cuet-polimemedecode

the notebook for our main approach to the cuet polimemedecode datathon

problem statement

given any meme (as an image), the model classifies the meme as either "Political" or "NonPolitical"

pipeline

setup

training dataset used:

https://www.kaggle.com/datasets/zarifishmam/polimemedecode

this was the training dataset used (mainly for the RAG-retrieval)
the dataset is scalable and additional memes can be added directly under the Train/Image/ directory (the respective image names and classifications must also be included under the Train/Train.csv/ directory)

working directory:

kaggle/
├── input/
│   └── traindata/
│       └── PoliMemeDecode/
│           ├── Train/
│           │   ├── Image/
│           │   │   └── train0001.jpg      # Example meme file
│           │   └── Train.csv              # Labels for training
│           ├── Test/
│           │   ├── Image/
│           │   │   └── test0001.jpg       # Example test file
│           │   └── Test.csv               # Metadata for testing
│           └── sample_submission.csv       # Required submission format
└── working/
    └── submission.csv                     # Final generated output

sample setup:

the dataset can be directly used in the kaggle environment, or you can download it and configure DATA_DIR path as required

output

after setting the appropriate directories, the notebook will succesfully generate a submission.csv file where it classifies each of the images under the Test/Image/ as "Political" or "NonPolitical"

further details:

• for a detailed overview of our approaches to this problem, you're encouraged to read the report
  report

authors

• Zarif Ishmam
• Shafnan Wasif
• Zarif Mahir
• Ananta Debnath