I understand how custom loss functions work in tensorflow. Suppose in the following code , a and b are numbers.
def customLoss( a,b):
def loss(y_true,y_pred):
loss=tf.math.reduce_mean(a*y_pred + b*y_pred)
return loss
return loss
But what if a and b are arrays which have the same shape as y_pred. let's say
y_pred= np.array([0,1,0,1])
a= np.arange(4)
b= np.ones(4)
I would like to the value of loss function is equal to 6:
np.mean(a*y_pred + b*y_pred) #element-wise.
I feel like my loss function is wrong now. It should be two additional inputs or weights of each sample. Can anyone help please? Thank you.
Assuming that a and b are fixed numbers across all loss computations, you can do something similar to your original loss function:
import numpy as np
import tensorflow as tf
y_pred = np.array([0, 1, 0, 1])
y_true = np.array([0, 1, 0, 1])
a = np.arange(4)
b = np.ones(4)
def get_custom_loss(a, b):
a = tf.constant(a, dtype=tf.float32)
b = tf.constant(b, dtype=tf.float32)
def loss_fn(y_true, y_pred):
y_true = tf.cast(y_true, dtype=tf.float32)
y_pred = tf.cast(y_pred, dtype=tf.float32)
loss = tf.math.reduce_sum(
[tf.math.multiply(a, y_pred), tf.math.multiply(b, y_pred)])
return loss
return loss_fn
loss_fn = get_custom_loss(a, b)
print(loss_fn(y_pred, y_true))
tf.math.multiply(foo, bar) will perform element-wise multiplication of two tensors (see docs). I assume you also want to sum the result of the two products, rather than take the mean (which would be 1.5).
Just a note: you aren't currently using y_true in your loss.