Univariate optimization with PyTorch

1

What’s the proper way to do univariate optimization with PyTorch?

Currently, what I do is something like the following which correctly finds the mean of given sample pts as 2.5.

import torch
from torch.autograd import Variable

mean = Variable(torch.ones(1, 1), requires_grad=True)
criterion = torch.nn.MSELoss()
optimizer = torch.optim.SGD([mean], lr=0.01, momentum=0.9)
pts = torch.Tensor([1, 2, 3, 4])

for _ in range(10**3):
    optimizer.zero_grad()
    loss = criterion(mean, pts)
    loss.backward()
    optimizer.step()
    print(mean.data)
2

Hi,

Pytorch is not really built to do univariate optimization.
But your approach looks like the right was to do this.

Note that Variable is not needed anymore and you can simply do mean = torch.ones(1, 1, requires_grad=True).
And replace print(mean.data) by print(mean) or print(mean.detach()).

1 Like