Hyperparameter Optimization: This Tutorial Is All You Need

[@abhishekkrthakur]

Please subscribe to help me keep motivated to make awesome videos like this one. :)

[@keshavbansal5148]

there is so much information in one video, i love it... I started following your channel a few days back, and its just crazy how much information and skill you pack in each and every video.. thank you so much sir

[@abhishekkrthakur]

I'm glad you like it! Please consider subscribing too :D

[@keshavbansal5148]

@@abhishekkrthakur already subscribed sir

[@rajeeevranjan6991]

I have heard about you earlier and now I know you. You are awsm. So much to learn in just 1 hr.

[@DIYGUY999]

I like the way you are covering various topics in DL.

[@anassbellachehab9821]

I got the book. One of my best purchases, makes everything straightforward.

[@rajacspraman1791]

Indha video partha en English marandhurum pola (just for fun). Great video! Nicely explained!

[@AIPlayerrrr]

Is it possible to make videos on past kaggle competitions solution walkthrough? They are really valuable to kagglers.

[@abhishekkrthakur]

any competitions in mind?

[@AIPlayerrrr]

Abhishek Thakur anything is appreciated. Maybe start by walking through completions that just ended? I think this will bring the most views and your channel will for sure be the hottest channel of Data Science in the near future. I’ve been competing for a long time now on kaggle and some winning solutions are still very hard for me to understand. I’d love to learn from GM like you and I am sure many are like me :)

[@AIPlayerrrr]

@@abhishekkrthakur also i would like to thank you for your SIIM video, I was able to reach top 30 at the moment by tweaking your code(used EfficientNet instead). Also noticed that you changed your WTFML library to version 3.0.0 and I need to re-learn it:/

[@abhishekkrthakur]

@@AIPlayerrrr ohh nice. yeah, im changing it quite fast. you can just do "pip install wtfml==0.0.2"

[@abhishekkrthakur]

@@AIPlayerrrr Lemme think of some competitions :)

[@caminerin]

Incredible video, so much better than many webs.

[@nitishthakur3644]

Really a great Video! Bought your book. Was really helpful watching the implementation in video. Thanks a lot Abhishek.

[@imakonkonvicted]

Hi Abhishek. I am new here but within a day of watching and understanding I have learnt a lot! :D Thanks!

[@SidVanam]

Such a helpful tutorial for a new kaggler like me, thanks for taking the time to put this together!

[@cetrusbr]

Very nice video! Indeed, this tutorial is all you need to be minimally efficient with Hyperparameter Optimization...

[@amitsrivastava3744]

Dear Abhishek, Thanks so much for this video. You explained so beautifully. So much learnt.

[@leodzingirai6775]

Hi Abhishek. Your videos have helped me a lot. You inspired me to write code using VS Code as I was only used to notebooks. I am looking forward to your book. Thanks

[@abhishekkrthakur]

Thanks Leo.! Do you like the IDE way of development?

[@leodzingirai6775]

@@abhishekkrthakur Yes, I realized that using an IDE makes it so much easier

[@dhristovaddx]

You're amazing and I love your book. Very well-written and easy to understand! Great job!

[@abhishekkrthakur]

🙏

[@srilekhavinjamara8732]

Learning a lot exponentially by watching your videos. Can't get any better, this is really awesome!! Thanks a ton for sharing your knowledge :)

[@allieubisse470]

Another killer Tuts... informative walk through

[@narendraprasath5745]

I really liked the way you explain. Thanks for introducing lot of new concepts. I am expecting some tips for reading ml research papers.

[@zhubarb]

Terrific coverage, thanks a lot for taking the time to share this.

[@abhishekkrthakur]

Thank you so much!

[@emmanuelphilibus1035]

Thank you for this excellent presentation/tutorial! Are there other optimization methods that can be used apart from the ones you have covered in this video? Do you also work on metaheuristic optimization methods? If yes, I would like you to drop a video on it.

[@pushkarajpalnitkar1695]

Great video. In gridsearchcv mentioning integer for parameter cv creates that many STRATIFIED fold so no need to worry about the stratification there. Thanks for the video anyways.

