Autograd

Autograd

import torch 

x = torch.tensor([1,5,2],dtype=torch.float32,requires_grad=True)
print(x)
# tensor([1., 5., 2.], requires_grad=True)

Consider the following function : $ y = x^{2} + 2x +1 $

a =  x**2 
b = a+ 2*x 
y = (b +1).sum() 
print(y)
y.backward()
print(x.grad)

output

tensor(49., grad_fn=)

tensor([ 4., 12., 6.])

---

Hidden

import torch 
from torch import optim 
from matplotlib import pyplot as plt 

def derivative_viz(plot:bool=False):
    def middle(fun):
        def inner(*a,**k):
            start = k.get('start',-13)
            end = k.get('end',17)
            spacing = k.get('steps',100)
            if plot:
                x = torch.linspace(start,end,spacing,requires_grad=True)
                print(f'Graph of ({fun.__name__})')
                y = fun(x)
                plt.plot(x.detach().numpy(),y.detach().numpy())
                plt.show()
                print(f'Graph of d/dx({fun.__name__})')
                y = y.sum()
                y.backward()
                plt.plot(x.detach().numpy(),x.grad.detach().numpy())
            return fun(*a)
        return inner 
    return middle

@derivative_viz()
def sin(x):
    return torch.sin(x)

xa = torch.tensor([45])
print(sin(xa,start=0,stop=20))