TF-like-framework
A reconstruction of bare bones TensorFlow 1
My implementation of a computational graph to train neural networks. Similar to TensorFlow V1. The main guts are in the module Framework.py.
There, I define the main pieces of the computational graph, as well as implement a number of operators on layers, including convolution, pooling, dropout, batchnorm, etc. Numerical gradient checking on networks can be performed by importing the module gradient_checking.py
There are two example notebooks. The first, main.ipynb, is using the Framework.py module to train a simple two-three layer neural network on a toy spiral dataset, where the goal is to classify different arms of a spiral. In the notebook conv_main.py, we load part of the CIFAR-10 dataset and train both a three layer neural net and a three layer convolutional net on the training set.
The framework is paired with the graphviz module to automatically generate and display the computational graph associated with a net.