csxeba/csxnet

My Neural Networking library with evolution optimization and a data wrapper class

csxnet's Issues

Optimizers need some love

The module -> optimizers need some refactoring, or maybe a complete rethinking.
The accumulated gradients / memory matrices / momentum terms are not transferable at the moment.

Fix CNN

CNN is completely broken
Implement Convolution or use the fast(ish) SciPy convolution!
Figure out the math behind backpropagation in the ConvLayer!

Do a major refactor!

A major refactoring is underway...

Merge CData and RData

CData and RData should really be merged into a big data wrapper class which can organize categorical and continuos independent variables along with the actual dataset.

Problem / session wrapper

The datamodel and the fitted network should be wrapped in a class together, so multiple datasets and multiple models can be bundled together

Finish Sequence datamodel

Write a dataframe for sequence type data. Text should be a subclass of this.

Finish LSTM

Create LSTM layer in thNets!

Speed it up!

Convolution and MaxPooling is very slow. Find a way to speed things up!

Create more tests!

Need more tests especially in higher-level modules, like high_utils.
Figure out some way to write tests for neural layers... Maybe hack around the stochastic weight inicialization?

Connect with sklearn

Find a way to add an sklearn model as the last layer (or maybe a hidden layer???) of a network.
Write wrappers?

"Preacher" class

Or something to hold a predictor, a stripped-down version of a neural network, uncapable of learning, only used for prediction.

Finalize the Autoencoder

Implement encapsulate() for autoencoders.

Create Base Class for all Networks

Create an abstraction for every neural network class... It's getting quite chaotic now -> some methods are implememented in some, not in others