computational-graph

A simple library that builds computational graphs and provides gradient computation capabilities. Currently supports only simple datatypes such as int and float.

Please note that this code has been tested only with Python3.6 and above. The code is not guaranteed to work correctly on lower versions of Python.

Currently supported operations

• Addition
• Subtraction
• Multiplication
• Division
• Exponentiation

Other feratures-

• Automatically casts ints, floats to Var
• gradient function provides autodiff capabilities

Examples

Import the library-

>>> import computational_graph as cg
>>> from computational_graph import Var, gradient

Basic ops

x = Var(5)
y = Var(6)
z = x + y	# Addition
print(z) 	# 11

type(z)		# <class 'computational_graph.basic.Var'>

# Also note that, we can also perform an operation with a Var and an int or float-
z = x + 6	# 11
type(z)		# <class 'computational_graph.basic.Var'>

# Other supported basic ops
z = x - y	# Subtraction
z = x * y	# Multiplication
z = x / y	# Division
z = x ** y	# Exponentiation

Other ops

Sigmoid

import computational_graph as cg
from computational_graph import Var, gradient


def sigmoid(x):
	exp_ = cg.exp()
	return 1 / (1 + exp_(-x))

x = Var(0.5)
z = sigmoid(x)
print(z)		# 1.6065306597126334

Gradient computation

import computational_graph as cg
from computational_graph import Var, gradient


def sigmoid(x):
	exp_ = cg.exp()
	return 1 / (1 + exp_(-x))

x = Var(0.5)
z = sigmoid(x)
print(z)				# 1.6065306597126334

# Calculate dz/dx (Gradient of z w.r.t x)
print(gradient(z, x)) 	# -0.6065306597126334

Future work

• Extend to numpy arrays
• Add functionality for plotting the computational graph
• Add support for detaching gradients
• Add support for commonly used neural network functions (e.g. sigmoid, softmax, ReLU, etc.)
• Detailed documentation (if things work out and I'm able to add numpy arrays support )