Skip to content

Latest commit

 

History

History
17 lines (10 loc) · 1.2 KB

README.md

File metadata and controls

17 lines (10 loc) · 1.2 KB

UjipenChars2 handwritten letters classifier with Gated Recurrent Unit (GRU)

This repository supplements stm32f429-chars repository to train a recurrent neural network that will be used later on in a microcontroller.

A small list of manually picked examples from train data which confuse classifiers is put in dropped.txt. All test samples from UjipenChars2 dataset are used during the model validation.

The main file is gru.py, where the training procedure of GRU is defined alongside with the test (validation) score.

Initially started with DTW as a baseline algorithm to find the closest pattern from the train data, given an input sample. DTW-related implementation is moved to dtw branch.

To give you the rough approximation of performance of both classifiers,

GRU DTW
Validation accuracy 98.3 % 81.9 %

But the main difference between those two is their inference time: GRU is much faster than DTW due to parallel computation.