DeepFusion is a highly modular and customizable deep learning framework.

It is designed to provide strong and explicit control over all data, operations and parameters while maintaining a simple and intuitive code base.

Table of Contents

• Table of Contents
• DeepFusion Framework
• Basic Usage
• Highlights
  • 1. Customizable training
  • 2. Gradient Checking
• Installation
  • 1. Basic Installation
  • 2. GPU Training
  • 3. Network Visualization
  • 4. All Dependencies
• Resources
• Contribution Guidelines
• License
• Acknowledgements
• Credits

DeepFusion Framework

In DeepFusion, all networks are composed by combining 3 basic types of components:

• Data
• Module
• Net

Data objects hold the network activations and Module objects perform operations on them. The Net object forms a thin wrapper around the Data and Module objects and is used to perform the forward and backward passes.

A simple neural network is shown below, where, ellipses represent Data objects and rectangles represent Module.

Note the alternating sequence of Data and Module. The scheme is Data -> Module -> Data. Red represents nodes with updatable parameters.

Every node (Data or Module) has a unique ID (for eg: z1 or MatMul1) using which it can be accessed and modified thus providing explicit access and control over all data and parameters.

More details on Data, Module and Net functionalities can be found in their respective readmes in deepfusion/components.

This is the basic idea behind deepfusion and any and all neural networks are created using this procedure of attaching alternating Data and Module nodes.

Basic Usage

As described before, in DeepFusion, all networks are composed by combining 3 basic types of components:

• Data
• Module
• Net

The codebase follows the same intuitive structure:

deepfusion
└── components
    ├── net
    ├── data
    └── modules
        ├── activation_functions
        ├── loss_functions
        └── matmul.py

To construct a neural network we need to import the Net, Data and required Module objects.

# Import Net, Data and necessary Modules
from deepfusion.components.net import Net
from deepfusion.components.data import Data
from deepfusion.components.modules import MatMul
from deepfusion.components.modules.activation_functions import Relu
from deepfusion.components.modules.loss_functions import MSELoss

The codebase is designed in an intuitive manner. Let's see how we would think about the above imports. "Okay, to create a neural network I need components (deepfusion.components). What kind of components do we need? Net, Data and Modules (import these). What kind of modules (operations) do we need? we need matrix multiplication, an activation function and a loss function (import these). That's it!"

To connect Data and Module objects we need to keep in mind the following 2 things:

• Data objects are used to specify the activation dimensions.
• Module objects require the inputs and output data objects to be specified.

Now, let's construct the simple network we saw above.

# Basic structure: x -> Matmul -> z1 -> Relu -> a -> Matmul -> z2, + y -> MSE -> loss
x = Data(ID = 'x', shape = (1, 3))

z1 = Data(ID = 'z1', shape = (1, 5))
Matmul1 = MatMul(ID = 'Matmul1', inputs = [x], output = z1)

a = Data(ID = 'a', shape = (1, 5))
ActF = Relu(ID = 'ActF', inputs = [z1], output = a)

z2 = Data(ID = 'z2', shape = (1, 1))
Matmul2 = MatMul(ID = 'Matmul2', inputs = [a], output = z2)

# Add target variable, loss variable and loss function
y = Data('y', shape = (1, 1))
loss = Data('loss', shape = (1, 1))
LossF = MSELoss(ID = 'LossF', inputs = [z2, y], output = loss)

# Initialize the neural network
net = Net(ID = 'Net', root_nodes = [loss])

For Data the first dimension is the batch size. This is specified 1 during initialization. Eg: a length 3 vector would have shape = (1, 3) and a conv volume (C, H, W) would have shape = (1, C, H, W). During training any batch size (B, 3) or (B, C, H, W) can be used, the Net object takes care of it.

Module parameter dimensions are inferred from connected data objects.

Examples introducing the basics and all features of the library can be found in the demo directory or in other resources.

To have a look at the codebase tree have a look at Codebase Tree.

Highlights

1. Customizable training

Let's say we make the simple neural network as before:

And train it. During training only the red portions of the network receive updates and are trained. Therefore, the matrix multiplication modules will be trained.

Let's say we have trained the network and now we want to find the input that optimizes the function that we have learnt. This also falls under the same forward-backward-update procedure with the following simple twist:

net.freeze() # Freezes all modules
x.unfreeze() # Unfreezes the input node

After this we obtain the following network:

Now when we train the network only the input node value will get updates and be trained!

2. Gradient Checking

When developing new modules, the implementation of the backward pass can often be tricky and have subtle bugs. Deepfusion provides a gradient checking utility that can find the derivatives of the loss function(s) w.r.t. any specified data object (data node or module parameter). Eg:

# Compare analytic and numeric gradients with a step size of 1e-6 for:
# Input node: x
gradient_checker(net, data_obj = x, h = 1e-6)
# Matrix multiplication parameter W
gradient_checker(net, data_obj = Matmul1.W, h = 1e-6)

Note

Other features such as forward and backward pass profiling, multiple loss functions, automated training, gpu training etc. can be found in the demo directory or in other resources.

Installation

1. Basic Installation

To install the core part of deepfusion use:

$ pip install deepfusion

2. GPU Training

To use GPU training capabilities you will require CuPy which needs the CUDA Toolkit. If the CUDA Toolkit is installed then use:

$ pip install deepfusion[gpu]

3. Network Visualization

For visualizing networks you will require the Graphviz software and the graphviz package. If Graphviz is installed then use:

$ pip install deepfusion[visualization]

4. All Dependencies

If all dependencies are pre-installed use:

pip install deepfusion[gpu,visualization]

Important

Make sure to select add to PATH options when downloading dependency softwares.

Resources

• DeepFusion documentation
• DeepFusion demo
• DeepFusion Tutorials

Contribution Guidelines

Contributions for the following are encouraged and greatly appreciated:

• Code Optimization: Benchmark your results and show a clear improvement.
• Visualization: Currently requires graphviz which is usually a pain to install. Structured graph visualization using say matplotlib would be a clear win.
• More Modules: Most scope for contribution currently in the following modules: loss_functions, pooling, normalizations, RNN modules etc.
• More Features: Some ideas include adding multiprocessing, working with pre-trained models from other libraries etc.
• Testing: Incorporating testing codes.
• Improving Documentation: Improving doc-string clarity and including doc tests. Also perhaps making a website for API reference.

We'll use Github issues for tracking pull requests and bugs.

License

Distributed under the MIT License.

Acknowledgements

Theoretical and code ideas inspired from:

• CS231n: Deep Learning for Computer Vision
• Coursera: Deep Learning Specialization
• Google Python Style Guide
• Udemy: Python Packaging

Credits

• Logo design by Ankur Tiwary