DeTeXt, a year later
6 November 2021
A year and 2 months ago, I released DeTeXt – my first iOS app that I had built using SwiftUI, Combine, CoreML and PencilKit. Thinking back, I believe I went from no-code to an app in the App Store in 6 weeks? I can hardly believe it now, but I am proud of building DeTeXt – I’m glad it helped some people find the LaTeX symbol they wanted and saved perhaps a few seconds of their day. Building and releasing DeTeXt emphasized one aspect of myself – I love building tools and working on projects that help other people be successful at their work, even in very small ways.
Initially, I had trained the MobileNetV2 neural network for DeTeXt’s symbol recognition engine on a 4 GPU cluster. Today, I did the same on my M1 Pro Macbook Pro running on battery, thanks to tensorflow-metal and the 16 core GPU. I wish I had documented the training better on the cluster, but I do have the training time for the M1 Pro: 13 minutes per epoch, for training a MobileNetV3_small model on over 200,000 images. In 2 epochs, the network reached an accuracy of 68%. Half an hour to train a complex neural network with over 2 million parameters, on battery power. Wow.