Loss values seem absurd

The loss values are close to 16 or close to zero.
The resulting gradient tensor is filled with only zeros.
The image is fully black.
This is before training, but looks nothing like the image provided in the notebook.

then go back and check your model, read instructions, refer the ungraded lab. What changes you can make in your model to see a better loss value.

1 Like

I’ve been checking for many hours. I can’t find any ways I changed it. The inputs look appropriate. The predictions are taken directly from the model the course provides- and they don’t look that great.

How can I figure out how to proceed?

Have you passed all the grader cell?

Huh? Do you mean the last cell?

No I mean every grader cell where you have written or suppose to write codes.

1 Like

What do you mean by “passed”, then?

I mean the cell have run successfully without creating any error log.

It is Cats vs Dog assignment right?

My cells run without creating an error. But they don’t produce appropriate output. I have not moved on past where I am having trouble.

Can you share a screenshot where you are having trouble. Please make sure the screenshot doesn’t have any part of the codes. It is against community guidelines.

Without seeing the code, my images look like black squares.

This image doesn’t give much of the description of your outcome to understand one what was your loss value other than two black images

Hello Dan,

There are few ungraded labs did you try referring to those labs before starting this assignment?

If yes and still not able to get the solution, then send your saved copy of your assignment via personal DM.


I figured it out.

So you cleared the grader. Let everyone know how you resolve your issue!!!

I had created my model with the functional API. While the output looked the same and had the same number of trainable variables, it didn’t work with the rest of the code. I changed it to the sequential model.

im strruggling with this i had sequential from get go too lol - can i ask to pass did u have to get into cv2 adjustments or tweak superimposed optional section or was it straight code and you pulled the 95 or better ssim