mlboydaisuke/U2Net-ExecuTorch

u2net — ExecuTorch XNNPACK

• Source: xuebinqin/U-2-Net + Carve/u2net-universal weights
• License: Apache-2.0
• Input: [[1, 3, 320, 320]] — RGB, ImageNet norm, 320x320
• Output: saliency mask [1,1,320,320] 0-1 (sigmoid); min-max normalize then resize to the source image

Variants

All variants take and return fp32 tensors — swap the .pte file, keep your app code.

precision	file	size (MB)	parity vs fp32 eager (worst corr)	Mac median (ms)*
fp32	u2net_xnnpack_fp32.pte	176.0	1.000000	56.5

*Mac arm64, single process, median of 10 — a reference point for relative cost only, not a device number (torch eager fp32 on the same machine: 138.5 ms).

Precisions that did not earn a slot

• fp16 is not shipped: it comes out at 100% of the fp32 file (176.0 MB vs 176.0 MB), so it buys nothing. XNNPACK serializes convolution weights as fp32 no matter what dtype the graph carries, so on a conv-heavy model fp16 saves no disk and only adds cast operations. Reach for int8 here, not fp16.
• int8 is not shipped: mask IoU 0.814 median and 0.212 at worst against fp32. The int8 build systematically shrinks weak saliency: on a photo where fp32 marks 2.2% of pixels foreground it marks 0.5%. On images with an unambiguous subject it agrees closely (0.99), but that is not a promise this model can make.

Verification (executorch 1.4.0, torch 2.13.0)

Parity is measured against the fp32 eager model on real image input; corr is the correlation over all elements of each output tensor.

output	shape	max_abs_diff	corr
0	[1, 1, 320, 320]	5.239e-05	1.000000

XNNPACK delegate coverage (fp32): 100.0% (477/477 ops)

Conversion

torch.export -> to_edge_transform_and_lower(XnnpackPartitioner) -> .pte (conversion scripts: executorch-models)