| import torch, math |
| def find_sd(c, prefer='act_net_state_dict'): |
| if isinstance(c, dict): |
| if prefer in c and isinstance(c[prefer], dict): return c[prefer] |
| for k in ['act_net_state_dict','model_state_dict','state_dict','model','policy','nets']: |
| if k in c and isinstance(c[k], dict): return c[k] |
| ten = {k:v for k,v in c.items() if torch.is_tensor(v)} |
| if ten: return ten |
| return c |
| BC='/data0/hongyi/act_runs/open_laptop_100demos_chunk25/policy_epoch_4000_seed_0.ckpt' |
| G='/data0/hongyi/act_sim_ppo_v6_gtsim/run_openlaptop_grpo_bc_start/checkpoints' |
| bc=find_sd(torch.load(BC,map_location='cpu',weights_only=False)) |
| g64=find_sd(torch.load(f'{G}/iter_00064.pt',map_location='cpu',weights_only=False)) |
| g256=find_sd(torch.load(f'{G}/iter_00256.pt',map_location='cpu',weights_only=False)) |
| print('bc keys:',len(bc),'| g64 keys:',len(g64)) |
| common=set(bc)&set(g64) |
| print('common keys:',len(common)) |
| if len(common) < 0.4*len(g64): |
| print(' bc sample:',list(bc)[:3]); print(' g64 sample:',list(g64)[:3]) |
| |
| import os |
| def cmp(a,b,label): |
| bb_w=bb_d=tr_w=tr_d=0.0; nbb=ntr=0; biggest=[] |
| for k in common: |
| if not torch.is_tensor(a[k]) or not a[k].is_floating_point(): continue |
| nw=a[k].float().norm().item(); d=(b[k].float()-a[k].float()).norm().item() |
| rel=d/(nw+1e-9) |
| if 'backbone' in k.lower(): bb_w+=nw*nw; bb_d+=d*d; nbb+=1 |
| else: |
| tr_w+=nw*nw; tr_d+=d*d; ntr+=1; biggest.append((rel,k,nw,d)) |
| print(f' {label}') |
| print(f' frozen backbone: |Δ|={math.sqrt(bb_d):8.4f} |W|={math.sqrt(bb_w):8.1f} rel={math.sqrt(bb_d/(bb_w+1e-9)):.5f} (n={nbb})') |
| print(f' transf+heads: |Δ|={math.sqrt(tr_d):8.4f} |W|={math.sqrt(tr_w):8.1f} rel={math.sqrt(tr_d/(tr_w+1e-9)):.5f} (n={ntr})') |
| biggest.sort(reverse=True) |
| for rel,k,nw,d in biggest[:4]: print(f' top-Δ {k:40s} rel={rel:.3f} |W|={nw:.2f} |Δ|={d:.3f}') |
| cmp(bc,g64,'BC -> ep64 (peak 85%)') |
| cmp(bc,g256,'BC -> ep256 (degraded 70%)') |
|
|
