CFlow was developed as an extension to the C project at the Department of Computing at Imperial College London.
We chose this project to gain an understanding of how deep learning models function under the hood and develop our C programming and team working skills.
CFlow attempts to emulate the Keras Functional API. What this means is that developers should be able to use CFlow to make neural networks with many input and output heads.
Like Tensorflow, CFlow represents neural networks as computational graphs, with data nodes and operation nodes. Data nodes are used to store both weights and biases and to store input and output data. Operation nodes represent different matrix operations such as: the dot product, convolution, the hadamard product and addition.
Graphs can be created by users, either through programming them, or by writing a .graph file, and reading it with graphFileRead. Once a network has finished training, it can be stored as a .graph file while its weights and biases would be stored in an associated .data file. graphFileWrite stores the graph node by node using breadth first search, to make it easier for people to read its output.
As mentioned before, nodes represent either data (in the form of matrices) or matrix operations. You may notice that nodes have the matrices: poolingMatrixGrad and optimiserMatrix, these are used during backpropagation. PoolingMatrixGrad is used to store the indices used during max pooling. OptimiserMatrix is used by optimisers to store the gradient accumulations during training.
Nodes also store pointers to their input and output nodes, which allow graphs to be traversed bidirectionally.