leeloolee / Kaggle-Bengali.AI-Handwritten-Grapheme-Classification

Kaggle-Bengali.AI-Handwritten-Grapheme-Classification 167th Place Solution

Dependencies

pytorch == 1.4.0
cudatoolkit == 10.1
efficientnet
numpy == 1.18.1
pandas == 1.0.3
scikit-learn == 0.22.1
apex == 0.1
tqdm == 4.44.1

Solution

https://www.kaggle.com/c/bengaliai-cv19/discussion/144549

Dataset

All the directories should be created manually before running the codes as mentioned in Directory_structure.txt file. All the folders should have both read and write access. After creating the directories with suitable permissions, competition data(only the train parquet files) must be placed in the '/data/' directory
• Image Size: 137x236 ( No preprocessing )

Augmentation

• CutMix

Model

• EfficientNet-B5 with three heads

Training:

• 5 fold Configuration
• Data split on the basis of grapheme root labels
• Loss: Cross Entropy Loss
• Optimizer: Over9000
• Scheduler: Reduce On Plateau
• Gradient Accumulation
• Batch Size 100
• Initial Learning Rate 0.03

Inference:

• Best Average recall checkpoints were used
• Simple Average of the outputs from 5 folds
• Inference kernel: https://www.kaggle.com/mohammadzunaed/efficientnet-b5-inference-kernel-pytorch?scriptVersionId=32245517

Things that did not work for me:

• Preprocessing
• GridMask, Cutout, AugMix
• Label Smoothing Criterions
• Single head instead of three heads
• Activation functions and Convolutional layers in the heads