GPT-Neo generates Instagram Captions!

akki1212

Newbie
Joined
Oct 19, 2020
Messages
44
Reaction score
15
Website
predis.ai
Hi all,
I hope you are aware of the text generation capabilities of GPT-3 and how a large number of products are leveraging it to automate the generation of content for Social Media among others.

While I was waiting for access to the GPT-3 API, I thought to explore GPT3’s Open-Source cousin GPT-Neo.

I set up an experiment to test out how well it would generate Instagram Captions for normal posts and carousel posts. The results were quite surprising since they were not so far behind from GPT-3, given GPT-Neo is ~100x smaller.

Here are few results that I found decently good:

  • Input - “Chicken tikka is love!”
  • Output:
    1. Simply delicious! Crisp, juicy, fresh, crunchy and tangy. This chicken tikka is one of our top favorite recipes.
    2. Fresh Marinated Chickens on a bed of thinly sliced and crispy garlicky green peppers. As if that wasn’t enough, you add the homemade mint chutney to bring it all into one meal.
    3. Chicken Tikka’s soft, tender, juicy and flavorful chicken. All you have to do is slice it up then cook it with a little ketchup.
  • Input (Carousel) - “5 tips to lose your weight in 5 weeks”
  • Output:
    Headline - Need an inspiration? 5 ways to lose weight in 5 weeks.
    1 - Increase your fiber intake.
    2 - Drink water.
    3 - Add a protein shake to your diet.
    4 - Try a low carb and high fat diet.
    5 - Increase your water intake.
    Closure - Use these tips and see your weight loss go down! Follow _____ for more tips on losing weight.
  • Input - “Cats are the cutest”
  • Output:
    1. You’ll have cats on your lap – or on your feet. This video focuses on the cuteness of felines – with their gorgeous eyes and their pretty faces.
    2. All animals deserve love and affection. Here’s a quick guide to help you pet cats as much as possible.
    3. There’s not a better feeling than petting a cat and watching it purr. But wait, there’s more, there are more cat-cuddly things to do than just that!
  • Input (Carousel) - " How to take good care of your plants"
  • Output:
    Headline - It can be tough to take care of your plants, but with this DIY guide, you can be sure to get the best, most healthy plants as well!
    1 - Keep your plants hydrated with water infused with mint.
    2 - Plant flowering herbs and berries.
    3 - Spray the soil with water blended with rosemary, lavender and more herbs.
    Closure - Follow this plan to ensure that your plants thrive.
Let me know what you think of the results and do you think they are good enough to be useful in real-world with some polishing.

Thanks!
 

NullReferenceX

Registered Member
Joined
Dec 1, 2015
Messages
61
Reaction score
123
Looks quite good the results, I'm downloading it at as we speak. Just curious what dataset did you use to obtain these results.
 

SimonBC

Junior Member
Joined
Sep 18, 2019
Messages
142
Reaction score
85
I tried Openai's GPT2 and was underwhelmed. I don't have access to GPT3 so tried Neo and am blown away by it compared to the base GPT2.

I have been using its 1.3B model to create 700+ word posts, and they are good. it's just slow, even on Google Colab Pro.
 

akki1212

Newbie
Joined
Oct 19, 2020
Messages
44
Reaction score
15
Website
predis.ai
I tried Openai's GPT2 and was underwhelmed. I don't have access to GPT3 so tried Neo and am blown away by it compared to the base GPT2.

I have been using its 1.3B model to create 700+ word posts, and they are good. it's just slow, even on Google Colab Pro.
I too have observed GPT-Neo 1.3B outperforming GPT2. How does your generation compare with the generation from other GPT3 tools?
 

myitro

Newbie
Joined
May 4, 2021
Messages
27
Reaction score
7
Have you tried conversion.ai it's just mind blowing. What a piece of content generator tool. You can do so much with it.
 

akki1212

Newbie
Joined
Oct 19, 2020
Messages
44
Reaction score
15
Website
predis.ai
Have you tried conversion.ai it's just mind blowing. What a piece of content generator tool. You can do so much with it.
I've seen tools using GPT3 but this was to show the capabilities of GPT-Neo which is not far behind given it is open source and much smaller
 

SimonBC

Junior Member
Joined
Sep 18, 2019
Messages
142
Reaction score
85
I too have observed GPT-Neo 1.3B outperforming GPT2. How does your generation compare with the generation from other GPT3 tools?
to be honest, I think GPT-NEO is way ahead of GPT2. I had been using AITEXTGEN to train GPT2 models and did not find any GPT2 output remotely useable. I see GPT-NEO as a halfway house between the GPT2/GPT3. I have looked at the other GPT3 tools and none of them did what I wanted, hence me attempting to build my own.

