2023 Samsung AI Challenge : Image Quality Assessment - Mos

• 카메라 영상 화질 정량 평가 및 자연어 정성 평가를 동시 생성하는 알고리즘 개발

• 종합 등수 : 6th

• coworker : 화질 정성 평가 code

Member🔥

박주용	허건혁

---

Index

• Competition imformation
• Code reproduction

---

Competition imformation

• 주관 : 삼성전자 SAIT
• 운영 : 데이콘
• 대회 : link

대회 기간 : 2023-08-21-11:00 ~ 2023-10-02-10:00

Input : 사진
Output : 화질 평가 점수(float)

평가 산식 : $(PLCC + SLCC)\over2$

---

Set Environment

1. clone repo

2. build image

docker build -t my_image:1.0.0 .

(you can change name and tag)

3. docker run

docker run -it --name {name_you_want} --gpus all --ipc=host my_image:1.0.0 /bin/bash

---

Code reproduction

※ raw_data를 data folder에 저장

# Data preprocessing
train_image_path : /data/train
test_image_path : /data/test

# Model Train
cd /lightning-hydra-template/src
python train.py model=maniqa_{model_size}_model data=maniqa_{model_size}_data trainer.devices={num_your_device} trainer=ddp

{model_size} = 384,448,640

# Another fold for training
default fold is fold0. (There are fold0,1,2,3,4)

if you want another fold for train,
you should change train,val_csv_file directory at 

'/lightning-hydra-template/configs/data/maniqa_{model_size}_data.yaml' 
{model_size} = 384, 448, 640

Ex) fold1
train_csv_file: ${paths.data_dir}/train_only_mos/train_df_fold1.csv
valid_csv_file: ${paths.data_dir}/train_only_mos/val_df_fold1.csv 

---

Insert your weight.ckpt into /lightning-hydra-template/weight

# Model Inference
cd /lightning-hydra-template/src
python eval.py ckpt_path=../weight/{your_weight_file} model=maniqa_{model_size}_model data=maniqa_{model_size}_data model.name={name_of_your_inference_csv_file}

---

Score Ranking

Type	score	Rank
Public	0.7874	6
Private	0.57225	5

---

report

코드 재현을 위한 상세 설명