💊 Challenge Chat: My Solution Using Simple Trans...

To Vaccinate or Not to Vaccinate: It’s not a Question by #ZindiWeekendz

Helping Africa

$300 USD

Completed (almost 6 years ago)

Skills you will learn

Natural Language Processing

Classification

Sentiment Analysis

212 joined

126 active

Info Data Chat Leaderboard

Start

Apr 24, 20

Apr 26, 20

Reveal

Apr 26, 20

devnikhilmishra

My Solution Using Simple Transformers

Help · 27 Apr 2020, 09:43 · 17

Transformers are the state of the art for NLP tasks now, and as @Johnowhitaker mentioned, libraries like SimpleTransformers and Fastai do all the heavy lifiting of being able to utilize transformers for our use case. I used Roberta base and large models and in the end did an ensemble. Thanks to Zindi for organizing such an amazing competition. Hope there are many more competitions to come. Also congratulations to everyone who participated and made an effort to submit their solutions, you all did a great job.

Here's my solution.

https://github.com/nikhilmishradevelop/zindi-winning-solutions/blob/master/To%20Vaccinate%20or%20Not%20to%20Vaccinate%20It%E2%80%99s%20not%20a%20Question%20by%20%23ZindiWeekendz/zindi_weekendz_nlp_final_solution.ipynb

Discussion 17 answers

aninda_bitm

Thanks for sharing the winning solution

27 Apr 2020, 09:45

Upvotes 0

devnikhilmishra

You're welcome Aninda.

replied to aninda_bitm27 Apr 2020, 10:03

Upvotes 0

JavaBlack

Congratulations!! nice solution. Thank you.

27 Apr 2020, 09:52

Upvotes 0

devnikhilmishra

Thanks JavaBlack.

replied to JavaBlack27 Apr 2020, 10:03

Upvotes 0

Shankar

Indian institute of technology madras

Congratulations!! Nikhil

27 Apr 2020, 09:59 (edited less than a minute later)

Upvotes 0

devnikhilmishra

Thank you Shankar

replied to Shankar27 Apr 2020, 10:03

Upvotes 0

Rajat_Ranjan

Allstate

Thanks Nikhil for the solution, i had a similar solution using roberta but not the distant score as you have

Rank 4

https://github.com/rajat5ranjan/Zindi-Solutions/tree/master/To%20Vaccinate%20or%20Not%20to%20Vaccinate%20%23ZindiWeekendz

27 Apr 2020, 10:00 (edited less than a minute later)

Upvotes 0

devnikhilmishra

Thanks Rajat for the solution, its very elegant.

replied to Rajat_Ranjan27 Apr 2020, 10:02

Upvotes 0

RenierBotha

Thanks for sharing @devnikhikmishra. I see you've used blending in the final stage - would you care to point me in the right direction for finding weights when blending?

27 Apr 2020, 10:01

Upvotes 0

devnikhilmishra

Keep weights so that your overall validation score is maximized but don't fine tune it too much, as you can in my solution, the same weights I have applied to validaiton data is being applied to the test data as well.

replied to RenierBotha27 Apr 2020, 10:05

Upvotes 0

Paul_Okewunmi

Obafemi awolowo university ile-ife

Great solution Mishra.. I've always paid attention to your solutions and approach, Always giving us something to learn from..Thank you. I want to ask, when blending do you intuitively assign weights to the individual models and see how well it scores on the LB ? or how do you do it.

27 Apr 2020, 10:07

Upvotes 0

Paul_Okewunmi

Obafemi awolowo university ile-ife

You've answered my question already. Thanks

replied to Paul_Okewunmi27 Apr 2020, 10:09

Upvotes 0

devnikhilmishra

@Enigma, I set the weights so that, my validaiton score is maximized, then use it on the leaderboard, yes its a bit of intuition and experience with many competitions, that I have attended, but I don't fine tune it too much, not more than the 2nd place, too much fine tuning can lead to overfitting.

replied to Paul_Okewunmi27 Apr 2020, 10:10

Upvotes 0