A flexible API for instantiating pytorch neural networks composed of sequential linear layers (torch.nn.Linear). Additionally, makes use of other elements within the torch.nn module.

Test implementation 1: Sequential linear neural network

import flexinet

nn = flexinet.models.NN()

# example
nn = flexinet.models.compose_nn_sequential(in_dim=50,
                                           out_dim=50,
                                           activation_function=Tanh(),
                                           hidden_layer_nodes={1: [500, 500], 2: [500, 500]},
                                           dropout=True,
                                           dropout_probability=0.1,
                                           )

Test implementation 2: vanilla linear VAE

Installation

To install the latest distribution from PYPI:

pip install flexinet

Alternatively, one can install the development version:

git clone https://github.com/mvinyard/flexinet.git; cd flexinet;

pip install -e .

Example

import flexinet as fn
import torch

X = torch.load("X_data.pt")
X_data = fn.pp.random_split(X)
X_data.keys()

dict_keys(['test', 'valid', 'train'])

model = fn.models.LinearVAE(X_data,
                            latent_dim=20, 
                            hidden_layers=5, 
                            power=2,
                            dropout=0.1,
                            activation_function_dict={'LeakyReLU': LeakyReLU(negative_slope=0.01)},
                            optimizer=torch.optim.Adam
                            reconstruction_loss_function=torch.nn.BCELoss(),
                            reparameterization_loss_function=torch.nn.KLDivLoss(),
                            device="cuda:0",
                           )

model.train(epochs=10_000, print_frequency=50, lr=1e-4)

model.plot_loss()

Contact

If you have suggestions, questions, or comments, please reach out to Michael Vinyard via email