"NodeNet" is a project aims to bulid neural network. It's oriention is to collect plenty of types of neural networks and provide manager to organize and test it, and the neural network libary can work without needs of manager isolatedly. It is a project belongs to NOOXY. There still lots of miles to complete it. Article about it might be established some day. Visit us https://nooxy.org.
Testers part provides fully connected and convolutional neural net testing. Besides, it also test NodeNet’s other functions like graphing.
Variables part provides default parameters for each gradient optimizer, backward configuration, forward configuration, default files path and framework version etc.
Utilities part provides commons utilities that can be used widely, such as shuffling datasets, getting mini batch randomly or generating dataset for testing like sin function or MNIST handwriting datasets. Even data preprocessing abilities will be added in future.
We only provide mini batch training session now. You can adjust lots of parameters like batch size, target loss, training profile and dropout keep rate etc.
It provides container to contain nodelayers, netlayers, convolutional layers and others. Which constructs neural net.
It provides pre design neural net with already determited activation functions, layers.
It provides different backpropagation optimization, such as momentum, nesterov momentum, adagrad, adadelta, rmsprop and adam.
IO provides access between neuralnet file and neuralnet objects. And it also gives access of datasets.
It provides nodes, convolutional, fully connected nets, pooling layers and convertor layers.
We provide graphing and console interface to graph training loss, graph neural net output and colored console log and provide console input interface.
Functions part provides vary types of activation functions like RELU, softmax, sigmoid. And loss functions like MSE, cross entropy error.
This part we haven’t finished it yet.
Providing friendly console interface to access NodeNet’s neuralnet. Which can build neuralnet basiclly by command.