woodongk / story-blender

๐Ÿค– โžก๏ธŽ ๐Ÿง‘๐Ÿปโ€๐Ÿฆฐ Human-Like Story Generation from Caption Using Seq2Seq Model

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

seq2seq ๋ชจ๋ธ์„ ์ ์šฉ์‹œํ‚จ ์Šคํ† ๋ฆฌ ๋ฌธ์žฅ ์ƒ์„ฑ ์—ฐ๊ตฌ

  • ๏ธŽHuman-Like Story Generation from Caption Using Seq2Seq Model ( ๐Ÿค– machine-like โžก๏ธ ๐Ÿง‘ human-like )
  • 2018๋…„๋„ ์•„์ฃผ๋Œ€ํ•™๊ต ๋ฏธ๋””์–ดํ•™๊ณผ ์กธ์—… ํ”„๋กœ์ ํŠธ ์ตœ์šฐ์ˆ˜์ƒ ์ˆ˜์ƒ
"the fireworks are shooting off in the sky" -> [Seq2Seq model] -> "the fireworks were beautiful"

Requirement

  • tensorflow 2.2.0
  • keras 2.4.3

ํ”„๋กœ์ ํŠธ ๋ชฉ์ 

  • Sequence-to-Sequence (Seq2Seq) ๋ชจ๋ธ์€ ์ฃผ๋กœ ํ•œ ๋„๋ฉ”์ธ์ธ(์˜ˆ: ํ•œ๊ตญ์–ด ๋ฌธ์žฅ)์—์„œ ๋‹ค๋ฅธ ๋„๋ฉ”์ธ(์˜ˆ: ์˜์–ด๋กœ ๋ฒˆ์—ญ๋œ ๋™์ผํ•œ ๋ฌธ์žฅ)์˜ sequence๋กœ sequence๋ฅผ ๋ณ€ํ™˜ํ•˜๊ธฐ ์œ„ํ•œ ๋ชจ๋ธ์„ ๋งํ•œ๋‹ค.
  • "๊ธฐ๊ณ„๊ฐ€ ์ƒ์„ฑํ•œ ๋”ฑ๋”ฑํ•œ ๋ฌธ์žฅ์„ ์ธ๊ฐ„์ด ์“ด ๋“ฏํ•œ ์–ธ์–ด๋กœ ๋ณ€ํ˜•ํ•˜๋ฉด ์–ด๋–จ๊นŒ?"๋ผ๋Š” ๋‹จ์ˆœํ•œ ์ƒ๊ฐ์—์„œ ์‹œ์ž‘ํ•˜๊ฒŒ ๋œ ํ”„๋กœ์ ํŠธ

1. ๋ฐ์ดํ„ฐ์…‹ : VIST

  • ๋ณธ ํ”„๋กœ์ ํŠธ๋ฅผ ์œ„ํ•œ ๋ฐ์ดํ„ฐ๋กœ ๋งˆ์ดํฌ๋กœ์†Œํ”„ํŠธ ์‚ฌ์—์„œ ์ œ๊ณตํ•˜๋Š” VIST(Visual Storytelling Dataset)์„ ์‚ฌ์šฉํ•จ
  • VIST๋Š” ์ฃผ๋กœ image captioning task์— ์“ฐ์ด๋Š” ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ, ํŠน์ • ์ด๋ฒคํŠธ๋กœ ๋ฌถ์ธ ์ˆœ์ฐจ์ ์ธ ์ด๋ฏธ์ง€๋“ค์„ ๊ฐ๊ฐ ์บก์…˜ ๋ฌธ์žฅ(descriptions for images in isolation, DII)๊ณผ ์ˆœ์ฐจ์ ์ธ ์Šคํ† ๋ฆฌ ๋ฌธ์žฅ(stories for images in sequence, SIS)์˜ ์Œ์œผ๋กœ ์ œ๊ณต
  • image captioning task์— ์“ฐ์ด๋Š” ๊ธฐ์ˆ ์€ ํ˜„ ์‹œ์ ์—์„œ ๋งค์šฐ ๋ฐœ์ „๋˜์–ด ์žˆ๊ธฐ์— ๋ฐ์ดํ„ฐ์…‹ ๋˜ํ•œ ์‰ฝ๊ฒŒ ๊ตฌํ•  ์ˆ˜ ์žˆ์—ˆ์Œ
  • [Code]

2. seq2seq ๋ชจ๋ธ

Reference

About

๐Ÿค– โžก๏ธŽ ๐Ÿง‘๐Ÿปโ€๐Ÿฆฐ Human-Like Story Generation from Caption Using Seq2Seq Model


Languages

Language:Jupyter Notebook 100.0%