Help with running this code on Windows 10

[nitzankoll]

How can I disable cudnn for batch_norm in Windows 10? I tried to run the code with out doing it, I got this output: File "C:\Users\pc\Desktop\pose-project\human-pose-estimation.pytorch-master\pose_estimation..\lib\core\function.py", line 52, in train loss.backward() File "C:\Users\pc\Anaconda3\envs\tensorflow\lib\site-packages\torch\tensor.py", line 102, in backward torch.autograd.backward(self, gradient, retain_graph, create_graph) File "C:\Users\pc\Anaconda3\envs\tensorflow\lib\site-packages\torch\autograd_init_.py", line 90, in backward allow_unreachable=True) # allow_unreachable flag RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

In the "main" function I added -"with torch.no_grad():" to solve a memory error. Maybe that is the problem? Any help will be appreciated.

[aceimnorstuvwxz]

Does pytorch not have this file on Windows10?

${PYTORCH}/torch/nn/functional.py

(I'm going to run this code in Win10 too, but not already)

[pallsopp]

Looks like all that is doing is replacing "torch.backends.cudnn.enabled" with "False" in functional.py. That file should be found in your python site-package folder, under "torch/nn/". For example, for me that file is located at: E:\Python37\lib\site-packages\torch\nn\functional.py or if you use git-bash as I do: /e/python37/lib/site-packages/torch/nn/functional.py

So you can open that file, find all occurrences of "torch.backends.cudnn.enabled" (in the function returns), and simply return False in place.

For example, change:

return torch.instance_norm(
  input, weight, bias, running_mean, running_var,
  use_input_stats, momentum, eps, torch.backends.cudnn.enabled
)

to:

return torch.instance_norm(
  input, weight, bias, running_mean, running_var,
  use_input_stats, momentum, eps, False
)

HTH