정리왕 지제로사

This article is provided by...

Stop doing this on ChatGPT and get ahead of the 99% of its users | by Josep Ferrer | Geek Culture | Apr, 2023 | Medium

[0] Stop doing this on ChatGPT and get ahead of the 99% of its users

• Unleash the Power of AI writing with Effective Prompts
• have you ever felt frustrated with AI-generated content?
  • you think the ChatGPT output is completely horrible
  • and doesn’t satisfying your expectations?
• getting quality output from AI writing tools
  • depends heavily on the quality of your prompts
  • by training ChatGPT
    • you can get a personal writing assistant for free!
• time to discover The art of crafting powerful prompts
• to make the most of this cutting-edge technology.

[1] Why the bad results?

• The problem often lies not with the AI itself,
  • but the limitations and vagueness of the input provided.
  • instead of expecting the AI to think for you
  • you should be the one doing the thinking and guiding the AI to perform the tasks you need
• Bad results?
  • you are feeding ChatGPT poorly written and short prompts
  • and expecting some magical output to happen
    • ”Generate a LinkedIn post about AI for my account”
    • “Suggest me some ideas to write about programming” … etc.
  • you should stop!
  • giving prompts like this, ChatGPT has to make too many decisions
    • which generates some poor output
  • Poor instructions = Poor results !!!

[2] 4 main steps to assess the prompts and results

#1. Understand your needs and requirements:

• to order something, you first need to know what you want.
  • what you want from the AI
  • why you want it
  • how you want it delivered
• This clarity will help you create better prompts and enhance the quality of the output.
• start with standardizing all types of outputs you require from AI.
• (for examples)
• “I am starting to be active on my Twitter account – so I would like both tweet ideas and Twitter structures.”
  • Ideas to write tweets for Twitter.
  • Threads structures for Twitter.

#2. Treat AI like a digital intern:

• imagine you’ve hired an intern
• you wouldn’t give them only one brief explanation and expect them to do everything great at first.
  • you should provide your AI with a detailed checklist
  • explain the purpose behind the task
  • be open to clarify any doubts the AI might have.
• “Hey ChatGPT. Write me a Twitter thread about the Google Cloud Platform” 
  • this prompt is way too vague. >> letting an AI too many decisions to make for you.
    • How many tweets do you want?
    • What writing style?
    • What subtopic should ChatGPT emphasize?
    • What is my language tone – friendly, professional..?
• Always use AI tools to leverage your work – not to substitute you.

#3. Create constraints and avoid assumptions:

• This is the key to the process.
• To get specific and accurate output >> provide your AI with clear and well-defined information.

1. A good contextualization
   • what kind of output do you want?
2. A specific topic
   • with subtopics to emphasize.
3. A specific structure
   • like how may tweets, words..
4. A specific format for the output
   • what writing style to use, what tone..
5. A specific list of thing to avoid
   • what you do not want to mention.

for example.....

• Want ChatGPT to generate a Twitter Thread for me.
  • Hey ChatGPT. I am going to train you to create Twitter threads.
• Make sure ChatGPT understands the mean by the Twitter Thread.
  • Twitter threads are a series of tweets that outline and highlight the most important ideas of a longer text or some specific topic.
• Want ChatGPT to write with a specific topic.
  • The twitter thread will talk about Google Cloud Platform free tier services.
• Outline what I want it to mentions for sure and what to emphasize.
  • I want you to talk about Google cloud platform environment, all its services and its utility for data science. However, you need to emphasize all these advantages are free for ever when your usage is contained within some tiers. Mention google big query and cloud functions, that are two of the most important services for analytics.

• Specify the whole structure I want to receive from ChatGPT.
  • A first tweet with a short but concise message, letting people know what’s the thread about. It is important not to be more than 30 words, use key hashtags and convince people to read the whole thread. Emphasize the utility of the thread for them.
  • A second tweet that makes a short intro and let the reader get contextualized and understand why are they still reading the thread. It is important to keep the reader reading.
  • 4 or 5 tweets that outline and describe the most important parts of the article. These should summarise the main ideas of the topic I explained to you before.
  • A last tweet with some conclusions and letting people know why your thread is worth it.
  • A final tweet inviting them to retweet your thread and follow you.

• Specify the format of the output I want.
  • I would like you to behave as a technology and data science writer. Use natural and engaging language. It is important to use easy-to-understand vocabulary — remember I want to break complex concepts down to everyday words.
• A specific list of things to avoid.
  • Avoid using complex vocabulary.

#4. Iterate and refine your input:

• if the AI generates incorrect output.
  • probably due to an issue of your input.
  • Don’t be afraid to rework your prompts.
• to communicate with your AI machine
  • even though you use a natural language,
  • you should think you are writing a code for AI.
• Repeat and regenerate the response as many times as you want, so you get a some good result. 

