• Link: https://www.kaggle.com/c/cassava-leaf-disease-classification

Cassava Leaf Diease Classification

Objective : Classify each cassava image into four disease categories or a fifth category indicating a healthy leaf.

Solution

• BaseLine Code : I used my tensorflow tutorial code as a baseline code for this contest. You can see my code here Tensorflow-Tutorial

• Preprocessing

  • Image Size : 512
  • Augmentation : RandomResizedCrop, Transpose, HorizontalFlip, VerticalFlip, RandomRotate90, RandomBrightnessContrast, HueSaturationValue

• Loss

  • SparseCategoricalCrossEntropy

• Model

  • EfficientNetB3 pretrained with Imagenet with no top.
  • EfficientNetB3 top layer : GlobalAveragePooling + Dropout + Dense

• Optimizers & LR Scheduler

  • Adam + Lookahead + CosineDecay
  • Adam + SWA(Stochastic Weight Averaging) + CosineDecay
  • RectifiedAdam + Lookahead + CosineDecay
  • Adabelief + Lookahead + CosineDecay

• Training

  • Batch Size: 8
  • Epochs :10
  • Callbacks : EarlyStopping, ModelCheckpoint
  • I've also tried training my model twice by first training the newly added top layer, and then unfreeze them and fine-tuning the entire model. But it didn't show much improvements.

• Inference

  • Ensemble & TTA(Test Time Augmentation)

Models	TTA	Public Score	Private Score
Adam	6	0.8600	0.8626
Adam + SWA	6	0.8548	0.8581
Adam(5Fold)	6	0.8607	0.8652
RectifiedAdam	6	0.8571	0.8627
Adabelief	6	0.8620	0.8638
Adam, RectifiedAdam	6	0.8609	0.8661
Adam, RectifiedAdam, Adabelief	6	0.8600	0.8685

Improvements & Results

• Improvements
  • Ensemble is powerful! I used single model for the final submission because the public score was higher, but in private score ensemble models were higher.
  • I wanted to use CutMix and Mixup with label smoothing but OOM kept occur during training. Should study more about this.
  • Also I wanted to use ResNeXt50 model, but I couldn't find how to use it from keras.
  • Train dataset for this competition had few miss labeled data, and after the competition I've found out that I could have tried different loss, like bi-tempered loss. Also have to study more about this.
• Results
  • Private: 0.8652
  • Public: 0.8607
  • This was my first competition for image classification and I've learned many things during competition. Thanks for all kagglers and organizers for this competition.