[GitHub] [incubator-mxnet] apeforest commented on issue #15288: [MXNET-978] Higher order gradient for sigmoid

[GitBox <gi...@apache.org>]

apeforest commented on issue #15288: [MXNET-978] Higher order gradient for sigmoid
URL: https://github.com/apache/incubator-mxnet/pull/15288#issuecomment-507516653
 
 
   I verified the result is the same as pytorch
   
   ```
   import torch
   import numpy as np
   import math
   
   op = lambda x: torch.sigmoid(x)
   grad_op = lambda x: op(x) * (1 - op(x))
   grad_grad_op = lambda x: grad_op(x) * (1 - 2 * op(x))
   grad_grad_grad_op = lambda x: grad_grad_op(x) - 2 * ( grad_op(x)**2 + grad_grad_op(x) * op(x))
   
   x = torch.tensor(np.array([1, 2, 3]), dtype=torch.float32)
   head_grads = torch.tensor(np.array([1, 1, 1]), dtype=torch.float32) * 0.5
   head_grad_grads = torch.tensor(np.array([1, 1, 1]), dtype=torch.float32) * 0.6
   head_grad_grad_grads = torch.tensor(np.array([1, 1, 1]), dtype=torch.float32) * 0.7
   x.requires_grad = True
   head_grads.requires_grad = True
   
   y = op(x)
   x_grad = torch.autograd.grad(y, x, grad_outputs= head_grads, create_graph=True, retain_graph=True)[0]
   expected_grad_x = (grad_op(x) * head_grads).detach().numpy()
   print('expected_grad_x = {}'.format(expected_grad_x))
   print('grad_x          = {}'.format(x_grad.detach().numpy()))
   x_grad_grad = torch.autograd.grad(x_grad, x, grad_outputs= head_grad_grads, create_graph=True, retain_graph=True)[0]
   x_grad_grad.backward(head_grad_grad_grads)
   
   expected_grad_grad_x = (grad_grad_op(x) * head_grads * head_grad_grads).detach().numpy()
   expected_head_grad = (grad_op(x) * head_grad_grads).detach().numpy()
   expected_grad_grad_grad_x = (grad_grad_grad_op(x) * head_grads * head_grad_grads * head_grad_grad_grads).detach().numpy()
   
   print('expected_grad_grad_x = {}'.format(expected_grad_grad_x))
   print('grad_grad_x          = {}'.format(x_grad_grad.detach().numpy()))
   print('expected_grad_grad_grad_x = {}'.format(expected_grad_grad_grad_x))
   print('grad_grad_grad_x          = {}'.format(x.grad.detach().numpy()))
   ```

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services