🐨코알라의 Kaggle 도전일지🦜

Dogs vs Cats

---

#01. Keras CNN Classification

Ver.01 (0.9171)

• val_loss: 0.2044 , val_accuracy: 0.9171 @epoch 26, epochs : 36/50
• Keras CNN Classification
• Run on Colab
• earlystop = EarlyStopping(monitor='val_loss', patience=10)
• learning_rate_reduction = ReduceLROnPlateau (monitor='val_accuracy', patience=2, verbose=1, factor=0.5, min_lr=0.00001 )

Ver.02 (0.7952)

• val_loss: 0.5258 , val_accuracy: 0.7952 @epoch 4, epochs : 6/50

• Keras CNN Classification

• Run on Colab

• 학습조기종료, 학습률 자동감소 조건 변경

• earlystop = EarlyStopping(monitor='val_accuracy', patience=2)

• learning_rate_reduction = ReduceLROnPlateau (monitor='val_loss', patience=10, verbose=1, factor=0.5, min_lr=0.00001 )

Ver.03 (0.9143)

• val_loss: 0.2238 , val_accuracy: 0.9143 @epoch 32, epochs : 38/50

• Keras CNN Classification

• Run on Kaggle ( GPU )

• earlystop = EarlyStopping(monitor='val_loss', patience=10)

• learning_rate_reduction = ReduceLROnPlateau (monitor='val_accuracy', patience=2, verbose=1, factor=0.5, min_lr=0.00001 )

Ver.04 (0.9189)

• val_loss: 0.2465, val_accuracy: 0.9189 @epoch 37, epochs : 37/50
• Keras CNN Classification
• Run on Kaggle ( GPU )
• batch_size를 5에서 128로 변경
• earlystop = EarlyStopping(monitor='val_loss', patience=10)
• learning_rate_reduction = ReduceLROnPlateau (monitor='val_accuracy', patience=2, verbose=1, factor=0.5, min_lr=0.00001 )

Ver.05 (0.)

• val_loss: , val_accuracy: 0. @epoch , epochs : /50
• Keras CNN Classification
• Run on Colab
• batch_size를 5에서 128로 변경
• 학습조기종료, 학습률 자동감소 조건 변경
• (EarlyStopping과 ReduceLROnPlateau의 모니터 모두 val_accuracy)
• earlystop = EarlyStopping(monitor='val_accuracy', patience=10)
• learning_rate_reduction = ReduceLROnPlateau (monitor='val_accuracy', patience=5, verbose=1, factor=0.5, min_lr=0.00001 )

---

#02. Keras CNN, Transfer Learnings

Ver.01 (0.9468)

• val_accuracy: 0.9468
• Run on Kaggle ( GPU )
• Feature Extractor : VGG16 + LogisticRegression
• Fine Tuning : None
• cv_scores=cross_val_score(LogisticRegression(solver="lbfgs"), features, targets, cv=3 )

---

#03. Feature extractor fine tuning with Keras

Ver.01 (0.9844)

• val_loss: 0.0345 , val_accuracy: 0.9844 @epoch 5, epochs : 5
• Run on Kaggle ( GPU )
• MGD(Mini-Batch Gradient Descent)
• Feature Extractor : VGG16 + Logistic Regression
• Fine Tuning : ResNet50
• compare "Custom CNN / Transfer Leearn(VGG16) / Fine Tune(ResNet)

---

#04. Vanila CNN & ResNet50

Ver.01 (0.9892)

• val_loss: 0.0373, val_accuracy: 0.9892 @epoch 8, epochs : 13/15
• Run on Kaggle ( GPU )
• Building model for transfer learning on top of pretrained ResNet50 Model