README.md

hypermeme

======= inspired by NicsMeme®

"In the future memes will be able to generate themselves and propagate automatically"

see also the Dead Internet Theory

---

For a complete presentation of this project see hypermeme.ipynb

Project overview

Brief description

The aim of this project is to categorize memes into 4 categories or topics.

• pol: politics
• ent: entertainment
• sport: sports
• oth: other

The classification is based on their visual content (ocr and caption).

Project structure

Setup

0. Make sure you got docker and wget installed on your machine as they are required for this project

1. Download Kafka

   cd ./kafka/setup
   wget https://downloads.apache.org/kafka/3.7.1/kafka_2.13-3.7.1.tgz 

2. Build containers

   # if you want gpu acceleration
   docker compose -f gpu_compose.yaml build
   # otherwise (cpu)
   docker compose build

3. Start pipeline (see quickstart)

4. Import dashboard+data_view.ndjson file from /kibana directory into kibana

Quickstart

    # if you want gpu acceleration
    docker compose -f gpu_compose.yaml --profile pipeline up
    # otherwise (cpu)
    docker compose --profile pipeline up

Download the dataset

    docker compose --profile download_dataset up

Model training

    # Pretrained model is already included with this repo. If you want to rebuild it using your own data you can use this command
    docker compose --profile build_model up

Useful links

Container	URL	Description
kafka-UI	http://localhost:8080	Open kafka UI
kibana	http://localhost:5601	Kibana base URL

Dashboard demo

Technologies Used

• Ingestion:
  • Python Script: download memes from selected subreddits using Python Reddit API Wrapper PRAW
  • Logstash
• Image data extraction
  • Image captioning model: blip-image-captioning-large su huggingface
  • EasyOcr
• Streaming:
  • Apache Kafka
• Processing:
  • Spark Streaming
  • Spark MLlib
• Indexing:
  • ElasticSearch
• Visualization:
  • Kibana