static MLP GUI v0.31

MLP-GUI is manually coded multilayer perceptron / neural network using only numpy. The approach is function based. For a dynamic class based MLP implementation, checkout my MML repo - an improved approach for deeper access to the model.

Mainly initiated for learning python(3) and the math behind some mechanics used in neural networks, the project grow to a useful tool for simple classification/regression tasks and getting an intuition about the effects of hyperparameter choice.

A simple GUI is available, so little to no programming knowledge is required to get your hands dirty with some neural networks.

how to use

Dependencies:

• time
• datetime
• matplotlib
• PIL
• numpy
• appJar (Tk wrapper, works if appJar folder is in the same directory. no install needed)

1. clone repo
2. open console: python mmlp.py

NOTE: works on windows, to use on ubuntu(etc.) toggle line comment for use of "mmlp.ico"

Features  

Features implemented so far:

• loading/saving data (data, weights, label)
• 1-, 2-, 3- hidden layer network
  • free choice for number of nodes in each layer
  • classification and regression
• full hyper-parameter control:
  • Error-/Lossfunciton
    quadratic
    pseudo huber loss
    binary cross entropy
  • Activation function (out layer activation) / Transfer function (activation for all units except the out layer)
    tanh
    atan
    logistic
    softplus
    relu
    bent identity
    binary
    stochastic
  • Gradient descent optimizer
    vanilla gradient descent
    Adam / RMSprop
    SGD (minibatch)
    learning rate
  • Regularization
    Droptout
    L2/L1 norm
• PCA
  • to reduces the input dimensions
• Autoencoder
  • saves autoencoded data and weights

Visualized training process

!video->Video