mlboydaisuke/D-FINE-S-ExecuTorch

dfine_s_coco — ExecuTorch XNNPACK

• Source: ustc-community/dfine-small-coco
• License: Apache-2.0
• Input: [[1, 3, 640, 640]] — RGB/255 only (no mean/std norm), 640x640
• Output: logits [1,300,80] (sigmoid -> per-class score), boxes [1,300,4] cxcywh normalized 0..1; postprocess = sigmoid + top-k, NO NMS

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	dfine_s_coco_xnnpack_fp32.pte	41.5	1.000000	54.0

*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: 138.0 ms).

Precisions that did not earn a slot

• fp16 is not shipped: worst-output corr 0.223 against fp32 eager, below the 0.995 bar for this precision. The file converts and runs; the numbers do not hold up, so it is left out rather than shipped with a warning.

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, 300, 80]	1.420e-03	1.000000
1	[1, 300, 4]	4.858e-05	1.000000

XNNPACK delegate coverage (fp32): 77.5% (1391/1794 ops); ops left on the portable kernels: dim_order_ops._to_dim_order_copy.default x64, aten.select_copy.int x59, aten.expand_copy.default x36, aten.pow.Tensor_Scalar x29, aten.arange.start_step x21, aten.view_copy.default x18, aten.where.self x17, aten.eq.Scalar x16, aten.unsqueeze_copy.default x15, aten.split_with_sizes_copy.default x13, aten.native_layer_norm.default x12, aten.squeeze_copy.dims x12, aten.logical_not.default x10, aten.alias_copy.default x10, aten.grid_sampler_2d.default x9, aten.mul.Scalar x8, aten.full_like.default x8, aten.copy.default x6, aten.sum.dim_IntList x6, aten.any.dim x5, aten.mul.Tensor x4, aten.cat.default x3, dim_order_ops._clone_dim_order.default x2, aten.sin.default x2, aten.cos.default x2, aten.upsample_nearest2d.vec x2, aten.topk.default x2, aten.repeat.default x2, aten.gather.default x2, aten.full.default x1, aten.div.Tensor x1, aten.pow.Scalar x1, aten.reciprocal.default x1, aten.gt.Scalar x1, aten.lt.Scalar x1, aten.max.dim x1, aten.mean.dim x1

Conversion

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