/ Home

Torch Samples

Note: Basics Torch Samples come here

Check cuda available

import torch

torch.cuda.is_available()

Simple Autograd function

import torch
from torch.autograd import Function

class Exp(Function):

  @staticmethod
  def forward(ctx, i):
    result = i.exp()
    ctx.save_for_backward(result)

    return result

  @staticmethod
  def backward(ctx, grad_output):
    result, = ctx.saved_tensors()

    return grad_output * result

input = torch.rand(2, 2)

output = Exp.apply(input)

print(output)

Simple Jacobian

import torch
from torch.autograd.functional import jacobian

def exp_reducer(x):

  return x.exp().sum(dim = 1)

inputs = torch.rand(2, 2)

print(inputs)

jacobian(exp_reducer, inputs)

jacobian(exp_reducer, inputs, create_graph = True)

def exp_adder(x, y):

  return 2 * x.exp() + 3 * y

inputs = (torch.rand(2), torch.rand(2))

jacobian(exp_adder, inputs)

Ref :