Final Term — POC Dataset Classification using GoogLeNet

Development Process

---

1. Environment Setup

• Configured a PyTorch-based deep learning environment
• Implemented data loading using torchvision.datasets.ImageFolder
• Added augmentation and normalization pipelines
• Organized the project structure (googlenet.py, train.py, result/)

---

2. Implementation of GoogLeNet Architecture

• Reconstructed the Inception modules following the original paper
• Implemented Auxiliary Classifiers to stabilize gradient flow
• Built the complete GoogLeNet architecture from scratch
• Ensured auxiliary outputs are used only during training, not inference

---

3. Dataset Preparation

The POC dataset contains two folders:

POC_Dataset/
  ├── Training/
  └── Testing/

• Training → split into Train : Validation = 90 : 10
• Train set → augmentation applied
• Validation / Test sets → only resize + normalize
• Testing folder was strictly used only at the very end → prevents data leakage and ensures proper generalization evaluation

---

4. Baseline Training

The initial baseline training used a minimal pipeline:

• No augmentation
• No LR scheduler
• Direct train/test split

Baseline performance: ~50–56% accuracy

Issues detected:

• Unstable training
• High confusion in certain classes
• Sensitivity to class imbalance

---

5. Data Augmentation & Normalization

To improve generalization, the following augmentations were added:

• RandomHorizontalFlip
• RandomRotation
• ColorJitter
• Input normalization (mean=0.5, std=0.5)

Result: Validation accuracy increased significantly — reaching ~71%, with more stable loss curves. This showed augmentation was essential for this medical dataset.

---

6. Training Stabilization

To build a more reliable training procedure:

• Added auxiliary classifier loss (GoogLeNet aux branches)
• Introduced StepLR scheduler
• Implemented Early Stopping (patience = 5)
• Created a full train/val/test split
• Added automatic logging (CSV)
• Enabled intermediate Confusion Matrix visualization

Effect:

• Training stabilized
• Overfitting became easier to detect
• Best-performing model was saved automatically

---

7. Evaluation & Error Analysis

Confusion matrices were generated for both validation and final test sets.

Key observations

• Chorionic_villi ↔ Trophoblastic_tissue showed noticeable misclassification
• Hemorrhage was classified relatively accurately
• Visualization clearly revealed class imbalance and inter-class similarity issues

These insights guided tuning decisions throughout development.

---

8. Final Results

• Best Validation Accuracy: 87.23%
• Final Test Accuracy: 81.34%

• Exported results include:

  • Per-epoch validation confusion matrices
  • Final test confusion matrix
  • Training log CSV
  • Best model checkpoint

Despite limited dataset size, the model shows strong improvement compared to the initial baseline.

** Result Summary **

Final Test Performance Summary

Metric	Score
Accuracy	0.8134
Macro Precision	0.8214
Macro Recall	0.8045
Macro F1-score	0.8024
Weighted F1-score	0.8025

---

Per-Class Precision, Recall, F1-score

Class	Precision	Recall	F1-score
Chorionic_villi	0.8048	0.9359	0.8655
Decidual_tissue	0.8044	0.5186	0.6317
Hemorrhage	0.7619	0.9145	0.8317
Trophoblastic_tissue	0.9147	0.8489	0.8807

---

Performance Interpretation

• Chorionic_villi and Trophoblastic_tissue achieved strong performance, each with high F1-scores (0.86–0.88).
• Hemorrhage was also classified accurately (F1 ≈ 0.83).
• Decidual_tissue exhibited noticeable misclassification (F1 ≈ 0.63), consistent with the Confusion Matrix.
• The overall metrics (~0.80 macro/weighted F1) indicate solid generalization despite the dataset’s small size and inter-class similarity.
• Auxiliary classifiers and data augmentation contributed significantly to training stability and performance.

