I am lost here. If the angle calculation is 𝜃(𝑝𝑜𝑠,𝑖,𝑑)=𝑝𝑜𝑠/(10000*2𝑖/𝑑) and the first couple of i’s are zero, then you have division by zero. If i is a vector, then division of pos/i is not really defined, right?

You’re missing part of the equation, ‘i’ is used as an exponent, not multiplication. [image]

Yup, there’s a big difference between a * b and a ** b. :nerd_face: You can do “elementwise” division with a vector in the denominator, the same way there is elementwise multiplication.

:joy: That is a big revelation! You have to admit, it kind of looks like multiiplication rather than an exponent. Now it works. Thank you!

Here’s a “play” cell I added to my version of that notebook to show how elementwise divide works and “plays well” with broadcasting: A = np.arange(4)[:, np.newaxis] B = 2 * np.ones((1,6)) print(f"A.shape {A.shape}") print(f"A {A}") print(f"B.shape {B.shape}") print(f"B {B}") quotient = A / B print(…

Cool! Here is another question on the next exercise. I don’t get how to do the sin/cos on both sides of the equation for angle_rads[:, 0::2] = ??? I tried to use the same [:, 0::2] but apply sin and I get the following error: could not broadcast input array from shape (8,16) into shape (8,8) I…

Yes, you are taking a tensor and applying sin to it “elementwise” to get another tensor of the same shape. You must not have actually used the same tensor indices on both sides.

Notice that your angles_rad tensor is 8 x 16, so somehow you must have used the whole thing on the RHS instead of slicing it. Maybe this is a case in which you typed new code in the cell and then just called it again without actually clicking “Shift-Enter” on the modified cell. So WYRINWYS (What Yo…

Yes, that was it. I think I am going to just run all above every so often just to keep things current. BTW creating the pos column vector and k row vector was a new challenge for me. Not much guidance on that but I guess they are trying to wean us from the hand holding. I did some research and e…

An alternative method is using np.arange() and np.newaxis(). This is shown in the test cell just prior to Section 1.2.

Exercise 1 angle calculation

Course Q&A Deep Learning Specialization Sequence Models

paulinpaloalto January 4, 2026, 7:17pm 9

Here’s a post that talks in a bit more detail about this phenomenon.

Topic		Replies	Views
[week 4] Transformer Network - get_angles Sequence Models week-module-4 , coursera-platform	56	4682	January 10, 2026
[Week 4] Exercise 1 - get_angles Sequence Models coursera-platform	23	1993	February 16, 2026
Course 5 week 4 Sequence Models coursera-platform	1	784	June 27, 2021
Not sure the first unit test is right in week 4 lab Sequence Models week-module-4 , coursera-platform	8	48	May 25, 2025
Week 4 Scaled Dot Product Attention Sequence Models coursera-platform	10	825	October 31, 2021

Exercise 1 angle calculation

Related topics