C4 Week 4 Exercise 6 - train_step

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

I have some problems with exercise 6, the train_step function.

I get no error message, but every time I run the function I get new tensor values that do not correspond to the solution. For instance:

tf.Tensor(25700.293, shape=(), dtype=float32)
tf.Tensor(17777.773, shape=(), dtype=float32)

I expect the error is in one of the previous functions. But all pass the test and pass the grader function.

I do not really know where to look. Thank you for your help.

2

@lindhard

Share the screenshot of the error you are encountering without sharing any part of the grade cell codes.

Regards
DP

3

What does the grader say if you submit your code?

One thing to note is that the train_step function is stateful: you canā€™t just run the test cell multiple times. If you want consistent results, you have to start by running the train_step cell itself.

Your values are actually relatively close. It could be different rounding behavior caused by numerical stability issues. It might be worth having a look at this thread.

4

I have no error. I states it passes, even though the tensor values are wrong. When I submit it for grading it does not pass.

Everything runs :slight_smile:

5

Thanks for your input. It makes sense, just have to figure out where the problem is, as everything runs :slight_smile:

6

The grader says that it is wrong.
I can run everything and even generate fine pictures.
It is a little hard to know where the error is as i get no error message.

7

It is probably time to look at the code, but we canā€™t do that on a public thread. I will send you a DM ā€¦

8

you need to share screenshot of your submission grade where it provides information about the reason why it failed, click on show grader output and then take screenshot.

9

Code Cell UNQ_C1: Function ā€˜compute_content_costā€™ is correct.
Code Cell UNQ_C2: Function ā€˜gram_matrixā€™ is correct.
Code Cell UNQ_C3: Function ā€˜compute_layer_style_costā€™ is correct.
Code Cell UNQ_C4: Function ā€˜total_costā€™ is correct.
Code Cell UNQ_C5: Function ā€˜train_stepā€™ is incorrect. Check implementation.
If you see many functions being marked as incorrect, try to trace back your steps & identify if there is an incorrect function that is being used in other steps.
This dependency may be the cause of the errors.

10

In UNQ_CI i get a roundet of number:

Not sure if that is the problem.
J_content = tf.Tensor(7.056877, shape=(), dtype=float32) All tests passed

Expected Output:

J_content 7.0568767

11

Regarding UNQ_C5 i get:

tf.Tensor(25700.293, shape=(), dtype=float32) tf.Tensor(17777.783, shape=(), dtype=float32) All tests passed

Expected output

tf.Tensor(25700.346, shape=(), dtype=float32)
tf.Tensor(17778.389, shape=(), dtype=float32)

But they vary a little

12

To close the loop on the public thread, there was a bug in train_step. The logic used a_S as the input image for both the style cost and the content cost. Probably a classic ā€œcopy/pasteā€ error.

The scarier point is that the tests in the notebook do not catch this error, although the grader does.

1 Like
13

Improving the unit tests is always in-fashion.

1 Like