CNN and MNIST using Go
This repo sets up some Go code to read in MNIST images leveraging the GoMNIST repo. It loads in all the images and displays some examples in grayscale. See example image below:
After reading in the images, the next step was to build a model to make image predictions. Here are the steps for this project:
- Split the MNIST images into test and training
- Find a Go Package to build the neural network go-deep
a. This is not a CNN
b. Will start with a simple NN and can build after getting the code to work - Set up the params for the neural network and create functions for training
a. Normalize the images
b. Convert labels and build examples
c. One hot encoding - Run the code - go run main.go or double click the app - mnistNN.app
The model did not perform as well as I would have liked - coming in around 40 to 45%. I believe that with some more work on the set up and the helper functions, I could get closer to scores above 85%. Part of the results should also be compared to the isolation forest to see how anomaly detection was affecting my code.
To continue to build on this neural network one would need a new package, perhaps Gorgonia. I attempted to use this on the MNIST images, but was unable to load the data even following this example. The example does a great job of showing how to build a simple CNN, but it did not integrate well with the previous work around MNIST for isolation forests.