This is a simple MNIST demonstration using a neural network that is loosely based on Naimishnet CNN.
The results were very good, but this is mostly used as a Hello World
Python 3.x numpy pytorch with cuda92
There are a few basic networks in model.py
, but Net2 achieved the best results.
When run using 60000 training images, a batch size of 64, and 1000 epochs.
Training did not seem to overfit based on the loss over time, but the extent of was not fully explored.
![Loss]((Net2_1000_loss.png)
The accuracy is very high, but there are still a number of images that were predicted incorrectly