Author: László Szél
Date: 2021-02-01
This is a simple neural network for handwritten digit recognition, and it is implemented in pure SQL. The image transformation from MNIST sample data to database records are implemented in SQL too.
The trainig and the test samples are downloadable mnist training and test files from here: http://yann.lecun.com/exdb/mnist/
The files:
- http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz: training set images (9912422 bytes)
- http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz: training set labels (28881 bytes)
- http://yann.lecun.com/exdb/mnist/t10k-images-idx3-ubyte.gz: test set images (1648877 bytes)
- http://yann.lecun.com/exdb/mnist/t10k-labels-idx1-ubyte.gz: test set labels (4542 bytes)
Extract them to a folder, for example to /home/centos/mnist or you can copy them into the data directory of the mariadb directly.
You can check some enviroment settings with the sql script
This sql file loads the original mnist files into a table and split them to separate rows for faster use.
If the sample data is visible for the mariadb, than you can create the databas
- RectifiedLinearUnit
- Sigmoid
- Hiperbolictangent
- And the derivatives :)
These views are contains the base calculations for learning and backpropagation.
Main procedures and functions for learning and testing.
With the next sql you can set the size of layers, and it creates the neurons and the weigths.
If the network is ready, you can initiate the learning with a command:
call train(number of cycles, the batch size, learning rate);
for example: call train(20,100,0.01);