DeepTFUS: variant D (soft-argmax extreme + anchor)

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

Fine-tune of masonwang025/deeptfus-base that triples all the soft-argmax knobs vs C: 3× focal weight, sharpest temperature, 3× peak LR. Tests how far position-aware fine-tuning can push focal_position_error_mm before other metrics break.

⭐ Project's best focal_mm (−30% median, ~64% of the gap to paper closed), but the only variant to break the rel_l2 paper budget and the worst max_p of any variant.

Modification (vs base)

loss.focal_weight       0    to 1.5e-4   (3x C, 15x A)
loss.focal_temperature  :    to 0.02     (sharpest tested)
loss.grad_weight        0.1  to 0.1      (anchor kept)
train.lr                1e-3 to 1e-4     (1/10 of base, 3x C; addresses LR cratering)

3-epoch warmup ramp on focal_weight. Fine-tune ran 10 epochs from base ckpt; shipped ckpt is ckpt_epoch_006.pt (best val_focal_mm).

Test results (n = 597)

metric	paper	base	D (this model)	Δ vs base
`relative_l2` mean ± std	0.414 ± 0.086	0.384 ± 0.078	0.422 ± 0.082 ⚠️	+0.038 (above paper mean)
`relative_l2` median	0.394	0.369	0.404	+0.035
`focal_position_error_mm` mean ± std	2.89 ± 2.14	6.49 ± 4.58	4.19 ± 2.93 ★	−2.30 mm (−35%) project best
`focal_position_error_mm` median	2.45	5.15	3.61 ★	−1.54 mm (−30%) project best
`max_pressure_error` mean ± std	0.199 ± 0.158	0.225 ± 0.116	0.283 ± 0.105 ⚠️	+0.058 (worse)
`max_pressure_error` median	0.166	0.217	0.287	+0.070
`focal_pressure_error` median	:	0.528	0.475	−0.053
`focal_iou_fwhm` median	:	0.143	0.121	−0.022
`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-d-softargmax-extreme", "ckpt_best.pt"),
    map_location="cpu", weights_only=False,
)

Model code: github.com/masonwang025/deeptfus.