simple-pytorch-wrapper 0.0.1

A lightweight PyTorch wrapper for fast and easy neural network training and evaluation.

Simple-PyTorch-wrapper: assess fast and easy your neural network

Importance of easy and quick assessment of quality

A lightweight PyTorch wrapper that can be used to fasten process of training and setting up arbitrary Neural Network to quickly test an idea/setup. The wrapper provides an interface for both standard neural networks and CNNs, but can be extended to any architecture, with built-in visualization and performance tracking capabilities. The wrapper is customizable and aims to be used on any dataset.

Usage

pip install simple_pytorch_wrapper

This package is aimed to be as simple as possible. The following example trains on the example dataset in this package, using a simple feedforward network.

def main():
    # For reproducibility
    set_seed(0)

    # Example data
    X, Y = load_language_digit_example_dataset()

    # Transform data accordingly (vectorizing) + squeezing for batching
    X, Y = PytorchWrapper.vectorize_data(X, Y, NetworkType.FNN) 

    # Pytorch wrapper
    wrapper = PytorchWrapper(X, Y)  
    
    network = FNNGenerator(
    input_size=64*64,  # flattened value of image of example dataset
    output_size=10,    # number of outputs
    hidden_layers=[100, 500],  # sizes of hidden layers
    hidden_activations=[nn.ReLU(), nn.ReLU()],  # activation functions for each hidden layer
    )
    
    # Create custom pytorch network
    wrapper.upload_pyTorch_network(network) 

    wrapper.setup_training(batch_size=32, learning_rate=0.01, epochs=10) 

    # Training
    wrapper.train_network(plot=True)  # Enable plotting

    # Visualization
    wrapper.visualize()

You see that training a network from start to finish, with a clean visualization in less than 15 effective lines!

Upload your own network!

The network variable above is there to be defined by you. This is aimed to take in any Neural network design. The only caveat is to correctly shape your input. For a feedforward and CNN, these are already implemented by the vectorize_data(*args) function.

To get you started on the networks, there has been provided already NN generator classes for a feed-forward NN and a convolutional one. These can be used respectively by calling FNNGenerator(*args) and CNNGenerator(*args).

The function signature of the FNNGenerator is:

network = FNNGenerator(input_size: int,
              output_size: int,
              hidden_layers: List[int],
              hidden_activations: Optional[List[nn.Module]]
              )

While for the CNNGenerator, this is:

network = CNNGenerator(input_channels: int,
                      conv_layers: List[Dict[str, int]],  
                      fc_layers: List[int], 
                      output_size: int,
                      batch_size: Optional[int] = None,
                      use_pooling: bool = True)

Example runs

It can be hard to know the power of a tool, without having a solid example. By calling FNN_example_run() with the training params as function arguments, you get a run on the example data set for a FNN architecture. The CNN_example_run() does it with a CNN architecture.

Dataset

The data used for the example is the Sign Language Digits Dataset from the Turkey Ankara Ayrancı Anadolu High School students. This dataset is available at ardamavi/Sign-Language-Digits-Dataset.