amartya18x / easyGrad

An easy implementation of Symbolic gradient framework

Easy Grad

EasyGrad is now available via pip and it is the easiest way to use it. To install easyGrad via pip

   pip install easyGrad

To install it from source, clone this repository and run

   python setup.py install

This library aims to provide an easy implementation of doing symbolic operations in python

Operators allowed

• exp(x) and log(x)

def testOps():
    x = Integer('x')
    y = ops.log(x)
    z = ops.exp(y)
    graph = GradGraph(z)
    graph.getOutput({x: 1})
    graph.getGradients(wrt=x)
    print x.gradient

• Sigmoid(x)

def activ_fns():
    x = Double('x')
    z = ops.sigmoid(x)
    graph = GradGraph(z)
    graph.getOutput({x: 110.5})
    graph.getGradients(wrt=x)
    print x.gradient

• Tanh(x)

def activ_fns():
    x = Double('x')
    z = ops.tanh(x)
    graph = GradGraph(z)
    graph.getOutput({x: 110.5})
    graph.getGradients(wrt=x)
    print x.gradient

Testing

def gradTestSimple():  
    a = Integer("a")  
    b = Integer("b")  
    c = a + b  
    d = b + 6  
    e = c * d  
    graph = GradGraph(e)  
    graph.getOutput({a: 32,  
                     b: 11})  
    graph.getGradients(wrt=b)  
    print a.gradient, b.gradient  

Here is a more complex example.

def gradTest():  
    x = Integer("Int1x")  
    y = Integer("Int2y")  
    z = Integer("Int3z")  
    p = Integer("Int4p")  
    k = p * z  
    t = y * k  
    m = k + t  
    n = m * z  
    graph = GradGraph(n)  
    graph.getOutput({x: 9,  
                     y: 9,  
                     z: 9,  
                     p: 2})  
    graph.getGradients(wrt=z)  
    print x.gradient, y.gradient, z.gradient, p.gradient

This is the same examples as above but the commands are not three op commands.

def gradTestLong():  
    x = Integer("Int1x")  
    y = Integer("Int2y")  
    z = Integer("Int3z")  
    p = Integer("Int4p")  
    k = p * z  
    n = (k + (y * p * z)) * z  
    graph = GradGraph(n)  
    graph.getOutput({x: 9,  
                     y: 9,  
                     z: 9,  
                     p: 2})  
    graph.getGradients(wrt=z)  
    print x.gradient, y.gradient, z.gradient, p.gradient

Tensor Operations

def dotProduct():
    x = DoubleTensor("Tensor1")
    y = x.dot([3, 4])
    z = y.dot([4, 5])
    graph = GradGraph(z)
    output = graph.getOutput({x: [3, 4]})
    graph.getGradients(wrt=x)
    assert(np.all(output == [100, 125]))
    assert(np.all(x.gradient == [[ 12., 16.], [ 15., 20.]]))
	


def TensorOp():
    x = DoubleTensor("Tensor1")
    y = x - [3, 4]
    z = ops.log(y * x)
    graph = GradGraph(z)
    output = graph.getOutput({x: [10]})
    assert(np.all(np.isclose(output, np.log(10 * (10 - np.asarray([3, 4]))))))
    graph.getGradients(wrt=x)
    a = 2 * 10 - np.asarray([3, 4])
    b = 1.0/np.exp(np.asarray(output))
    assert(np.all(np.isclose(x.gradient, a * b)))