bhamz-1 / fgvc6-kaggle-cassava-classification

Cassava disease classification using architectures such as ResNet50, MobileNet, VGG16, and SE-ResNeXt.

Cassava disease classification using:

1. ResNet50V1
2. MobileNetV2
3. VGG16
4. SE-ResNeXt101-32x4d

Data augmentation:

1. Zoom (0.1)
2. Horizontal Flip
3. Vertical Flip

These augmentations were chosen because for the purposes of the images, these augmentations do not interfere with the integrety of the image. For example, a plant flipped horizontally can still be seen as a plant and therefore, disease feature extraction would not be affected.

ResNet50V1 (Val_acc = 0.95, 10 epochs)

https://arxiv.org/abs/1512.03385

Trained using 336x336 image size, 3x augmentation. Parameters: 25,636,712 (25,583,592 trainable)

Loss and accuracy curves

MobileNetV2 (Val_acc = 0.93, 10 epochs)

https://arxiv.org/pdf/1801.04381.pdf

Trained using 336x336 image size, 3x augmentation. Parameters: 2,257,984 (2,223,872 trainable)

Loss and accuracy curves

VGG16 (Val_acc = 0.94, 10 epochs)

https://arxiv.org/abs/1409.1556

Trained using 336x336 image size, 3x augmentation. Parameters: 14,717,253 (14,715,461 trainable)

Loss and accuracy curves

SE-ResNeXt101-32x4d (Val_acc = 0.85, 10 epochs)

https://arxiv.org/abs/1611.05431

Trained using 224x224 image size, 3x augmentation. Parameters: 47,054,517 (46,916,661 trainable)

Loss and accuracy curves

Model Improvements

There are many improvements that can be made to the model but were not implemented due to time and computation purposes:

1. Increase image size For reference, each epoch takes:

Image Size: 500x500 - Colab GPU: 15 minutes, i5-8300H CPU: 9 hours

Image Size: 336x336 - Colab GPU: 8 minutes, i5-8300H CPU: 5 hours

Image Size: 224x224 - Colab GPU: 5 minutes, i5-8300H CPU: 1 hour 30 minutes

2. Potentially add data augmentation
3. K-fold cross validation
4. Test time augmentation
5. Callbacks (though may not be useful due to the low training epochs) ReduceLROnPlateau EarlyStopping
6. Add additional Dense layers to the base model output including DropOut.
7. Potentially remove the mean from training data.

Appendix

Sample training (top to bottom - ResNet50, MobileNet, VGG16, SE-ResNeXt)

ResNet

MobileNet

VGG16

SE-ResNeXt