Kaggle's Kuzushiji Recognition
- Mobilenet backend => small, super fast (also, works on mobiles)
- U-net => pix2pix upsampling with mobilenet backend
- Centernet with max-pooling vs NMS (speed vs accuracy, more training, no bounding boxes, e2e trainable)
- Pretrain backbone for classification
- Thresholded images => (6% improvement)
- Larger images (256 vs 512)
- Patches of original images for training
- Imprinting to alleviate generalization loss
- Cyclic learning rates
- Display Callbacks in Keras
- F1 score => box recall vs character recall
- Monitoring
- Training time => 2 hours