Python – Xiaoxu Meng

Install OpenEXR for Python (Windows and Ubuntu)

Windows

Download the wheel for openEXR from:

Python Extension Packages for Windows – Christoph Gohlke (uci.edu)

python -m pip install OpenEXR-1.3.0-cp36-cp36m-win_amd64.whl

1 2	python -m pip install OpenEXR-1.3.0-cp36-cp36m-win_amd64.whl

Ubuntu:

sudo apt-get install zlib1g-dev
sudo apt-get install libopenexr-dev
sudo apt-get install openexr

sudo apt-get install zlib1g-dev

sudo apt-get install libopenexr-dev

sudo apt-get install openexr

How to print the gradient of intermediate variables in Pytorch

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'])

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:

b = nn.Parameter(torch.ones(batch_size, 1)
a = torch.norm(b, dim=1)

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.

Error about multiply with different precision

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:

c = a * b

c = a * b

The result of “c” is 0.

The error will be fixed when b is converted to float32.

Numpy Precision

value = 0.0205862828046875
print("%.16f" % value)
value = value.astype(np.float32)
print("%.16f" % value)

value = 0.0205862828046875

print("%.16f" % value)

value = value.astype(np.float32)

print("%.16f" % value)

#output
0.0205862828046875
0.0205862820148468

#output

0.0205862828046875

0.0205862820148468

Precision changed!