There's plenty of excellent (tinygrad) and minimalist (micrograd) built-from-scratch, deep learning frameworks out there, so the goal of fauxgrad is to sacrifice some of the full functionality, and focus on the general idea and building blocks for writing your own.
The walkthrough/tutorial can be found in this notebook.
pip install fauxgrad
Calculating some gradients:
from fauxgrad import Value
a = Value(2.3)
b = Value(-1)
c = (-a * b).log()
l = -(c.sigmoid() + b) + a
l.backward()
print('The derivative that we computed before, dl/da:', a.grad)
>>> 0.91Plotting the backward pass graph:
from fauxgrad.utils import plot_graph
plot_graph(l) # green node is l, light blue nodes have no parents
