Once upon a time, there lived a little Data Scientist. He had quite a
good understanding of LightFM model for recommender systems and
wanted to reimplement it using PyTorch since the model's architecture
was very close to those of neural networks. And our little Data
Scientist wholeheartedly believed that his code in deep learning
framework strictly replicated the original LightFM behaviour. But the
model didn't train well. And the little Data Scientist spent several
working days looking for a reason why and trying different things.
Then he wrote a simple unit-test to check the coherence of the neural
model with the classical one and found the bug, really stupid and
ugly one, in his code. The end.
A word to the wise: if don't have a test for some property of your
code, with high probability the property doesn't hold.