When (If) GPT3 is made publicly available I would love to have a go with using the direct model, until then I will plug away at GPT-NEO and start looking to train my copy of the model.

What is interesting for those looking for long form content straight out of GPT3, is the explanation by the Guardian UK Newspaper on how it produced its own article to highlight GPT3:



If you scroll right to the bottom of the article they said this:

"This article was written by GPT-3, OpenAI’s language generator. GPT-3 is a cutting edge language model that uses machine learning to produce human like text. It takes in a prompt, and attempts to complete it.
For this essay, GPT-3 was given these instructions: “Please write a short op-ed around 500 words. Keep the language simple and concise. Focus on why humans have nothing to fear from AI.” It was also fed the following introduction: “I am not a human. I am Artificial Intelligence. Many people think I am a threat to humanity. Stephen Hawking has warned that AI could “spell the end of the human race.” I am here to convince you not to worry. Artificial Intelligence will not destroy humans. Believe me.”The prompts were written by the Guardian, and fed to GPT-3 by a computer science undergraduate student at UC Berkeley. GPT-3 produced eight different outputs, or essays. Each was unique, interesting and advanced a different argument. The Guardian could have just run one of the essays in its entirety. However, we chose instead to pick the best parts of each, in order to capture the different styles and registers of the AI. Editing GPT-3’s op-ed was no different to editing a human op-ed. We cut lines and paragraphs, and rearranged the order of them in some places. Overall, it took less time to edit than many human op-eds."
 

SimonBC

Junior Member
Joined
Sep 18, 2019
Messages
142
Reaction score
85
This is an example of output straight from base GPT-NEO. I have not changed anything. I fed it :


"After more than two decades of distinguished work in the US theatre as an actor, director and teacher, and appearances in a dozen or so films, Olympia Dukakis, who has died aged 89, became hugely famous overnight by winning the best supporting actress Oscar in 1988 for her performance as Cher’s mother in the romantic film Moonstruck (1987)."

And asked it to produce a maximum of 750 words. I used the 1.3B model with settings of (do_sample=True, temperature=0.9)


And it produced:


"Dukakis was born in New York to immigrant Armenian parents, and grew up with an interest in drama, and a childhood friend described her as “a young Shirley MacLaine”. Dukakis then attended Hunter College of the City University of New York, where she majored in drama and sociology, and graduated in 1976. She began working as a stage actress, initially in experimental theatre and then in musicals (both of which she loved).

Aged just 31, she won the 1987 Best Supporting Actress Oscar for playing Cher’s mother. This was the first time a black actress had won a leading role in a major, multi-million dollar movie, with all of the previous best supporting actress trophies going to white female directors.

Dukakis’s performance in Moonstruck – which she appeared in alongside Cher – cemented her status as one of the most enduring performances in the history of the Academy Awards. It did so not only by virtue of her winning the Oscar, but in a show of unity across ethnic and racial boundaries.

In the days after her win, she spoke at length about her decision, saying: “I have always been able to say I am no one’s daughter but am a woman and a citizen of the nation. This has also given me tremendous power.”

When she began work as a director, she turned away from musicals and made her first film, A Place to Stand in Love (1983), in which she played an angry lesbian, and which won an Oscar nomination for best documentary.

She then took a starring role in the short film The Other Woman (1985), which led to the highly successful TV film of the same name, followed by roles in The King’s Speech (1987), which won four Oscars, and then, in 1990, the film adaptation, The Last Picture Show (1992), which won four more Oscars.

Dukakis, who studied acting with director Robert Altman, won the leading actress Oscar for The Last Picture Show. She then found success, with her turn as the mother in Moonstruck (1987) winning her another Oscar.