[3] Conclusions

• It’s not the AI that’s falling short, but rather the way we interact with it.
• To make the most of ChatGPT and similar tools,
  • we must refine our approach
  • and focus on becoming thinkers who guide the AI in executing.
• By following these tips and taking responsibility for the input,
  • you’ll find that AI-generated content
  • can be a valuable asset in your content creation arsenal.
• So, let’s start crafting effective prompts and unlock the full potential of AI writing!
• And always remember…
  • Using AI won’t make you a cheater — but not using it will leave you behind!

Hyperparameters in Deep RL with Optuna

Hyperparameters in Deep RL with Optuna

[1편 문제: Hyper-parameters]

[2편 해결: Bayesian Search]

[3편: Hyper-parameter Search with Optuna]

총 3편으로 나눠서 Deep Reinforcement Learning에서의 Hyperparameter 의 중요성과

다양한 Hyperparameter를 선택하는 Search 방법을 소개하고

Optuna 사용 방법을 익혀보겠습니다.

[1] Optuna

* Optuna

https://optuna.org/

[2] Start Hyper-parameter Search with Optuna

git clone https://github.com/Paulescu/hands-on-rl.git

pip install –r requirements.txt

pip install mlflow
pip install jupyter
pip install optuna
pip install torch

load_ext autoreload
autoreload 2
pylab inline
config InlineBackend.figure_format = 'svg'

2) env load

import gym
env = gym.make('CartPole-v1')

3) terminal에서 03_cart_pole 폴더로 이동

각자 환경에 맞는 디렉토리로 작성

cd ~/hands-on-rl/03_cart_pole

4) MLflow

• ML 모델 운영화 목적 (MLOps)
• MLFlow Tracking
  • 모델 개발 중에 필요한 모든 것을 추적
  • 많은 하이퍼-파라미터로 run하는 실험 진행 시, 매 run마다 정확한 구성을 Log
• terminal에서 다음 명령어 실행

mlflow ui --backend-store-uri mlflow_runs/

ex. Log 예시

* MLflow

https://mlflow.org/

5) MLflow Setup and URL

아래 python 파일을 run

연결되면 URL 클릭 (필자의 경우 아래의 링크)

http://127.0.0.1:5000

[3] Optimize Hyper-parameters with Optuna

• hyperparameter sampling
• training
• and evaluation of RL agent

• The objective function: returns the metric we want to optimize
  • 리워드 값 최대화 목표

[4] Objective function

• Samples hyperparameters
• trains the agent 200 episodes and evaluates it 1000 episodes
• output is the average reward in 1000 episodes

6) Optuna study object 생성

[5] Optimize Hyper-parameters with Optuna (con.)

7) Optuna에게 study.optimize() 로 search 시작 (100번 시도)

• perfect agent가 나온다면 평균 reward가 500점일 것
• 그럼 hyperparameter 값들 도출 성공

• agent_id : 298
• mean_reward : 500
• optimized result with information of good hyperparameter

8) CartPole을 완벽하게 풀이하는 agent=298

9) 결과 이용하여 agent 평가

agent 평가 진행하기

[6] Summary

Hyperparameters in Deep RL with Optuna

Hyperparameters in Deep RL with Optuna

[1편 문제: Hyper-parameters]

[2편 해결: Bayesian Search]

[3편: Hyper-parameter Search with Optuna]

총 3편으로 나눠서 Deep Reinforcement Learning에서의 Hyperparameter 의 중요성과

다양한 Hyperparameter를 선택하는 Search 방법을 소개하고

Optuna 사용 방법을 익혀보겠습니다.

[1] 해결: Bayesian Search

https://towardsdatascience.com/a-conceptual-explanation-of-bayesian-model-based-hyperparameter-optimization-for-machine-learning-b8172278050f

• Bayesian Search
  • superior to random search
  • perfect choice to use in Deep RL

• differ in how they build the surrogate model
• will use Tree-structured Parzen Estimator (TPE) method.
  •  uses applying Bayes rule
  • probability of the hyperparameters given score on the object function

 

 

 

Hyperparameters in Deep RL with Optuna

Hyperparameters in Deep RL with Optuna

[1편 문제: Hyper-parameters]

[2편 해결: Bayesian Search]

[3편: Hyper-parameter Search with Optuna]

총 3편으로 나눠서 Deep Reinforcement Learning에서의 Hyperparameter 의 중요성과

다양한 Hyperparameter를 선택하는 Search 방법을 소개하고

Optuna 사용 방법을 익혀보겠습니다.

[1] 개요: Hyper-parameters in Deep RL

[2] 문제: Hyper-parameters

1. choose a set of hyperparameters
2. train the agent
3. evaluate the agent
4. if we are happy with the result, we are done.
   Otherwise, we choose a new set of hp and repeat the whole process.

[3] Grid Search

• trying all possible combinations and select the one that works best.
• method called grid search
• works well for many supervised ML problems

[4] Grid Search Problem

[5] Random Search