TensorSlow is a minimalist machine learning API that mimicks the TensorFlow API, but is implemented in pure python (without a C backend). The source code has been built with maximal understandability in mind, rather than maximal efficiency. Therefore, TensorSlow should be used solely for educational purposes. If you want to understand how deep learning libraries like TensorFlow work under the hood, this may be your best shot.
I have written an article in my blog at deepideas.net
that develops this library step by step, explaining all the math and algorithms on the way: http://www.deepideas.net/deep-learning-from-scratch-theory-and-implementation/