These notebooks detail the method used to pre-process data and construct a convolutional neural network (CNN) whose aim is to identify house numbers from photos. Those photos are made available in the SVHM dataset, which is from google streetview.
The final model is 96% accurate, which is surprisingly good considering the 'entry level' PC and free software used. The hardware is a Windows 10 PC 64bit with Core i5-7500 (3.4GHz) 16Gb RAM, NVIDIA GeForce GTX1050Ti (4GB). The 1050Ti retails for GBP 135.00. Training in 20 epochs took 8hrs.
The work is part of the capstone project for the Udacity 730 Tensorflow course. It follows the approach of the google research team in the paper Goodfellow, et al 2013.
Thanks to Tensorflow.rstudio.com, the project was completed in R, not Python. This may be the first attempt to complete Udacity 730's capstone in R.