Code for explaining fundamentals of back propagation on a compute graph. This uses computational primitives e.g., additions and multiplication, to build the graph for forward computation and back propagation. It results in a slighlty more verbose code while setting up the network but everything also becomes very explicit, hence easy to understand.
I am
Main code is in core and core.np packages. core.np package contains code that works with matrices, core package has code for simple functions.
Tests are in "tests" directory.
I have also committed some juputer notebooks that I am using to run equivalent pytorch computations to compare the results. These are in "jupyter" folder.
For the code, you primarily need numpy but I have also used networkx and pyvis for visualization which I should probably remove
I have manually changed the contents and removed most of the items listed when using pip freeze. Create a new conda or basic virtual environment, activate it and then run the following command from within the environment
pip install -r pip-requirements.txt
Using conda does not yet work
Using conda - this has been created using
conda list --export > conda-requirements.txt
You can create a new conda environment using the command
conda create --name <envname> --file conda-requirements.txt
Jupyter notebooks require pytorch for most of the notebooks. There is one notebook that uses symy. You can skip this notebook if you do not wish to install sympy
python -m unittest tests.core.np.TestDenseLayer.DenseLayerStandAlone.test_linear_optimization