pytorch-functional 0.7.0

Provides functional API for model creation in PyTorch.

Pytorch Functional

Pytorch Functional is MIT licensed library that adds functional API for model creation to PyTorch.

Defining complex models in PyTorch required creating classes. Defining models in Keras is easier. Pytorch Functional makes it just as easy.

With Pytorch Functional, you can create neural networks without tedious calculations of input shape for each layer.

Features:

• Small extension of PyTorch
• No dependencies besides PyTorch
• Produces models entirely compatible with PyTorch
• It is faster or just as fast as equivalent models (see benchmarks)
• Reduces the amount of code you write
• Works well with complex architectures
• Code and documentation is automatically tested

Example

To create a functional model, you'll use symbolic tensors and nn.Modules. You can add any layer inheriting from nn.Module to your model by calling layer(symbolic_tensor) or equivalently symbolic_tensor(layer).

You can create functional model just like in Keras: by calling the modules and symbolic tensors as if they were normal tensors. Layers will be automagically registered in your model.

from torch import nn
from pytorch_functional import Input

x = Input(shape=(1, 28, 28))  # Input is a SymbolicTensor
print(x)

x = nn.Flatten()(x)  # Every layer returns another SymbolicTensor
x = x(nn.Flatten())  # This is equivalent
print(x)

<SymbolicTensor at 0x7faf5779d130; child of 0; parent of 0>
<SymbolicTensor at 0x7fafea899f10; child of 1; parent of 0>

Using symbolic tensors, we can define a working classificator in a few lines of code:

from torch import nn
from pytorch_functional import FunctionalModel, Input

inputs = Input(shape=(1, 28, 28))
x = nn.Flatten()(inputs)
x = nn.Linear(x.shape[1], 10)(x)(nn.ReLU())
model = FunctionalModel(inputs=inputs, outputs=x)
model

FunctionalModel(
    (module000_depth001): Flatten(start_dim=1, end_dim=-1)
    (module001_depth002): Linear(in_features=784, out_features=10, bias=True)
    (module002_depth003): ReLU()
)

See more examples in Quick Start.

Gentle introduction

There's a notebook showing the basic usage of Pytorch Functional. With it you can:

• Learn Pytorch Functional in an interactive way
• See visualizations of graphs that are created under the hood
• Try the package out before installing it on your computer

Click:

Installation

Install easily with pip:

pip install pytorch-functional

Links

• See Documentation
• See on GitHub
• See on PyPI

Create an issue if you have noticed a problem! Send me an e-mail if you want to get involved: [e-mail].