nhtlongcs / ucc-ai-quest-2023 Goto Github PK

Cork is blessed with breathtaking landscapes and serene greenery. This year, UCC AI Quest will focus on stunning aerial images of a high-resolution camera to recognise vegetation patches in Irish natural places... It includes the release of a new dataset of realistic drone images for benchmarking semantic segmentation from various above ground levels.

Source: competition page

This repository contains the solution of team FSGWL (i don't know what it stands for lmao) for the competition.

Quick access to the inference notebook:

By the way, we are one of the teams that achieved top places in the leaderboard, so feel free to ask us anything about the competition. The slides for the presentation can be found here and the technical report can be found here.

Using conda, recommend mamba for faster solving time

conda env create -f environment.yml
conda activate ucc

By downloading the data from the competition, you agree to the competition's terms and conditions.

Unzip *.zip in data folder so it has the following structure

data/public/img/train
data/public/img/valid
data/public/ann/train
data/public/ann/valid
data/private/img/test

To ensure the data and the environment is setup correctly, run the following command. It should run without error

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. pytest tests/pkg/classics/
CUDA_VISIBLE_DEVICES=0 PYTHONPATH=. pytest tests/pkg/transformers/

Join team using WanDB https://wandb.ai/ucc-quest-23/

$ ucc-train -c <config-file-path> -o <override_arg1>=<value1> <override_arg2>=<value2> ...

More details in notebooks/train.ipynb

$ ucc-pred -c <inference-config-file-path> -o <override_arg1>=<value1> <override_arg2>=<value2> ...

More details in notebooks/make_submission.ipynb

Some special flags:

• global.find_lr=True : This will find the optimal learning rate for the config file, rerun when have minor change
• global.wandb=True: In the training code include some visualize code using wandb, please not set this value to False in the trainning mode.

After training, the checkpoints are stored in folder PROJECT_NAME/RUNID/checkpoints. We need to prepare a file named "results.json" for submission on CodaLab. Use the notebook notebook/make_submission.ipynb and replace the checkpoint path

there should be a file results.zip generated in the output directory. You should be able to submit the file results.zip now.