Exercise 3 Triplet Loss - cannot progress & conflicting instructions

[Damieno]

Hi,
I am stuck. I have spent over a day on Exercise 3 - TripletLossFn. The instructions conflict, for example the detailed instruction say to use " triple_loss is the fastnp.mean of the sum of the two individual losses." while the code states " # add the two losses together and take the fastnp.sum of it ".

I do not understand this requirement with the masks. I find it confusing.

create a composition of two masks:

# the first mask to extract the diagonal elements, 
# the second mask to extract elements in the negative_zero_on_duplicate matrix that 
# are larger than the elements in the diagonal 

For the line " # the first mask to extract the diagonal elements, " → extract from what and in what shape ?

Again, for the second mask, what shape is it supposed to be. If I extract I reduce the shape. To extract elements that "are larger than the elements in the diagonal " - does this mean the the diagonal element in a row or the max of all diagonal elements or the diagonal i have already extracted ? How do you compare to elements ie the plural ?

The instructions are confusing and I am stuck. I have most of exercise 3 correct ( I hope ). I need some help with finalising this exercise please… so I can move on to the next exercise.

Thanks

[arvyzukai]

Hi @Damieno

Check this post and compare your values. It might help you find the solution.

Good luck!

[Damieno]

Hi, you have not clarified the instructions. Can you please explain and expand on

create a composition of two masks:

the first mask to extract the diagonal elements,

the second mask to extract elements in the negative_zero_on_duplicate matrix that

are larger than the elements in the diagonal

Also, is the final calculation fastnp.sum or .mean ? The instructions contradict each other.

Thanks

[arvyzukai]

Ok, to clarify the instructions but without revealing the whole solution :):

@Damieno

# the first mask to extract the diagonal elements,

@Damieno

# the second mask to extract elements in the negative_zero_on_duplicate matrix that 
# are larger than the elements in the diagonal 

@Damieno

create a composition of two masks:

This (mask) is a way to help you exclude the diagonal elements and elements that are larger than the diagonal elements. This is the heart of Triplet Loss function.

@Damieno

Also, is the final calculation fastnp.sum or .mean ? The instructions contradict each other.

It’s the former - fastnp.sum. The instructions will be corrected.

Cheers

[Damieno]

Hi, thank you for the extra info and explanation. I have worked through the exercise variable by variable and successfully completed it.