nn-builder 1.0.5

nn_builder lets you build neural networks without boilerplate code. You specify the type of network you want and it builds it.

Install

pip install nn_builder

Support

Examples

On the left is how you can create the PyTorch neural network on the right in only 1 line of code using nn_builder:

Similarly for TensorFlow on the left is how you can create the CNN on the right in only 1 line of code using nn_builder:

Usage

See this colab notebook for lots of examples of how to use the module. 3 types of PyTorch and TensorFlow network are currently supported: NN, CNN and RNN. Each network takes the following arguments:

Each network type has slightly different requirements for input_dim and layers_info as explained below:

1. NN

• input_dim: # Features in PyTorch, not needed for TensorFlow
• layers_info: List of integers to indicate number of hidden units you want per linear layer.
• For example:

from nn_builder.pytorch.NN import NN   
model = NN(input_dim=5, layers_info=[10, 10, 1], output_activation=None, hidden_activations="relu", 
           dropout=0.0, initialiser="xavier", batch_norm=False)            

2. CNN

• input_dim: (# Channels, Height, Width) in PyTorch, not needed for TensorFlow
• layers_info: We expect the field layers_info to be a list of lists indicating the size and type of layers that you want. Each layer in a CNN can be one of these 4 forms:
  • ["conv", channels, kernel size, stride, padding]
  • ["maxpool", kernel size, stride, padding]
  • ["avgpool", kernel size, stride, padding]
  • ["linear", units]
• For a PyTorch network kernel size, stride, padding and units must be integers. For TensorFlow they must all be integers except for padding which must be one of {“valid”, “same”}
• For example:

from nn_builder.pytorch.CNN import CNN   
model = CNN(input_dim=(3, 64, 64), 
            layers_info=[["conv", 32, 3, 1, 0], ["maxpool", 2, 2, 0], 
                         ["conv", 64, 3, 1, 2], ["avgpool", 2, 2, 0], 
                         ["linear", 10]],
            hidden_activations="relu", output_activation="softmax", dropout=0.0,
            initialiser="xavier", batch_norm=True)

3. RNN

• input_dim: # Features in PyTorch, not needed for TensorFlow
• layers_info: We expect the field layers_info to be a list of lists indicating the size and type of layers that you want. Each layer in a CNN can be one of these 4 forms:
  • ["lstm", units]
  • ["gru", units]
  • ["linear", units]
• For example:

from nn_builder.pytorch.CNN import CNN   
model = RNN(input_dim=5, layers_info=[["gru", 50], ["lstm", 10], ["linear", 2]],
            hidden_activations="relu", output_activation="softmax", 
            batch_norm=False, dropout=0.0, initialiser="xavier")

Contributing

Together we can make something that is useful for thousands of people. Anyone is very welcome to contribute via a pull request. Please see the issues page for ideas on the best areas to contribute to and try to:

1. Add tests to the tests folder that cover any code you write
2. Write comments for every function
3. Create a colab notebook demonstrating how any extra functionality you created works