DeepTFUS: variant A (soft-argmax mild fine-tune)

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

Fine-tune of masonwang025/deeptfus-base that adds a soft-argmax focal-position L1 term to the loss at gentle weight, to test whether any position-aware aux signal closes the focal_position_error_mm gap that the base reproduction left open.

⭐ Conservative recipe: the only fine-tune variant where max_p improves over baseline, but focal_mm plateaus at a modest 10% median gain.

Modification (vs base)

Single new loss term added on top of the paper recipe:

L_focal = || soft_argmax(P̂_norm, τ=0.05) − argmax(P_gt_norm) ||_1
total   = L_paper + 1e-5 · L_focal

3-epoch linear warmup ramp from 0 to 1e-5. Paper's gradient_L1=0.1 anchor kept on. Fine-tune ran 12 epochs from base ckpt at lr=3e-5 (1/30 of base's peak); shipped checkpoint is ckpt_epoch_007.pt (best val_focal_mm in plateau).

Test results (n = 597)

metric	paper	base	A (this model)	Δ vs base
`relative_l2` median	0.394	0.369	0.372	+0.003 (within budget)
`relative_l2` mean	0.414	0.384	0.389	+0.005
`focal_position_error_mm` median	2.45	5.15	4.64	−0.51 mm (−10%)
`focal_position_error_mm` mean	2.89	6.49	5.60	−0.89 mm
`max_pressure_error` median	0.166	0.217	0.200	−0.017 ✅
`max_pressure_error` mean	0.199	0.225	0.204	−0.021
`focal_pressure_error` median	:	0.528	0.487	−0.041
`focal_iou_fwhm` median	:	0.143	0.148	+0.004
`inference_latency_s` median	:	0.233	0.232	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-a-softargmax-mild", "ckpt_best.pt"),
    map_location="cpu", weights_only=False,
)

Model code: github.com/masonwang025/deeptfus.