PyTorch how to generate sample normal distribution? - pytorch

Question

Please may someone review this code, I cannot seem to use this as the fit
function does not exist?
import torch
import torch.distributions as dist
print(torch.cuda.is_available())
# Sample data
data = torch.randn(1000)
# Fit the data to a log-normal distribution
mu, std = dist.Normal(0, 1).fit(torch.log(data))
# Create a log-normal distribution with the fitted parameters
lognormal = dist.LogNormal(mu, std)
# Get the probability density function (PDF) of the log-normal distribution
pdf = lognormal.log_prob(torch.log(data)).exp()
# Plot the data and the PDF
import matplotlib.pyplot as plt
plt.hist(data, bins=30, density=True)
plt.plot(data, pdf, 'r')
plt.show()
I receive the following error:
AttributeError: 'Normal' object has no attribute 'fit'

Answer 1

Neither the PyTorch Distribution base class nor the Normal distribution offer a fit() method (thus the error).
If you want to fit a normal distribution to given data, you could use the scipy.stats package.
import numpy as np
from scipy.stats import norm
import torch
data = torch.randn(1000)
# ... do other thinngs with the data
data = data.numpy() # torch tensor must be converted to numpy array first
mu, std = norm.fit(data)
# ...
Please also note: By using torch.randn() you generate random numbers from a normal distribution with mean 0 and variance 1.
This means approximately half the numbers are negative. This will cause problems if you take the logarithm of these values in the next step, since logarithms of negative numbers are not defined.
data = torch.randn(1000)
data = torch.log(data) # log(data) produces nan any time a value is less than zero
# --> tensor([ 6.4755e-01, 7.7064e-01, -3.5208e-01, 1.2873e-01, nan,
# nan, -6.5469e-01, 5.8180e-02, nan, nan, ...])
data = data.numpy()
mu, std = norm.fit(data) # this won't work now as the data contains non-finite values.