_pred,mask1))))
  value=K.mean(tf.boolean_mask(y_pred,mask))

  return value*mean_square + check_class

Here I am masking other values to obtain the last value which is 1000--> object present 0 --> object absent. Is there any other better way to do this?

The value of loss when I am running this in Kaggle decreases rapidly , by 2nd epoch the loss becomes 0.

Answer 1

First of all i would recommend using 1 and not 1000 for the "image exist" parameter.
You can manipulate y_true and y_pred.

penalty = 100

def lf(y_true,y_pred):
    mean_square = tf.keras.losses.mean_squared_error(y_true[:,0:4], y_pred[:,0:4])
    check_class = tf.subtract(y_true[:,4], y_pred[:,4])
    check_class = check_class * -penalty
    check_class = tf.keras.backend.mean(check_class)
    return mean_square + check_class

The above function first check the mean squared error for the first 4 parameters.
the second part then check for the "present" parameter.
If they are different it will output -1, if they are the same it will output 0.
Then it use the penalty for punishing it for a wrong parameter.

Using "punish" by some constant can be difficult to train. I would recommend changing the optimizer to SGD, adam will not work well in the situation, and playing with the penalty until you reach satisfying results.