Playing around with toy data to understand how tensorflow works
Code was written using Jupyter and Tensorflow 1.1.0-rc0 using the official docker image
This project is simply a collection of things I am working through to get a feel for how to work with Tensorflow.
- Simple linear regression, nothing fancy.
- Set up so that all controls are at the top.
- Implemented regularization
- Generic method of dealing with powers, so additional powers only requrire the change of 1 numbner
- Fully connected neural network
- Set up so that there can be any number of hidden layers of any size
- Utility functions in getMNIST.py to download and extract the images. The original code came from the tf tutorials
- Saving and loading model
- Confusion plots for training and test data
- Fully convolutional neural network
- Is intended to be generic to the number of and shapes of layers, but it really finicky
- Experimented with tensorboard. Started later in the project as a debugging activity so only prints out the net. The way it is done makes the graph really messy. Need to look into how to control the outputs better.
- I found it really interesting how small changes in the initialization caused major issues in training. I knew that it was an issue, but it was tough to track down.