DeepTFUS: variant C (soft-argmax cranked + anchor restored)

A reproduction attempt of DeepTFUS, proposed by Srivastav et al. (arXiv:2505.12998).

Fine-tune of masonwang025/deeptfus-base that's a single-knob counterfactual to variant B: same aggressive soft-argmax recipe, but with the paper's gradient_L1=0.1 anchor restored. Tests whether the anchor matters at this aggression level.

⭐ Statistically tied with B on the three paper headline metrics; the anchor exists to suppress secondary hot-spots (median secondary-lobe volume drops 99% vs B), not to move focal_mm.

Modification (vs base)

loss.focal_weight       0   to 5e-5     (+soft-argmax term)
loss.focal_temperature  :   to 0.03     (sharper than A's 0.05)
loss.grad_weight        0.1 to 0.1      (anchor KEPT, single knob diff vs B)

3-epoch warmup ramp on focal_weight. Fine-tune ran 10 epochs from base ckpt at lr=3e-5; shipped ckpt is ckpt_epoch_008.pt (best val_focal_mm in plateau).

Test results (n = 597)

metric	paper	base	C (this model)	Δ vs base	Δ vs B
`relative_l2` mean ± std	0.414 ± 0.086	0.384 ± 0.078	0.388 ± 0.077	+0.005 (in budget)	0.000 (tied)
`relative_l2` median	0.394	0.369	0.372	+0.003	0.000
`focal_position_error_mm` mean ± std	2.89 ± 2.14	6.49 ± 4.58	5.11 ± 3.76	−1.38 mm	+0.05 (≈ tied)
`focal_position_error_mm` median	2.45	5.15	4.15	−1.00 mm (−19%)	−0.03 (≈ tied)
`max_pressure_error` mean ± std	0.199 ± 0.158	0.225 ± 0.116	0.239 ± 0.106	+0.014	0.000
`max_pressure_error` median	0.166	0.217	0.239	+0.022	0.000
`focal_pressure_error` median	:	0.528	0.496	−0.032	−0.006
`focal_iou_fwhm` median	:	0.143	0.136	−0.007	0.000
`inference_latency_s` median	:	0.233	0.232	unchanged	unchanged

Other variants and discussion

See the Collection for the other 5 variants, and the project page for the full reproduction story, interactive viewer, and discussion of trade-offs.

Usage

from huggingface_hub import hf_hub_download
import torch

ckpt = torch.load(
    hf_hub_download("masonwang025/deeptfus-ft-c-softargmax-anchored", "ckpt_best.pt"),
    map_location="cpu", weights_only=False,
)

Model code: github.com/masonwang025/deeptfus.