Welcome to BlueBird’s documentation!

logo

BlueBird is a small deep learning library. It was developed to help me (the author) learn better about deep learning.

I thought that it was a cool project so I decided to share it with others.

Down bellow you will find installation instruction and a basic tutorial.

Note: This documentation was generated automatically so there may be some bugs and it’s not easy to navigate.

Install Guide

BlueBird can be downloaded with pip (make sure you set up virtual environment). If you want to use the latest version, get it from github.

pip install bluebird-stoick01

Quick Start

Simple module looks like this:

from bluebird.nn import NeuralNet
from bluebird.activations import Relu
from bluebird.layers import Input, Linear

# create the module
net = NeuralNet([
   Input(10), # input layer
   Linear(50),  # hidden layer
   Relu(), # activation function
   Linear(10) # output layer
])

# build the module
net.build()

# train your model
net.fit(X_train, y_train, num_epochs=20)
Let’s explain this a bit:

  1. Imports: From nn you need to import your model. We also import activation functions (Relu in this example) and layers (Input and Linear)

  2. Then we create the model. The model accepts a list of layers where the first layer has to be of the input type. After that you can put layers in any order. You have to define how many neurons each layer has, except the input, where you define number of inputs.

  3. After you have created your model, you can build, during this step you can define optimizer and loss functions, which you can find in bluebird.optimizer and bluebird.loss

  4. Final step is to train your model. Where you pass it your data and define number of epochs.

Layers and Activations

You can find all the layers and activations in bluebird.layers and bluebird.activations.

Every layer and activation is explained in the mode detail on it’s own page, but in general:

  • All models begin with input type layer

  • All other layers must have defined number of neurons, aka the output dimension, (input in handled during build faze of the module)

  • Activations only need to be initialized

Building the model

After creation each model must be built. This is a crucial step, where all the weights in layers are initialized and each layer is built.

You can also pass the loss function and optimizer which can be found in bluebird.loss and bluebird.optimizer

Contribution Guide

I don’t expect anyone to contribute but just in case you want to, contact me, or open an issue on github before committing anything.

bluebird

Indices and tables