| |
| """Error attribution ViT vs MANO head on the 32 acc.py crops: the device ViT feature map fed to the CPU fp32 |
| head, and the device full model, both vs the CPU fp32 reference; prints the signed camera-scale error of each.""" |
| import os, sys, statistics |
| from pathlib import Path |
| sys.path.insert(0, str(Path(__file__).resolve().parents[1])) |
| sys.path.insert(0, str(Path(__file__).resolve().parent)) |
| os.environ.setdefault("TT_FUSED_HASH", "0") |
| import torch, ttnn |
| from hamer import Hamer, build_from_reference, load_hamer_checkpoint |
| from hamer import vit as tv |
| from acc import crops |
|
|
| hf = Path.home() / ".cache/huggingface/hub/models--changh95--hamer-weights/snapshots/e8217b46d54b4e8afaaf0dd6efb6b62341c97d32" |
| torch.manual_seed(0) |
| ref = Hamer().eval() |
| load_hamer_checkpoint(ref, hf / "hamer.ckpt", hf / "mano_mean_params.npz") |
| xs = crops() |
| tt = build_from_reference(ref, device_id=0, strict=True) |
| rows = [] |
| with torch.no_grad(): |
| for n, x in xs: |
| fr = ref.backbone(x) |
| o_ref = ref.head(fr) |
| tok = tt._upload_tokens(x) |
| f = tv.forward(tok, tt._vit_params, fused=True) |
| ft = ttnn.to_torch(f).float().reshape(1, 16, 12, 1280).permute(0, 3, 1, 2) |
| ttnn.deallocate(f) |
| o_vit = ref.head(ft) |
| o_tt = tt(x) |
| fe = (ft - fr) |
| rows.append((float(o_vit[0, 154] - o_ref[0, 154]), float(o_tt[0, 154] - o_ref[0, 154]), |
| float((fe * fr.sign()).mean() / fr.abs().mean()))) |
| tt.close() |
| for i, name in enumerate(("devViT+cpuHead", "device full", "feat mag-bias")): |
| v = [r[i] for r in rows] |
| print(f"{name:16s} mean {statistics.mean(v):+.5f} pos {sum(t > 0 for t in v)}/{len(v)}") |
|
|
