Zalo AI Challenge 2023 - Advertising Banner Generation
T-sharp Team

We are pleased to announce our 1st-place ranking on both the public leaderboard and private leaderboard.

The poster shown at Zalo AI summit 2023.

Docker deployment

1. Docker build

   $ docker build -t zac2023:env_code_v15 .

2. Docker run

   $ export WEIGHT_PATH= path to pretrained models (folder structure described in Appendix)
   $ export INFO_PATH= path to info.csv
   $ export RESULT_PATH= path to result folder
   
   $ docker run --rm --gpus="device=2"  \
   -v $WEIGHT_PATH:/code/weights \
   -v $INFO_PATH:/data/private/info.csv \
   -v $RESULT_PATH:/results \
   zac2023:env_code_v15 \
   /bin/bash /code/predict.sh

3. Results

   After running the script, you can find the results in RESULT_PATH

   

   Sample 91.jpg

   

Appendix

Pretrained models team used in this challenge are: Kandinsky2.0 and Finetuned Qwen7B.

To gain better speed in keywords extration, we quantized the model Qwen into 8bit using AutoGPTQ

After all, we placed models in folder WEIGHT_PATH as following structure

Authors

• Trần Hữu Thức
• Trần Hậu Tuấn
• Nguyễn Hữu Thắng