My solution for the following Kaggle Competition:

Cassava Leaf Disease Classification https://www.kaggle.com/c/cassava-leaf-disease-classification

kaggle competitions download -c cassava-leaf-disease-classification

I also used the images from the 2019 competition as additional training data:

kaggle competitions download -c cassava-disease

After extensive cross-validation settled on:

• Efficientnet B4
• Larger image size (446x446)
• Standard augmentations + cutmix (cutmix made a big difference)
• Self-distillation on out-of-fold predicted soft targets
• Cosine annealing learning rate schedule, model converges after 10 epochs

Two-step training gcp_446.py and gcp_distillation_final10.py for the final distillated model.

In notebook leaf_inference.ipynb, usign light test-time augmentation.