Kernel dying consistently in DLS 4/week 3 U-Net assignment

Course Q&A Deep Learning Specialization Convolutional Neural Networks
21

Don’t ask me why but it works fine when you do not execute the “display” function (this
function is used in the 2 cells before the training).

Best,
Alexandre

22

Commenting out the display function I got the training to succeed once, but the notebook kernel dies further down. I still can’t get the submission to succeed - I get the same error

Cell #8. Can’t compile the student’s code. Error: AssertionError(‘Error in test’)

23

The course staff reports that Coursera engineering is working on the “dead kernel” issue.

24

That error may also indicate that your code is incorrect: it passes the tests in the notebook, but fails (at least) one of the tests that the grader uses. Perhaps your code is not general: e.g it hard-codes some assumptions about dimensions or references global variables from the body of some of your functions, rather than the actual parameters passed.

25

Can confirm this is still an issue. I got my assignment graded at 100%. But I still really wanted to see the network in action.

26

@ paulinpaloalto yes you’re right: I had an error in my cell 7; specifically: I made the mistake of setting the kernel_initializer to the TF class

kernel_initializer=tf.keras.initializers.HeNormal

when I should have just used the string ‘he_normal’ per the instructions.

Fixing this, I can submit and get a 100% grade from the auto-grader. But I still get the kernel dying message.

1 Like
27

Thank you, @vorpalsnark ! After the same fix grader gave 100% to my submission. Kernel issue still persists though.
Is there an actual difference between ‘he_normal’ and tf.keras.initializers.HeNormal? I thought the first is just an alias for the latter.

28

Whether they are functionally identical is one issue. But the issue that breaks the grader is that in some cases, it checks for a specific implementation by looking at the exact code you entered.

29

Thanks for the explanation. It seems like the tests here (the ones which are part of the exercise, not the autograder) are a little weak or inconsistent?

30

Agreed, the unit tests and the grader look for different things, and they use different checking methods.

31

Don’t ask me why but it works fine when you do not execute the “display” function (this
function is used in the 2 cells before the training).

That didn’t work for me, tried it a few times.

32

OK. Yet, I tested several times (~5 times) w/ and w/o the “display” function. And the training was only successful w/o the display function (even for 40 epochs).
I have just checked twice more, and obtained the same outcome: training ok only w/o prior executions of the “display” function. So maybe it can be helpful for some folks.

33

That’s what Coursera is having difficulty with sorting out. I’ve never seen the “kernel dying” issue personally. So it’s not a consistent issue that’s easy to investiate.

The kernel_initializer and the dying kernel appear to be different issues.

The kernel_initializer is a known discrepancy between the tests the notebook and the grader use. I’ve raised an issue for that with the course staff.

34

Hi all,

The Kernel dying issue has been fixed. “Coursera has increased the memory required by the GPU assignments.”

“In order for the change to take effect for the leaners they would have to reboot their server.”

‘They can do this by clicking on the “Help” on the top right and when the panel opens, click on “Reboot”.’

If the issue still persists, let us know.

Best,
Saif.

35

Fantastic - thanks for letting us know.

36

Worked for me after I restarted the server. That’s a critical step.

37

Thanks Saif, I have the same issue. “Kernel dying” message when executing training cell. Epoch #1 never started. Tried several times.

I then read these posts, clicked on “help” and clicked on “reboot”. Server rebooted. After that, epoch#1 was able to start during training, but kernel died again shortly after that. Any suggestions ? Problem came back ?

38

Oh! That’s strange, even after rebooting, you are getting kernel dying error. There might be the following reasons:

  1. Are you sure you didn’t add any print statements in the inner loops that are causing the output to be more voluminous than necessary?
  2. Try doing “Kernel → Restart and Clear Output, and then run all.
  3. It could be random “bad weather in the cloud”. It means that server is busy. You may try later when server isn’t busy. Try tomorrow.

Kindly let us know if the above don’t solve your problem.

One more thing you can do is, just finish your code, without running the cells and if you are sure that your codes are correct, you can submit your assignment. Grader don’t need the output.

Best,
Saif.

39

Hello @Mariano_Domenech!

Try running it at a different time. This user tried for a different time and succeed. Or perhaps try rebooting by clicking on the “Help” on the top right and when the panel opens, click on “Reboot”.

Meanwhile, you can submit your assignment as grader doesn’t need the output of your code.

Best,
Saif.

40

I’m also having this kernel dying issue. It occurs every time I reach section 4 and attempt to train my model. It tries to run the first 1/5 Epoch and then suddenly dies. I was able to turn in my assigned and did receive a score of 100%. I am just unable to run any Jupyter cells after “4 - Train the Model”.