QRNet

A simple neural network that predicts the first character that is stored in a QR-code such as this:

At training time, QR codes are created randomly from alphanumeric strings of length ten (00bLF6qvQF in the example above). The validation set also contains 5000 randomly generated QR codes [1].

Training

To start the training, simply call:

python train.py

Then, you can run tensorboard to view the test-set accuracy:

tensorboard --logdir /tmp/qrnet-log --reload_interval 5

Results

After approximately 50.000 iterations (with 200 QR codes per batch), it reaches a test-set accuracy of over 0.999.

Footnotes

[1] Note that there are (2 × 26 + 10)^10 ≈ 10^18 possible QR codes, so chances of collisions between the training and the test set are vanishingly small.

[2] Unfortunately, the online training is rather slow. Most of the time is actually spent in the (parallelized) QR code generation.