The official implementation for the paper titled: "Concept Drift Guided LayerNorm Tuning for Efficient Multimodal Metaphor Identification", ICMR 2025.

📌 Note: We strongly recommend referring to the [arXiv] version, which contains updated content and refinements beyond the ICMR 2025 publication.

📌 注意：我们强烈建议参考 [arXiv] 版本，该版本包含 ICMR 2025 出版版本之外的更新内容和改进。

conda create -n CDGLT python==3.8.0
conda activate CDGLT
pip install -r requirement.txt

Download image data: here

Download label file: here

Download OCR content of memes: here

Download the dataset division file: here ( The 6/2/2 dataset division given by Vincy2King/M3F-MEME adopted in our experiments.)

The structure of data directoty:

-data/
    -Eimages/
        -Eimages/
    -avg_test_label_E
    -avg_train_label_E
    -avg_val_label_E
    -E_text.csv
    -label_E.csv

Note：The label_E.csv and E_text.csv offered by kaggle have some flaw in encoding. It can be solved by openning them with VScode then clicking Select Encoding -> Save with Encoding -> UTF-8. The repaired files is provided by us in data directory of this repository.

The pretrained CLIP model we used: openai/clip-vit-large-patch14

The pretrained GPT-2 model we used: openai-community/gpt2

Obtain the 6/2/2 train/val/test label files of each specific tasks:

# {YOUR_PATH}/CDGLT$
cd ./utils
python split_train_val_test.py

Then the split files will be written in data/E_split/

Task 0: Sentiment Analysis; 1: sentiment analysis; 2: Intention Detection; 3: Offensiveness Detection; 4: Metaphor Identification

The first columns in the obtained .csv files represent the IDs of images (e.g. the ID of file named image_ (26).jpg is 26)

# {YOUR_PATH}/CDGLT$
cd ./utils
python write_clipText_feature.py
python write_clipViT_feature.py
python write_gpt2_prompt_tokenid.py

Embeddings and GPT2 tokenIDs will be written in feature/cache_E

# {YOUR_PATH}/CDGLT$
bash ./train_MI.sh

This bash script uses the nohup command, so the python program will run in the background and redirect std output to a log file.