torchluent 0.0.3

Build pytorch models in a fluent interface

PyTorch Fluent Models

A small package that provides a fluent interface for creating pytorch models.

Summary

A fluent interface is roughly one where you chain method calls. Read more about fluent interfaces here.

This library allows for dense layers, convolution layers, max pooling, and nonlinearities or other operators (i.e. normalization). This calculates the new shape after each layer, meaning you do not have to redundantly specify features.

Consider the following pure PyTorch code:

import torch.nn as nn

net = nn.Sequential(
    nn.Linear(28*28, 128),
    nn.Linear(128, 10)
)

The input to the second layer (128) must always match the output of the first layer. This redundancy is very small but can be improved. The issue becomes even more apparent when you consider convolution layers.

Furthermore, the official PyTorch library does not include some common glue code for extensive sequential blocks. One possible reason for this is that Fluent API's are unlikely to be as exhaustive as conventional API's so one will often have to fall back on the more verbose module definition anyway.

Finally, this has the extremely versatile then and then_with which work for transposed convolution layers and unpooling while still avoiding redundant layer sizes or channel numbers.

API Reference

https://tjstretchalot.github.io/torchluent/

Usage

Create an instance of torchluent.FluentModule with the shape of your input. There are a few meta functions on FluentModule, such as .verbose() which will print how the shape changes through progressive calls. For layers which change the number of features one can call .transform in the generic sense or use one of the provided functions such as .dense which will calculate the new number of features. For layers which do not change the shape of the data, rather than including a function for each one you may use .operator which accepts the name of the attribute in torch.nn as well as an arguments or keyword arguments.

Installation

pip install torchluent

Examples

from torchluent import FluentModule

print('Network:')
net = (
    FluentModule((1, 28, 28))
    .verbose()
    .conv2d(32, kernel_size=5)
    .maxpool2d(kernel_size=3)
    .operator('LeakyReLU', negative_slope=0.05)
    .flatten()
    .dense(128)
    .operator('ReLU')
    .dense(10)
    .operator('ReLU')
    .build()
)

print(net)

Produces:

Network:
  (1, 28, 28)
  Conv2d -> (32, 24, 24)
  MaxPool2d -> (32, 8, 8)
  LeakyReLU
  Reshape -> (2048,)
  Linear -> (128,)
  ReLU
  Linear -> (10,)
  ReLU

Sequential(
  (0): Conv2d(1, 32, kernel_size=(5, 5), stride=(1, 1))
  (1): MaxPool2d(kernel_size=3, stride=3, padding=0, dilation=1, ceil_mode=False)
  (2): LeakyReLU(negative_slope=0.05)
  (3): Reshape(2048)
  (4): Linear(in_features=2048, out_features=128, bias=True)
  (5): ReLU()
  (6): Linear(in_features=128, out_features=10, bias=True)
  (7): ReLU()
)

Wrapping and Unwrapping

One concept which is not in PyTorch by default is a way to consider the hidden state of an arbitrary network in an abstract way. The idea is basically that it is often nice if a module returns an array in addition to the transformed output, where each element in the returned array is a snapshot of the input as it propagated through the network.

The following is a contrived example that illustrates what such a module might look like:

import torch.nn as nn

class HiddenStateModule(nn.Module):
    def forward(self, x):
        result = []
        result.append(x) # initial state always there
        x = x ** 2
        result.append(x) # where relevant
        x = x * 3 + 2
        x = torch.relu(x)
        result.append(x)
        return x, result

This module means to expose this concept without having to modify the underlying transformations (i.e. nn.Linear) nor be forced to fallback on creating a custom Module just for this extremely common situation.

However, another problem that arises with this type of module is that this result will break much of your codebase if it expects a single output. This is most problematic when combined with some abstract training paradigm such as PyTorch Ignite. Luckily, it's very easy to just drop the second output from such a module, as if by the following

import torch.nn as nn

class StrippedStateModule(nn.Module):
    def __init__(self, mod):
        super().__init__()
        self.mod = mod

    def forward(self, x):
        return self.mod(x)[0]

By including the array in the main implementation and then using such an "unwrapping" module you can get the best of both worlds. For training and generic usage which does not need the hidden state, use the stripped version. For analysis which desires the hidden state, use the pre-stripped version.

With this context in mind, the following code snippet will produce both the wrapped and unwrapped versions of the network:

from torchluent import FluentModule

print('Network:')
net, stripped_net = (
    FluentModule((28*28,))
    .verbose()
    .wrap(with_input=True) # create array and initialize with input
    .dense(128)
    .operator('ReLU')
    .save_state() # pushes to the array
    .dense(128)
    .operator('ReLU')
    .save_state()
    .dense(10)
    .operator('ReLU')
    .save_state()
    .build(with_stripped=True)
)
print()
print(net)

Produces

Network:
  (784,)
  Linear -> (128,)
  ReLU
  Linear -> (128,)
  ReLU
  Linear -> (10,)
  ReLU

Sequential(
  (0): InitListModule(include_first=True)
  (1): WrapModule(
    (child): Linear(in_features=784, out_features=128, bias=True)
  )
  (2): WrapModule(
    (child): ReLU()
  )
  (3): SaveStateModule()
  (4): WrapModule(
    (child): Linear(in_features=128, out_features=128, bias=True)
  )
  (5): WrapModule(
    (child): ReLU()
  )
  (6): SaveStateModule()
  (7): WrapModule(
    (child): Linear(in_features=128, out_features=10, bias=True)
  )
  (8): WrapModule(
    (child): ReLU()
  )
  (9): SaveStateModule()
)

Limitations

For non-trivial networks there will likely be significant usage of the then and then_with functions which aren't quite as nice as the examples shown above, but I believe they are still a significant improvement.