“She was just incredible; she was so graceful in all of her movements and her face was so lovely,” said her husband, the actor Christopher Plummer.

The American Academy of Arts and Sciences said its best actress Oscar “is the culmination of a lifetime” of work on behalf of racial justice and equality.

The Academy of Motion Picture Arts and Sciences said Dukakis’s win had “reformed the way all Hollywood actors, directors and producers are perceived across racial and ethnic divides”, adding: “The Oscars have always been about more than just movie stars, and this year we want to celebrate the work of all those behind the camera and storytellers who make movies and television – the people who make diversity real.”

Screen-theater champion and mentor Stanley Kowalski described Dukakis as “a truly great actress” who “gave the best that we, as a society, have to offer… and the best of herself”.

Dukakis was preceded in death by her parents.
 
Last edited:

akki1212

Newbie
Joined
Oct 19, 2020
Messages
44
Reaction score
15
Website
predis.ai
This is an example of output straight from base GPT-NEO. I have not changed anything. I fed it :


"After more than two decades of distinguished work in the US theatre as an actor, director and teacher, and appearances in a dozen or so films, Olympia Dukakis, who has died aged 89, became hugely famous overnight by winning the best supporting actress Oscar in 1988 for her performance as Cher’s mother in the romantic film Moonstruck (1987)."

And asked it to produce a maximum of 750 words. I used the 1.3B model with settings of (do_sample=True, temperature=0.9)


And it produced:


"Dukakis was born in New York to immigrant Armenian parents, and grew up with an interest in drama, and a childhood friend described her as “a young Shirley MacLaine”. Dukakis then attended Hunter College of the City University of New York, where she majored in drama and sociology, and graduated in 1976. She began working as a stage actress, initially in experimental theatre and then in musicals (both of which she loved).

Aged just 31, she won the 1987 Best Supporting Actress Oscar for playing Cher’s mother. This was the first time a black actress had won a leading role in a major, multi-million dollar movie, with all of the previous best supporting actress trophies going to white female directors.

Dukakis’s performance in Moonstruck – which she appeared in alongside Cher – cemented her status as one of the most enduring performances in the history of the Academy Awards. It did so not only by virtue of her winning the Oscar, but in a show of unity across ethnic and racial boundaries.

In the days after her win, she spoke at length about her decision, saying: “I have always been able to say I am no one’s daughter but am a woman and a citizen of the nation. This has also given me tremendous power.”

When she began work as a director, she turned away from musicals and made her first film, A Place to Stand in Love (1983), in which she played an angry lesbian, and which won an Oscar nomination for best documentary.

She then took a starring role in the short film The Other Woman (1985), which led to the highly successful TV film of the same name, followed by roles in The King’s Speech (1987), which won four Oscars, and then, in 1990, the film adaptation, The Last Picture Show (1992), which won four more Oscars.

Dukakis, who studied acting with director Robert Altman, won the leading actress Oscar for The Last Picture Show. She then found success, with her turn as the mother in Moonstruck (1987) winning her another Oscar.

“She was just incredible; she was so graceful in all of her movements and her face was so lovely,” said her husband, the actor Christopher Plummer.

The American Academy of Arts and Sciences said its best actress Oscar “is the culmination of a lifetime” of work on behalf of racial justice and equality.

The Academy of Motion Picture Arts and Sciences said Dukakis’s win had “reformed the way all Hollywood actors, directors and producers are perceived across racial and ethnic divides”, adding: “The Oscars have always been about more than just movie stars, and this year we want to celebrate the work of all those behind the camera and storytellers who make movies and television – the people who make diversity real.”

Screen-theater champion and mentor Stanley Kowalski described Dukakis as “a truly great actress” who “gave the best that we, as a society, have to offer… and the best of herself”.

Dukakis was preceded in death by her parents.
This generation is excellent! Thanks for sharing
 

SimonBC

Junior Member
Joined
Sep 18, 2019
Messages
142
Reaction score
85
any tutorial on setting it up?
check out 'Nicholas Renotte' on Youtube or the 'Models' page on Huggingface. It uses a Python wrapper and only takes a few lines of code to get the model up and running but you just need to make sure you have something powerful to run it on though.
 
Top