kungtalon / adgc

This repo implements a deep learning framework and a standalone multi-array module, aiming to help everyone grasp a deeper understanding of tensor operations and automatic differentiation mechanisms in TensorFlow, PyTorch etc.

ADGC

This repo implements a tensor module and a deep learning framework. It aims to help everyone grasp a better and clearer understanding of tensor operations and automatic differentiation mechanisms in TensorFlow, PyTorch etc. The three-layer MLP built on top of my framework has achieved 86% accuracy on MNIST dataset.

Tensor Module

This part implements a standalone module for creating and manipulating multidimentional array like Numpy. The operations are based on CPU with BLAS.

Some of the methods implemented:

• dot, multiply, add: do some basic algorithm functions for matrices/tensors
• reshape: change the shape of the tensor
• transpose: switch two axes of the tensor
• slice: slice tensor along several axes with given indices
• sum: sum the values along an axis
• fill_diag: fill the diagonal entries with a vector
• map: accept a lambda function and transforms the value of each entry
• kron: do the matrix kronecker product

We are trying to learn BLAS and make full use of its extreme performance. Any advice would be very appreciated! :)

Deep Learning Auto-Differentiation Framework

This part contains all elements with regard to building an automatic differential graph and a deep learning model trainer. There are different types of nodes on the graph:

• Node: base class for all nodes; it is abstract and no backward or forward is defined
• Variable: leaf nodes controlling the user's input
• Parameter: trainable values of the neural networks
• Ops: algorithmic operations between different nodes
• Functional: activation functions and loss functions, including CrossEntropy, Softmax...

These are not nodes but important to building and training the model:

• Layer: some common layers built on top of the above nodes, currently including:
  • Fully connected layer
  • Conv2d layer
  • BatchNorm2d layer
• Optimizer: encapsulations of gradient-based learning algorithms, currently including:
  • Gradient Descent
  • Adam

Forward and Backward

Each node of the graph needs to implement the logics of do_forward() and do_backward(Node* parent_ptr), the first one computes its forward value and the other computes the nodes' jacobian matrix with regard to one of its parent nodes.

A code snippet about how to use the framework:

  Variable x = Variable({ 2, 2 });
Variable labels = Variable({ 2, 2 });

tensor::Tensor<double> value_x = tensor::Tensor<double>({ 2, 2 }, { 1, 2, 3, 4 });
tensor::Tensor<double> value_labels = tensor::Tensor<double>({ 2, 2 }, { 1, 0, 1, 0 });

// build graph
layer::Dense dense_layer_1(2, 10, "relu");
layer::Dense dense_layer_2(10, 2, "none");

auto loss = functional::cross_entropy_with_softmax(
dense_layer_2(dense_layer_1(x)), labels);

auto optim = optimizer::Adam(loss);

// forward
x.assign_value(value_x);
labels.assign_value(value_labels);
optim.zero_grad();
loss.forward();

// backward
optim.step();

// in the end
Graph::clear_graph();

Demo

There are runnable demos in the directory demo. You can build and run it yourself with the CMakeLists.txt given. Please make sure the dependencies are installed properly.

mkdir build
cd build
cmake -DSKIP_TEST=ON ..
make mnist_trainer
./mnist_trainer $BATCH_SIZE $EPOCH $TRAIN_DATA_PATH $TEST_DATA_PATH

Graph Visualization

We are using Graphviz to visualize the computational graph. Call graph->visualize("file.svg") to convert the graph into a DOT image and saves as SVG file.

Use -DUSE_GRAPHVIZ=OFF in cmake command line to turn off using graphviz.

Example output:

Dependency

Here are some software and packages used in this project:

mkl (w/ cblas)

graphviz

cgraph

googletest