micronets 0.1.2

A simple autograd engine along with a PyTorch/Tensorflow like API module to easily train neural nets

Micronets

Micronets is a python framework for the simple implementation and training of deep neural networks. The project consists of a small auto-gradient engine module which implements back propagation through a DAG (Directed Acyclic Graph) and a module resembling the PyTorch/Tensorflow APIs to create and train dense neural networks on top of it.

The package also provides additional functionalities such as plotting of the training curve, make predictions and visualisation of computational graphs after training

Installation

pip install micronets

Usage

1. Building Computational Graphs

from micronets.autograd_engine import Value
from micronets.graph import draw

#inputs
x1 = Value(2.0,label='x1')
x2 = Value(4.0, label='x2')

#weights
w1 = Value(-3.0,label='w1')
w2 = Value(1.0,label='w2')

#bias
b = Value(3.3,label='b')

x1w1=x1*w1; x1w1.label='x1*w1'
x2w2=x2*w2; x2w2.label='x2*w2'

x1w1x2w2 = x1w1+x2w2; x1w1x2w2.label='x1w1 + x2w2'

z = x1w1x2w2 + b; z.label='z'

output = z.tanh(); output.label='output'

output.backward()

draw(output)

2. Neural Net Implementation

from micronets.nn import DNN  # Dense Neural Network
from micronets.losses import binary_cross_entropy


xs = [
    [2.0, 3.0, -1.0],
    [3.0, -1.0, 0.5],
    [-0.5, -1.0, -1.0],
    [-1.0, -1.0, 0.5]
]

ys = [1.0, 0.0, 1.0, 0.0]

model = DNN(input_features=3, layers=[2,2,1], activations=['relu', 'relu', 'sigmoid'])  # creates model architecture

history, graph = model.train(X=xs, Y=ys, iterations=500, loss_function=binary_cross_entropy, learning_rate=0.1)  # trains network and stores loss function history and computational graph

predictions = model.predict([  # outputs predictions
    [2.0, -1.0, 3.5],
    [-1.0, -3.0, 0.5]
]) 

model.plot(history)  # plots learning curve

Additionally the playground.ipynb file demonstrates functionality through a more complex example along with its outputs.

Lastly, a big thank you to Andrej Karpathy for inspiring this project :)

License

MIT