nn.Parameters vs nn.Module.register_parameter

1

According to the document, nn.Parameter will:

they are automatically added to the list of its parameters, and will appear e.g. in parameters() iterator

and nn.Module.register_parameter will

Adds a parameter to the module.

I wonder since nn.Parameter will add tensor into parameters automatically, why we need register_parameter function?

7 Likes
2

I guess it is because there are some tasks to add parameters to given module. Here is an example code. (Below code is just example, but I did this kind of thing when I implemented spectral normalization layer)

import torch
import torch.nn as nn

class Test(nn.Module):
    def __init__(self, module):
        super(Test, self).__init__()
        self.module = module
        self.register_param()
    
    def register_param():
        exist_w = hasattr(self.module, 'w')
        if not exist_w:
            w = nn.Parameter(torch.ones(1))
            self.module.register_parameter(w) # register 'w' to module

    def forward(self, x)
        return x

conv = nn.Conv2d(3, 3)
conv_w = Test(conv)
4 Likes
4

nn.Module.register_parameter takes the tensor or None but first checks if the name is in dictionary of the module. While nn.Parameter doesn’t have such check.

12 Likes
5

Hey, thank you for that answer. In this example, what type does the self.module have? Is it an nn.Module?