Hello
i used the tf.keras.losses.categorical_crossentropy and also tf.reduce.sum
also logits and labels are in shape of (6, num_examples) , so i used tf.transpose on then so that i have them in expected shape for categorical_crossentropy function
At the end i devided the the result into unmber of example
still i get a big different
please help
thanks
In this function we are computing the sum, not the average of the costs. The other thing to check is to make sure you used the from_logits parameter correctly. We are passing the “logits” here and not the softmax output values, right? Here’s a thread which talks about that and why it is done that way.
Here’s a thread which talks about why it’s the sum, not the average.
2 Likes
Thank you, after using from_logits parameter and removing the deviding part of code, it went smooth.
The reason of my confusion was this part Text in Exercise:
The second link I gave you in my previous reply explains exactly that point. Please have another look at it.