Python Keras Custom Loss Function and Gradient Tape

In this somewhat longer video I step you through the process that I go through when I am learning new features in Keras, or any new machine learning library. We walk through style transfer which uses a custom multi-objective loss function, and uses the optimizer to modify the actual pixels of the image with gradient descent.

Follow Me/Subscribe:

[prof_shixo]

Keep going in this direction as it is unique than the other content and it is actually very beneficial for research students in DL. I would say having session of "lets reproduce a paper" might be really illuminating.

[saranshbhandari]

This is awesome, I learned more from this video than my collage courses.

[dinoscheidt]

A lot of respect. To do this, left alone to bring it to video so the audience can learn something, is incredible tedious. Thank you from Berlin — will share this for sure.

[FailTrainS]

This was excellent. I'm working on getting up to speed for a new career in ML and am finding myself studying code pretty frequently at this point. Love the insights!

[shivagurrala]

I have defined multiple losses and performed back prop on pytorch, but never tried that on tensorflow. I have written few custom models on tensorflow but never tried “compute gradient and loss”. Thanks for the code walk through, it totally makes sense to run few iterations and check out what loss functions really do and colab is best place for it.

[zakariaabderrahmanesadelao]

Style Transfer never seemed easier. Thanks sir!

[garrettosborne]

Like your videos because you take a project view of ML. Thanks Jeff.

[joobcoh]

This style of video is great and actually really helpful, thank you for doing these

[andrewh]

Another great video. Much appreciated. I've learned more with your videos then any others by far!! I found it very useful and like all of your videos I take my knowledge and style and add in things you teach to advance my knowledge. Thanks again.

[gauravprasad]

It's good to have such good content available on youtube.

[chyldstudios]

Just excellent. More of videos like this please.

[benignma]

This was great. Would love to see more like this

[danielt.]

This was very interesting, thanks. I'd be interested in something PyTorch-related

[lunistic]

i love to see more video like this; Sir

[joliver]

Great stuff but you have to increase the font size if showing code cause we can’t read it. Either zoom in or change your desktop resolution. That’s what I do in my videos. Also do you watch your own videos before publishing?

[tempdeltavalue]

it would be better if you zoomed in on parts of the code (+ you have a bunch of white space on the screen)

[bediosoro]

Tensorflow keras is now getting better in terms of functionality. Except gpu installation which is the very bad thing that will make everyone migrate to pytorch. I just spent all night not being able to make tensorflow shows the number of gpu always solver not found.

[rs]

hello author,

i want to train a model to predict heatmap (mean square error loss) and binary segmentation (binary cross entropy loss).
i tried to train model using multi branch (2 branch duplicates for 2 output). but the the final output will favour for only one type of output.
For example when i train using model.fit with equal loss weights, the output is good for heatmap, but binary mask output is wrong and gives pixels 1 for the regions similar to heatmaps.
And when i train using gradtape loop, the output is good for segmentation mask, but heatmaps are wrong and looks like masks.

how can i solve this, please give me your suggestion.

thank you

[rs]

hello,
i have some problem using custom dice loss function for multiclass segmentaion. My labels are one hoted.
my code works with keras inbuilt categorical cross entropy loss, but when I use custom dice loss function to iou, the loss reduces up to 3 epochs and remains almost same, even if decreases slowly for long epochs, the output is juts blank. What might be the reason? i have tried a lot of custom losses from the internet but all fails.

Thank you.

[jonwolgamott]

How is this different than a filter that say Snapchat or IG is using to apply to their videos? how do they process live streams?