Attention-based Neural Networks for Handwriting Recognition

Documentation

Recording of working system: https://smith.zoom.us/rec/share/841_Ne3snhwP3mSduKZu63ctTFzYvdDdCrwsdPvCQWOAFDxka9tsdDTwGGZM3fWw.n5T9sxD24vdlCBzQ Passcode: iZQc4=5s

Fall 2020

This honors thesis aims to improve current handwriting recognition through refining the use of attention mechanisms in sequence-to-sequence and Transformer models in HTR systems.

Week 1: 09/04 - 09/10

Goals:

- Install PyTorch and get something running

- Find a good starting point
  - Review more literature (particularly for sequence-to-sequence models)
  - Distill knowledge from currently cited papers in proposal

Summer 2021 Notes

Pretraining run notes:

model_2021-06-18_11_53_47 is first successful run, using ReLU on fully connected layers. 4x image subsampling on 32x64 basic unit means that letters don't fill much of the space.

model_2021-06-22_16_58_03 implements random pixel sampling to handle images that don't fit in memory. 2x image subsampling makes letters bigger, but decreases random sampling fraction. [SVN 2293]

model_2021-06-24_17_28_20 standard size run; added character frequency weighting.

6/24 Ford342-05: standard res 32x64 unit, x2 subsampling, no character weights

6/24 Ford342-06: lo-res 16x32 unit, x4 subsampling, no character weights

6/25 Ford 354: standard res 32x64 unit, x2 subsampling, linear character weights, argmax pixels