[@romananalytics2182]

Again informative content with live coding! Can't expect for anything else!!

[@vidulakamat6564]

So informative and great explanation. A lot to learn indeed. Thank you Abhishek !!

[@АнтонАртемьев-я8г]

Thanks a lot, really nice overview of hyperparameters optimization tools!

[@krocodilnaohote1412]

Thank you, this video is very helpful

[@Rohsn1]

Absolutely awesome tutorial on this topic. Love your videos.

[@abhishekkrthakur]

Thanks so much!

[@tanujpahuja4749]

Loving the videos and the book. Keep it up!!

[@abhishekkrthakur]

Thanks so much!

[@khaledsrrr]

Keep them coming ❤

[@aliabdelkader3995]

Hi Abhishek, thanks for the great content. I am wondering if you could share more about your thought process. Like what do you see in an ML problem that makes you use hyperparameter optimization technique over the other ? I hope that you would consider this also in your upcoming videos. I usually come to your channel looking forward to learning how does great ML engineers think which is more important than what ML tools or libraries are there.

[@Tajriiba]

Hi Abihshek! thanks again for sharing all this valuable contente! plz can you tell me how you have got the auto completion in kaggle, Thank you again

[@mdkhalidsiddique5052]

Amazing explanation sir, and am really happy that i purchased your book. I wish you had written a chapter on outlier treatment as well, if possible make a video on that.

[@abhishekkrthakur]

Will make one :)

[@mdkhalidsiddique5052]

@@abhishekkrthakur thank you sir, Will be waiting eagerly!

[@GurinderGhotra]

Hi, can you provide some guidance to use hyperopt for time series

[@ayuumi7926]

Hi Abhishek, can u kindly make a video on text summarization? Thanks!

[@jitendrakumarsah8683]

Great Abhishek Sir👍

[@diniamalia3119]

are this tutorial can be implemented to time series forecasting using ANN?

[@rushikeshdarge6115]

Thankyou sir! Killerrr learnings, from winner 🤟🙏

[@mahdihosseinali7492]

great video. When do you usually optimize hyper parameters? Is it the last step? For example do you do all the feature engineerings/pre-processing with the default values and then optimize the hyper parameters? What if the default space is better suited for a set of features?

[@bhavinmoriya9216]

Thank you very much for the video. One thing I want to ask. The pipeline with standard scaling you defined and you passed it the random search. The randomsearch is done with cv. So every time validation set with be transformed using standardscaler and fit of the training or it performs fit transform on validation?

[@joedoe4249]

Great video, thank you! Whenever I try the grid search I get an error at line of df=pd.read_csv("../input/mobile_train.csv") that says "FileNotFoundError: [Errno 2] No such file or directory: '../input/mobile_train.csv' ". Do you know how to fix this?

[@dikshantgupta5539]

Hello sir Can you please mention how one technique is better than other? And which one we should use?

[@lokesh6234]

BOMB....all hail Abhishek

[@jitendersinghvirk47]

Hi Abhishek, Can you please tell me your approach to learn a new concept? Suppose when the transformers concept came out, how did you learn that? btw, thanks for your awesome book.

[@manishsharma2211]

Read blog, page , articles,documentation

[@abhishekkrthakur]

First, i heard about it, then i checked the paper. tried to understand it. took some iterations and then I went with applications.

[@nouyed]

@Abhishek Thakur Are you using Pycharm to write the codes ?

[@shreyjain6447]

Hi Abhishek. Can you please make a video on Distributed Computing using Optuna or Ray or whatever you feel is well suited? I think most of the learners would really benefit from that as most of us face issues like long run times for complex models. Thank you!

[@Ajaysharma-yv7zp]

How to do parameterr tuning using gray wolf optimizer kindly explain if possible in deep learning ...

[@베타랩스]

Thank you, your lecture was very informative. The n_jobs specified does utilise only CPU. But how can I utilize the GPU too.

[@ahnafsamin3777]

Can I use a fixed validation set with GridSearchCV/RandomSearchCV?

