torch-composer 0.0.4rc0

API to compose pytorch neural networks

torch-composer

Compose pytorch neural networks with ease.

Installation (current version: v0.1.0)

pip install torch-composer

Sample use-cases of the API

import torch_composer
import torch

net = torch_composer.TorchNet(in_dim=2500, out_dim=10)

Sequential(
  (input): Linear(in_features=2500, out_features=200, bias=True)
  (activation_1): LeakyReLU(negative_slope=0.01)
  (hidden_1): Linear(in_features=200, out_features=200, bias=True)
  (output_activation): LeakyReLU(negative_slope=0.01)
  (output): Linear(in_features=200, out_features=10, bias=True)
)

As simple as you want (see above) or more complex with optional parameters:

torch_composer.TorchNet(
    in_dim=2500,
    out_dim=10,
    hidden={1: [800, 800], 2: [200, 200]},
    activation_function=torch.nn.LeakyReLU(negative_slope=0.01),
    dropout=0.2,
    input_bias=True,
    output_bias=True,
)

Sequential(
  (input): Linear(in_features=2500, out_features=800, bias=True)
  (activation_1): LeakyReLU(negative_slope=0.01)
  (hidden_1): Linear(in_features=800, out_features=800, bias=True)
  (dropout_1): Dropout(p=0.2, inplace=False)
  (activation_2): LeakyReLU(negative_slope=0.01)
  (hidden_2): Linear(in_features=200, out_features=200, bias=True)
  (dropout_2): Dropout(p=0.2, inplace=False)
  (output_activation): LeakyReLU(negative_slope=0.01)
  (output): Linear(in_features=200, out_features=10, bias=True)
)

Make an encoder

torch_composer.TorchNetDecoder(data_dim=2500, latent_dim=10)

Make a decoder

torch_composer.TorchNetDecoder(data_dim=2500, latent_dim=10)

Access and set initial parameters for the output layer:

torch_composer.tools.init_output_params(net)

Potential future plans

• Composition of torch.optim funcs.