Disclaimer

This project only contains simple implementations for very limited functionality and should not really be used for anything.

Installation

To install all necessary dependencies to run the example.py file

pip install -r requirements.txt

To install numpynet using pip

pip install .

Example usage

A very simple mnist example can be found in example.py, but to summarize the VERY limited current functionality

import numpy as np
from numpynet import Model
from numpynet import Linear
from numpynet import ReLU
from numpynet import MSE

# The model object takes a loss function object, so we create a MSE loss object
# to pass to the model constructor
loss_function = MSE()
model = Model(loss=loss_function)

# Add some linear layers and ReLU activation functions.
# Activation functions are added in the same way as layers
model.add_layer(Linear(4, 10))
model.add_layer(ReLU())
model.add_layer(Linear(10, 2)) # Linear activations for the output layer

X = np.random.uniform(-1, 1, (1, 4))
t = np.random.uniform(-1, 1, (1, 2))

# Forward pass
y = model(X)
# Compute loss
loss = model.loss(y, t)
# Backward pass
model.backward()
# Update parameters
model.update_step(lr=0.01)
# Set all grads to zero
model.zero_grad()