wernhervondata/ai-katie

The library that is meant to be my own Python package that provides utility tools for Data Science and AI using common libraries. It's used and extended by myself when I'm running to problem that might be more common than me.

ai-katie's Issues

Implement save and load of the replay memory

To provide more flexible ability to resume learning of models implement the save and load to the pickle file. Keep in mind that the loading should contain hash checker as loading not-known pickle may be dangerous: https://docs.python.org/3/library/pickle.html#module-pickle

Dataframe normalizer

Create a function that will allow to use a normalizer globally for columns

def apply_normalization(dataframe, columns=None)

Check performance of the ReplayMemory

Check how the current deque behaves under a big number of elements in the memory in hundreds od thousands and millions. Verify especially loading and adding new elements when the memory is full.

Implement eventual improvements

N step data gatherer

Create a way to gather play data in batches.

NeuralNetwork RandomSearch

Allow to wrap a network from #1 in a Random Search, that will allow you to find the best architecture to tackle the problem.

Multi-input model definition

Provide a model structure that was used in the Google Q&A Labelling challange.

NeuralNetwork GridSearch wrapper

Allow to wrap a network from #1 in a Grid Search, that will allow to find the best architecture to tackle the problem.

Configurable Keras Network

The goal of this task is to provide an API that will allow creating a custom neural network on Keras backend.

In scope:

• Object Containing fit and predict methods
• Means to create custom network size - one input - one output at this moment.

Implement Eligibility Trace as a general purpose function

Use the eligibility trace function from ai-doom, but make it input type agnostic

Implement utils from my ai-doom

Get everything that is useful from the ai-doom repository and implement it, so it can be used from here.

Also publish it on the test-pypi and when it will be satisfying - make a first publish on the real PyPi.