[@gustavoaparecido6981]

How did you install the hyperopt intellisense vscode?

[@rahulsingh5592]

Is there any order in which the the models do hypertuning? Like first max depth, then learning rate.... something like that?

[@MustafaCam-uv5fr]

is there hyperparameter for cnn

[@prashlovessamosa]

Thanks man

[@muhammadmohsin7554]

great video. can you do a video on how you decide which Hyperparameter to tune and what should be the search space for each parameter.

[@Falconoo7383]

Thank you for this video. Sir, i have defined a CNN+LSTM model for EEG signal classification but the training accuray goes to 90-95% and the validate, test accuracy 30-40%. Can you please help me? I also modify the architectures and change the hyperparameters values but never get the solution. I am facing this problem from last 2 months. Can you please help me to figure out this problem?

[@kuberchaurasiya]

Excellent tutorial!! But I found hyperopt is not able to use multi core. Any option to do that?

[@ramjondhale8758]

@Abhishek Thakur is there any method for selecting best model for the data in less time or we have to test it manually for every model

[@hosseiniphysics8346]

tnx

[@shashidharmuniswamy2620]

Hello, thank you for the video. I'm a physics student who is interested in implementing your lesson on Grid search, Random search, and Bayesian optimization for a physics-based dataset. I have a few questions, please. 1) Can optimal hyperparameters for a regression problem be informed by categorical values as features? And the best set categories be evaluated? If yes, what considerations do I have to keep in mind? 2) In grid search, you obtained the parameter values for n_estimators, max_depth, and criterion, but how does inform me of the best set of input parameters for my output target measure?

[@rajanlagah4529]

Did hyper opt for 2k max_evals and it perform lower than random search and BOTH PERFORM BAD THAN DEFAULT values of random forest. XD any idea why ?

[@s.sasisekhar4608]

very good content

[@alay9526]

sir can u explain about parameters? like verbos,n_jobs

[@anirudhg7861]

Is it possible to use Optuna with SKLearn pipelines. If so, is there a reference I can look at?

[@JJGhostHunters]

Hello...Great video! I hope to buy your book soon! How would you approach finding an optimal MLP model by finding the best combination of number of layers, neurons per layer, choice of activation function, learning rate etc? Then, how would this be extended to CNNs and LSTMs? That is to select overall best models by finding the best combination of number of convolutional layers, number of filters, stride, etc for CNNs and for LSTMs, the number of LSTM units, etc? I would like to use Bayesian Optimization as you have shown in this video and it looks like it may be capable, however I am not sure how to go from what you are doing in this video to what I described above for MLPs, CNNs, and LSTMs. Any feedback would be greatly appreciated!

[@harshgupta3641]

while doing optimization do we need also need to drop the unnescessary features from the dataset , or fit on whole dataset and if we have unbalanced dataset then what we have to do

[@erfannariman]

Question: this looks really handy when you have to train one model. But I am doing a project right now where we (re)train and predict prices for roughly 500 articles daily. These articles are different so there's a different model for each one of them. Doing hyperparameter optimization for all 500 of them would take quite long right? What would you advice here.

[@abhishekkrthakur]

yeah. that would be very expensive. well, i would try to make a generic model, e.g. an nnet over all articles. and then you can finetune it for individual articles. well, i know its not hyperparameter optimization but just a thought. what do you think about this?

[@erfannariman]

@@abhishekkrthakur Thank you, sounds like a logical approach, but not sure how the "finetuning" step would look like, are we using a technique for that, or finetuning by weights or something else? Can you elaborate on that, or maybe link to papers / articles maybe?

Hey Abhishek I bought the ebook version of your book. and it doesn’t open on the Amazon browser reader...only on devices that can run their Kindle app. Could you enable that ?

[@kumarvis01]

Do we have git link for the above tutorial

[@cientifiko]

why not use train_test_plit stratify?

[@clivefernandes5435]

Sir will a bert model benefit standard preprocessing steps like stemming , lemmatisation ?

