Applying RPU_config after loading a HWA-trained model (exmaple 24)

[kusionlion]

Hi,

I tried applying RPU_config after loading a HWA-trained model (exmaple 24) and found that the RPU_config doesn't seem to be used to represent my weights. The reason for this this doubt is that when I scaled the noise in RPU by 0 to 5000 times, the inference results (F1) stay the same.

Here is the steps I did (example 24):
1. performed HWA training and saved the trained model:
ARGS.train_hwa=True
code link
PARSER.add_argument(
"-c",
"--checkpoint",
help="File name specifying where to load/save a checkpoint",
default="./New_trained_results.pth",
type=str,
)

2. Loaded new weights and changed the programming scale
ARGS.load=True
code link
noise_model=PCMLikeNoiseModel(prog_noise_scale=5000),

The step 2 generates many F1 over time. I varied the prog_noise_scale, and the F1 seems to be independent to the prog_noise_scale. Am I doing it wrong?

Thanks in advance.

[maljoras-sony]

By default, the rpu_config from the checkpoint is loaded. It is not clear from the above, how you changed the rpu_config. You could use model.drift_analog_weights(noise_model=the_new_noise_model) that would apply the new noise model without the need to change the rpu_config of the model.

[PabloCarmona]

Hello @kusionlion! The answer from @maljoras-sony help you out? Need any further assistance? Thanks!