How to initialize weights in PyTorch?

Question

How to initialize the weights and biases (for example, with He or Xavier initialization) in a network in PyTorch?

Answer 1

If you want some extra flexibility, you can also set the weights manually.

Say you have input of all ones:

import torch
import torch.nn as nn

input = torch.ones((8, 8))
print(input)

tensor([[1., 1., 1., 1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1., 1., 1., 1.],
        [1., 1., 1., 1., 1., 1., 1., 1.]])

And you want to make a dense layer with no bias (so we can visualize):

d = nn.Linear(8, 8, bias=False)

Set all the weights to 0.5 (or anything else):

d.weight.data = torch.full((8, 8), 0.5)
print(d.weight.data)

The weights:

Out[14]: 
tensor([[0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000],
        [0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000],
        [0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000],
        [0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000],
        [0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000],
        [0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000],
        [0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000],
        [0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000, 0.5000]])

All your weights are now 0.5. Pass the data through:

d(input)

Out[13]: 
tensor([[4., 4., 4., 4., 4., 4., 4., 4.],
        [4., 4., 4., 4., 4., 4., 4., 4.],
        [4., 4., 4., 4., 4., 4., 4., 4.],
        [4., 4., 4., 4., 4., 4., 4., 4.],
        [4., 4., 4., 4., 4., 4., 4., 4.],
        [4., 4., 4., 4., 4., 4., 4., 4.],
        [4., 4., 4., 4., 4., 4., 4., 4.],
        [4., 4., 4., 4., 4., 4., 4., 4.]], grad_fn=)

Remember that each neuron receives 8 inputs, all of which have weight 0.5 and value of 1 (and no bias), so it sums up to 4 for each.