How to turn parameter to buffer?

1

In the context of having a trained model. Is there a way to turn a parameter into a buffer to keep it fixed, as an alternative to freezing that parameter?

2

Seems this can be done like this:

for name, param in module.named_parameters(recurse=False):
    delattr(module, name) # Unregister parameter
    module.register_buffer(name, param)

recurse must be set to false, because in case there is a submodule with its own parameters, doing this would fail. Instead, this should be repeated for all submodules.

Edit:
Here is a function that does this:

def param_to_buffer(module):
    """Turns all parameters of a module into buffers."""
    modules = module.modules()
    module = next(modules)
    for name, param in module.named_parameters(recurse=False):
        delattr(module, name) # Unregister parameter
        module.register_buffer(name, param)
    for module in modules:
        param_to_buffer(module)