Hyperhyperparameter optimization, the next big thing in machine learning
@mohammadelghandour1614
Жыл бұрын
Amazing. Thank you !!
@X_platform
5 жыл бұрын
This stuff is super cool! Did not know there is a library :)
@HeatonResearch
5 жыл бұрын
Yes, I highly recommend, has save me lots of time/tuning.
@fredericnicholson80
13 күн бұрын
Jeff, you are the best, simply love your content! going a little deeper, I see that Keras itself comes with a Keras Tuner that offers "Bayesian" as a search model. do you have any experience with it and if yes, how do these two modules compare?
@HeatonResearch
13 күн бұрын
I’ve seen that option, but have not been doing a lot with Keira’s lately, would love to hear how that goes.
@fredericnicholson80
5 күн бұрын
@@HeatonResearch still working on it myself. do be honest, the program runs, but the resulting parameters are not optimal .
@juliananegrinidearaujo5653
4 жыл бұрын
Thanks for this content! Just what I looking for on how to actually do it!
@Wjv1234
4 жыл бұрын
What if you want to have a logarithmic search space for learning rate or maybe a categorical for the number of filters? How would you implement that?
@kolakareem2441
2 жыл бұрын
Can anyone share any content or file about Bayesian optimization of LSTM regression problems (not classification) please? I understand this video addressed neural network model. Thank you.
@jiahao2709
4 жыл бұрын
How about the Keras Tuner Lib compared with this?
@kolakareem2441
2 жыл бұрын
I can't quantify my gratitude to you for the help you render to us all. Sadly, I have been stuck with hyperparameter optimization for my LSTM regression problem for two months. I have exhausted all possible manual tuning approaches, gridsearched for days but the results are just terrible. I have tried to apply Bayesian optimization to my LSTM regression problem but I experienced this error when I evaluated the model (Found input variables with inconsistent numbers of samples: [10127, 12784]). I guess the LSTM data structure formatting might be wrong. I really need your technical view please.
@JJGhostHunters
2 жыл бұрын
Hello Kola...Have you found a solution? Do you have any resources for how to implement LSTM for regression?
@kolakareem2441
2 жыл бұрын
@@JJGhostHunters No Jeremy. Would you like to help?
@veerasrikanth3556
4 жыл бұрын
Hi, It's a very nice Content, one small doubt here, Can we pass categorical Hyperparameters to "BayesianOptimization" function? Say(activation functions) I tried passing them but its throwing error, I tried searching online but couldn't find any relevant results.
@thepresistence5935
2 жыл бұрын
you can but before you need to know what objective function and surrogate function you are using.
@user-tf7yj7dj2d
2 жыл бұрын
so how to save the tuner? please tell me ,please
@ibanguniverse811
4 жыл бұрын
Can you do this on google colab please ...
@paulchristian1244
4 жыл бұрын
Jeff, Thanks for your videos. They are awesome. One question, from the following code it appears that layer is always 0? I do not see where this is incremented or how its value ever changes? Am I missing something? Thank you. layer = 0 while neuronCount>25 and layer
@HeatonResearch
4 жыл бұрын
Excellent point, let me take a deeper look at that. I am in the process of reviewing everything for next semester right now.
@DefensorVieira
4 жыл бұрын
Hi, thank you very much for your videos. they are awesome. One question: how would you combine this with cross validtion search for the best number epochs? thanks
@HeatonResearch
4 жыл бұрын
Best number of epochs will depend on the hyper-parms. So, I usually search for the hyper-params with an early stop, then once I lock in on the params, I then search for the best number of epochs.
@DefensorVieira
4 жыл бұрын
@@HeatonResearch Great that's what I was thinking. Thank you very very much again for the awesome videos.
@chenghungchou9521
3 жыл бұрын
Is Bayesian Hyperparameter Optimization similar to Bayesian Neural Network?
@jayshete5219
5 жыл бұрын
One question here... Can we apply this technique or hyperopt to a CNN model doing image classification?
@HeatonResearch
5 жыл бұрын
Yes, it works on any type. The trick is representing your CNN hyperparameters as a fixed-length vector. You would do it similar to the method that I did. It would now be filter counts, number of CNN and MaxPool layers, etc.
@leonhma
3 жыл бұрын
I think keras_optimizer is an easier option
@devonk298
4 жыл бұрын
I have nothing against Bayes as long as they don't flaunt their lifestyle around the children.
Пікірлер: 36