1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
|
"""
This global flag controls whether to assign new tensors to the parameters
instead of changing the existing parameters in-place when converting an `nn.Module`
using the following methods:
1. `module.cuda()` / `.cpu()` (for moving `module` between devices)
2. `module.float()` / `.double()` / `.half()` (for converting `module` to a different dtype)
3. `module.to()` / `.type()` (for changing `module`'s device or dtype)
4. `module._apply(fn)` (for generic functions applied to `module`)
Default: False
"""
_overwrite_module_params_on_conversion = False
def set_overwrite_module_params_on_conversion(value):
global _overwrite_module_params_on_conversion
_overwrite_module_params_on_conversion = value
def get_overwrite_module_params_on_conversion():
return _overwrite_module_params_on_conversion
|
