Windows
Download the wheel for openEXR from:
Python Extension Packages for Windows – Christoph Gohlke (uci.edu)
|
1 2 |
python -m pip install OpenEXR-1.3.0-cp36-cp36m-win_amd64.whl |
Ubuntu:
|
1 2 3 |
sudo apt-get install zlib1g-dev sudo apt-get install libopenexr-dev sudo apt-get install openexr |
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 |
grads = {} def save_grad(name): def hook(grad): grads[name] = grad return hook x = Variable(torch.randn(1,1), requires_grad=True) y = 3*x z = y**2 # In here, save_grad('y') returns a hook (a function) that keeps 'y' as name y.register_hook(save_grad('y')) z.register_hook(save_grad('z')) z.backward() print(grads['y']) print(grads['z']) |
Thanks to Adam Paszke’s post in Pytorch Discussion
I struggled with a problem today: My parameter “b” is not updating in the following code:
|
1 2 |
b = nn.Parameter(torch.ones(batch_size, 1) a = torch.norm(b, dim=1) |
There’s nothing wrong with the gradient of “a”. So what’s the problem?
The problem is: I used the wrong initialization of “b”. I init “b” with all zeros. and the gradient of the norm of an all-zero vector is always zero.
Today I met an error:
a is a float32 pytorch tensor; and b is a float64 pytorch tensor with all elements equals 1. When I perform the following multiplication:
|
1 |
c = a * b |
The result of “c” is 0.
The error will be fixed when b is converted to float32.
|
1 2 3 4 |
value = 0.0205862828046875 print("%.16f" % value) value = value.astype(np.float32) print("%.16f" % value) |
|
1 2 3 |
#output 0.0205862828046875 0.0205862820148468 |
Precision changed!