ar8372 / CutMix-Regularization-Strategy-to-Train-Strong-Classifiers-with-Localizable-Features

Through this project we will try to understand CutMix by implementing it on a simple problem of cat-vs-dog classification.

CutMix-Regularization

Through this project we will try to understand CutMix Augmentation by implementing it on a simple problem of cat-vs-dog classification.

Implementation

I have used keras to implement the cutmix augmentation. Dataset used: https://www.kaggle.com/competitions/dogs-vs-cats/data.
In CutMix we mix two images and their labels respectively in order to generate new data point. To implement this, I have created a custom datagenerator. This custom datagenerator takes batches from two seperate keras datagenerator and returns one batch by applying cutmix on it.

train datagenerator (custom)

train_dataset = CutMixImageDataGenerator(
    generator1= train_datagen1, #train_generator1,
    generator2= train_datagen2, #train_generator2,
    img_size=128,
    batch_size=32,
)

valid datagenerator (default)

valid_dataset = valid_datagen.flow_from_dataframe(
    dataframe=validate_df,
    directory=image_path,
    target_size=(128, 128), 
    x_col='filename', 
    y_col='category',
    batch_size=32,
    seed=42,
    shuffle=True,
    class_mode="categorical",  
    )

Results

Type	Without CutMix	With CutMix
Output
Training and Validation Accuracy plot
Score	loss: 0.7671 - acc: 0.6406 val_loss: 0.9630 - val_acc: 0.6445	loss: 0.1226 - acc: 0.9538 val_loss: 0.4688 - val_acc: 0.8438
#epochs	50	50
Early Stop at	28	15

Conclusion

We have two notebooks 1. Without CutMix and 2. With CutMix. Both notebooks have everything exactly same(seed, batchsize, #epochs, rescaling etc) only except the training datagenerator. And we notice that using CutMix augmentation has really improved the validation score making the model more regularized and robust in comparision to model which is trained without using CutMix augmentation.

Reference

• original paper: https://arxiv.org/abs/1905.04899