chr0nikler / kuzushijirecognizer Goto Github PK

• Do some exploratory analysis, so you can view any image and it's bounding boxes
• Setup fastai datablock api to read train/validation sets.
• Setup how to transform output of CNN into output for submission
• See how a simple RNN performs. Don't need online learning, so accuracy over speed.
• See where I am, rinse and repeat? Study more FastAI courses, see if I can improve this score
• Build an initial character recognition network from scratch, on subset
• Examine hooks network from scratch, see what is was looking for?
• Try to build a character recognition network for all chars
• Build object recognition from subset
• With small object recognition model, classified boxes, then feed them into small character recognition model
• Repeat for large model

The strikethroughs are inteded to show what I had PLANNED on doing, but as my learning evolved, so did my understanding of what I wanted to do next

exploratory_analysis takes a peek at the initial data, and renders a sample

prepare_character_set contains code for creating the dataset and exploring a 2 class cnn.

character_recognizer has a 10 class CNN

character_recognzier_v2 has the full CNN for all Kuzushiji characters

object_recognizer_v1 contains initial attempts to do full character object detection on a page