---

Confusion Matrix (Final Test)

9. Training Pipeline

• Loss: CrossEntropyLoss + weighted auxiliary losses

• Optimizer: Adam (lr = 1e-3)

• Scheduler: StepLR(step_size=7, gamma=0.1)

• Added tqdm progress bars

• Enabled Early Stopping (patience = 5)

• Saved best model to:

  result/googlenet_poc_best.pt
  

---

10. Metrics & Visualization

• Logged training loss and validation accuracy per epoch

• Saved validation confusion matrices:

  result/cm_val_epoch_XX.png
  

• Saved final test confusion matrix:

  result/cm_test_final.png
  

• Exported training log:

  result/training_log.csv
  

---

11. Final Test (Hold-out Evaluation)

After training, the reserved Testing dataset was used for unbiased evaluation:

• Final Test Accuracy was computed
• Final Confusion Matrix generated
• Confirms generalization performance on unseen data

---

Final Results Summary

Metrics

• Best Validation Accuracy: 87.23%
• Final Test Accuracy: 81.34%

Confusion Matrix Insights

• Some confusion between Chorionic_villi and Trophoblastic_tissue
• Hemorrhage was reliably classified
• Auxiliary classifiers helped stabilize training on a small dataset

---

Project Structure

ComputerVision/
 ├── googlenet.py
 ├── train.py
 └── result/
      ├── cm_val_epoch_01.png
      ├── cm_val_epoch_02.png
      ├── cm_test_final.png
      ├── training_log.csv
      └── googlenet_poc_best.pt

---

Summary

This project implements a complete training pipeline for classifying medical images using a reconstructed GoogLeNet architecture.

It includes:

• A clean train/validation/test workflow
• Auxiliary classifier integration
• Training stabilization techniques (scheduler, early stopping)
• Automated logging and visual analysis
• Rigorous evaluation on a dedicated hold-out test set

This work demonstrates practical deep learning engineering skills suitable for academic submissions or portfolio use.

[MIDTERM]K-Nearest Neighbors (KNN) on CIFAR-10 — Assignment Version

이 프로젝트는 CIFAR-10 이미지 데이터셋을 이용하여 K-Nearest Neighbors (KNN) 분류기를 구현하고, 세 가지 실험 모드에 따라 모델 성능을 평가하는 과제용 코드이다.

주요 기능

기능	설명
KNN 분류기 구현	scikit-learn의 KNeighborsClassifier 사용
데이터셋 로드	torchvision.datasets.CIFAR10로 자동 다운로드 및 변환
데이터 전처리	StandardScaler로 픽셀 단위 정규화 (거리 기반 성능 향상)
실험 모드 3종	train/test, train/validation/test, 5-fold cross-validation
평가지표	Accuracy, Precision, Recall, F1-score (macro 평균)
그래프 저장	k 값에 따른 정확도 변화를 시각화 (matplotlib)

파일 구성

knn_cifar10_assignment.py : 메인 코드 (모든 기능 포함)
plot_split_k.png : train/test 결과 그래프 (자동 생성)
plot_val_k.png : validation 결과 그래프 (자동 생성)
plot_cv_k.png : 5-fold cross-validation 결과 그래프 (자동 생성)

의존성

pip install torch torchvision scikit-learn matplotlib numpy

실행 방법

1. 단순 train/test split: CIFAR-10 데이터셋을 단순히 학습/테스트로 나누어 평가

python knn_cifar10_assignment.py --mode split --k_list 5
--train_size 10000 --test_size 5000 사용 데이터: train 10,000 / test 5,000

결과 그래프: plot_split_k.png

2. train / validation / test split: Validation 세트를 사용하여 최적의 k를 선택한 뒤, Test 세트에서 최종 성능 평가

python knn_cifar10_assignment.py --mode split_val --k_list 1 3 5 7 9
--train_size 10000 --val_size 5000 --test_size 5000

Validation set으로 best-k 선택

Test set에서 해당 k로 최종 평가

결과 그래프: plot_val_k.png

3. 5-fold cross-validation: StratifiedKFold로 각 fold에서 KNN 학습 후 평균/표준편차 계산

python knn_cifar10_assignment.py --mode cv --k_list 1 3 5 7 9 --folds 5

폴드마다 독립적인 전처리 및 평가 수행

k별 평균 정확도 ± 표준편차 계산

결과 그래프: plot_cv_k.png

생성되는 그래프 요약

모드	파일 이름	내용
split	plot_split_k.png	Test Accuracy vs k
split_val	plot_val_k.png	Validation Accuracy vs k
cv	plot_cv_k.png	5-Fold Mean Accuracy ± Std vs k