Download code/src/robottt_pi05/parameter_statistics.py from jisue/robottt-pi05-base-comparison: direct link, hf CLI and curl.
- Browser
- Download file 607 Bytes
-
https://huggingface.co/jisue/robottt-pi05-base-comparison/resolve/main/code/src/robottt_pi05/parameter_statistics.py
- Command line
-
hf download hf://jisue/robottt-pi05-base-comparison/code/src/robottt_pi05/parameter_statistics.py
-
curl -L -o parameter_statistics.py https://huggingface.co/jisue/robottt-pi05-base-comparison/resolve/main/code/src/robottt_pi05/parameter_statistics.py
607 Bytes
| """Scalar records for vector gates; no model mutation or loss change.""" | |
| import torch | |
| def module_statistics(modules): | |
| with torch.no_grad(): | |
| return {'gates':{k:float(m.alpha.detach().double().mean()) for k,m in modules.items()}, | |
| 'gate_max_abs':{k:float(m.alpha.detach().abs().max()) for k,m in modules.items()}, | |
| 'gate_norm':{k:float(m.alpha.detach().double().norm()) for k,m in modules.items()}, | |
| 'W0_norm':{k:sum(float(w.detach().double().square().sum()) for w in m.w0)**.5 for k,m in modules.items()}, | |
| 'inner_lr':{k:float(m.inner_lr) for k,m in modules.items()}} | |