neurgoo
Implementation of modular neural networks
Components
As of the version `0.1.0 has 3 main components to train any neural network for downstream tasks:
I) models
Models are the skeleton of neurgoo where we can add any arbitrary number of layers for training.
All the models are inherited from neurgoo._base.AbstractModel base class that has all the concrete abstraction required.
Each model is expected to be composed of N number of layers. Each layer is expected to derive from neurgoo._base.AbstractLayer base class.
See neurgoo.models.DefaultNNModel for default implementation.
II) losses
Loss components are required to train any model in a supervised manner.
All the loss layers are inherited from neurgoo._base.AbstractLoss base class for which loss and gradient methods are to be implemented.
neurgoo.losses.MeanSquaredError is one implementation in this version.
III) optimizers
Optimizers are the heart of the training where weights/biases and all the "trainable" parameters are to be updated after backpropagation.
For now, optimizers gradient update method (neurgoo._base.AstractOptimizer.step() method works on any parameters that are of type neurgoo._base.OptimParam
trainers
Once we have the above 3 things -- model, loss, optimizer -- we can use any trainer (see: neurgoo.trainers) to train the model.
A test implementation of linear regression exists at tests.test_models.test_linear_regression() function.
Installation
python setup.py install
pip install -e .