[패스트캠퍼스 Upstage AI Lab 2기 부트캠프] #08_미니 프로젝트 - Upstage 경진대회 #2

# 01. 프로젝트 소개 : Document Type Classification

Task : Classification

Evaluation Metric : Macro F1 score

 

프로젝트 기간 : April 11th, 2024 - April 23rd, 2024 (19:00)

 

Given Dataset : 17 classes of document images

• Train Data : 1570

• Test Data : 3140

 

 

# 02. Project Process

 

 

# 03. EDA

• Analysis on label distribution of train data
• Analysis on size/hue distribution of train and test data
• Analysis on augmentation techniques used in test data

 

# 04. Data Processing

• Train data : data cleansing, upsampling, offline augmentation with augraphy
• Test data : rectification
• Online augmentation with Albumentation in models

 

# 05. Model

• Integrated version : using ResNet 152 and EfficientNet b4 to classify 17 types
• Individual version : using EfficientNet b4
  • step1 : 3-class classification of car/dashboard/documents
  • step2 : classification of 15 types of documents

 

# 06. Submissions

- Submission #1 & #2 & #3 & #4

(2024-04-16)

• Augmentation with augraphy in test stage
  • augraphy (docs only) - fold_aug5, delauny_aug5, noisylines_aug5 를 train에 추가

 

• Submission #1
  • (Baseline 0.1992) -> Public LB 0.6263 -> Private LB 0.6131
  • training config
    model_name = 'resnet101d'
    img_size = 128
    LR = 1e-3
    EPOCHS = 100
    BATCH_SIZE = 32
    num_workers = 0
• Submission #2
  • Public LB 0.6256 -> Private LB 0.6174
  • training config
    model_name = 'resnet101'
    img_size = 128
    LR = 1e-3
    EPOCHS = 50
    BATCH_SIZE = 32
    num_workers = 0
    
    

• Albumentation (updated 0416_4pm)
• Submission #3
  • Public LB 0.6556 -> Private LB 0.6969
  • training config
    model_name = 'resnet152'
    img_size = 384
    LR = 1e-3
    EPOCHS = 15
    BATCH_SIZE = 32
    num_workers = 0

• Submission #4
  • Public LB 0.6941 -> Private LB 0.7227
  • training config
    model_name = 'resnet152'
    img_size = 384
    LR = 1e-3
    EPOCHS = 10
    BATCH_SIZE = 32
    num_workers = 0
    
    

- Submission #5 & #6 & #7

• VoronoiTessellation(noise_type="random", p=0.5)
• Submission #5
  • Public LB 0.6941 -> Private LB 0.7227
  • training config
    model_name = 'resnet152'
    img_size = 384
    LR = 1e-3
    EPOCHS = 30
    BATCH_SIZE = 16

• Submission #6
  • Public LB 0.9218  -> Private LB 0.8985
  • training config
    model_name = 'resnet152'
    img_size = 384
    LR = 1e-3
    EPOCHS = 30
    BATCH_SIZE = 16
• Submission #7
  • Public LB 0.9230  -> Private LB 0.9156
  • training config
    model_name = 'efficientnet_b4'
    img_size = 384
    LR = 1e-3
    EPOCHS = 30
    BATCH_SIZE = 16

 

- Submission #8

• Edge dilated
• Submission #8
  • Public LB 0.9329 -> Private LB 0.9243

 

# 07. Team's Final Result

• Public Score : 3rd (0.9578)

• Private Score : 2nd (0.9441)