boostcampaitech5/level1_imageclassification-cv-11

level1_imageclassification-cv-11 created by GitHub Classroom

Gender and age classification for wearing mask image

This project is the Naver Boost Camp CV11 team's submission code for the mask wearing status classification competition. When a picture of a person wearing a mask is given, it is a matter of classifying the person's age, sex, and whether he or she is wearing a mask.

Team Members

서지훈	이준하	이지유	이채원	최지욱

Environment

• OS : Linux Ubuntu 18.04.5
• GPU : Tesla V100 (32GB)

Usage

1. git clone

• git clone https://github.com/boostcampaitech5/level1_imageclassification-cv-11.git

2. Install Requirements

• pip install -r requirements.txt

3. Training

To solve this problem, we conducted 18 classifications including gender, age, and whether to wear a mask, experimented with multi-output-classicaiton by setting each as a separate output, and conducted many experiments in the training process, such as removing the background of a given image through segmentation or adding regression loss for age. The method of execution for this experiment is as follows.

3-1. single-output-classification

• python train.py

3-2. multi-output-classification

• python train_multi.py

Key Options

• --save_dir : save path
• --data_dir : input path
• --use_age : weight of mseloss(age)
• --seg : enable segmentation
• --model : model type (default: EfficientNet B0)
• --batch_size : batch size for training
• --criterion : criterion type
• --epoch : number of epochs to train

4. Inference

4-1. Inference without segmentation image

• python inference.py

Key Options

• --data_dir : input path
• --model_dir : model weight path
• --output_dir : output path
• --is_multi : enable multi output classification
• --model : model type (default: EfficientNet B0)
• --batch_size : input batch size for validing

result

Metric : f1 score

Model	f1 score
EfficientNet B0	0.93093
ResNet 18	0.88177
ResNet 34	0.93165
EfficientNet B1	0.94594
EfficientNet B2	0.95842
ViT Tiny (Patch 16, 384)	0.76766
ViT Small (Patch 16, 384)	0.77515

Best model : Efficientnet B2