Alvin Rachmat - Manually Tuned DenseNet201 Binary Eye Gender Classification With Image Augmentation

Note for myself next time : Try grayscale as rgb input but process it first through loop to create folder of files so modelling wont take more time (as it increase ~50% time). Try Kfold. Try featurewise and samplewise. Optimize width-height shift and zoom range, maybe using automated GridSearch-like. Add wget upload saved model from drive, add wget to export saved model to drive to speed up upload process instead upload from local. Learning : Tried each architecture and manual one till come to the conclusion DenseNet-201 good enough for this. Looks like many nodes doesnt mean higher accuracy (tried 7680 vs 1536 vs 512 in DenseNet-201). BatchNormalization speed up convergence process. Using class weight is good this time. Mengorbankan satu hal untuk hal lebih penting adalah bijak. Kamu melakukan hal yang benar dan jangan disesali. 12 Ags 21

8/12/2021
382 views

Tags:  

#deep-learning