[@petergibson7209]

This is great! You don't happen to have a version of the code in a Github repo, do you?

[@abhishekkrthakur]

I'm sorry. I don't have github for this. This is more of a code-along video :)

[@nishantkumar3997]

HI @Abhishek, Thanks for you're great videos. However, data cleaning is one of the more important parts of a Data Science Project at companies. Can you make a video regarding data cleaning or data integration ? The initial phases of a Data Science Project .

[@SanjogMehta]

I am consistently getting this error even after installing, unstalling, reinstalling hyperopt through conda and even through pip, but every time it's the same during import. Any prerequisite before installing or anything am I missing? from hyperopt import hp, fmin, tpe, Trials ImportError: cannot import name 'hp' from 'hyperopt'

[@VjayVenugopal]

hello sanjog, change your file name from hyperopt to someother thing, then run the command pip uninstall bson && pip install pymongo i think it will help

[@HipHop-cz6os]

How are you coding in vscode without seeing the dataset parallelly?

[@LanguageHelp]

So I ran hyperopt the same way you ran it on my Pan Shared Task data 2019 (you can google it). It told me that it saw a configuration where the accuracy is 91%. I entered the best parameters hypeopt suggested on the same data and got 88% accuracy. I tried the default classifier on the same data and got 0.9016129032258065. Why is the default classifier better than the tuned one? Also, the data is balanced and split according to a certain ratio not randomly.

[@ansylpinto2301]

great video

[@vishalsiram1305]

how you're using vs code on localhost?

[@liocasa1974]

Hi there! Has anyone done any hyperparameter tuning using Isolation Forest on any dataset? I can't manage to get results via jupyter notebooks. I am learning Machine learning and applying it to my research. Any information or guidance would be highly appreciated. Thanks

[@_._gh0st__]

Sir is there any possible way to contact You I need your help for my project

[@smritisingh4090]

How should I integrate Bayesian optimization with sklearn pipeline

[@abhishekkrthakur]

Have you gone through the video?

[@smritisingh4090]

@@abhishekkrthakur yes..but didn't get where to exactly put that integration code for pipeline containing preprocessing steps and models

[@hiteshvaidya3331]

While I really appreciate you creating this content, I found it really hard to conceptually understand it. Next time onwards, in addition to syntax, may I request you to please explain things at least on a higher level so that things become pretty clear? Again, thank you for making such type of content.

[@bii710]

Why did you keep n_job=-1. Please guide

[@joao_ssouza]

n_jobs=-1 means that the model can use all CPU's cores to predict new instances. That is, the prediction of various new instances can be parallelized.

[@titashbboy]

Okay so the stream ended when I was asking the question 😂 so the question is, if i am using the using same pipeline as you showed with a StandaradScaler and the PCA, is there a data leakage between the validation fold and the training folds in the cv of the search algorithm?

[@abhishekkrthakur]

Hey, if you use pipeline then it is okay. if you fit pca separately, then its not okay.

[@regivm123]

We notice that you have used the entire data (X, y) for tuning the model. This makes lot of sense. However, there are many other professionals who disagree. They insists that tuning should be performed on the Train data to prevent data leak. I feel this is a weak argument. If we proceed this way, it looks like we are developing a model to fit the Train data. I wonder if you or viewers would have any comment on this.

[@abhishekkrthakur]

in case you watched the video, you will hear me saying that im doing this only because its faster for me while making this video. all optimizations must be done in a cross validated fashion.

[@regivm123]

@@abhishekkrthakur Thanks Abhishek. I agree with CV. But the clarification required is; use entire data or only Train data for Hyperparameter tuning. It will be nice to know your opinion.

[@richarddawkins9369]

😍 😍 😍

[@anirbancts2796]

Excellent tutorial, where can we find the code used in this tutorial ? Is it the book given here ? thanks

[@arpanghosh3801]

Can you share the code

[@abhishekkrthakur]

Not for this tutorial. Its easy to watch and implement.

[@bhaskartripathi]

Where is the link to code ?

[@abhishekkrthakur]

code along. no links

[@martinstu8400]

indian...