Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- executorch
|
| 5 |
+
- xnnpack
|
| 6 |
+
- pte
|
| 7 |
+
- on-device
|
| 8 |
+
- object-detection
|
| 9 |
+
base_model:
|
| 10 |
+
- PekingU/rtdetr_v2_r18vd
|
| 11 |
+
---
|
| 12 |
+
# rtdetrv2_s_r18vd — ExecuTorch XNNPACK
|
| 13 |
+
|
| 14 |
+
`rtdetrv2_s_r18vd_xnnpack_fp32.pte` (80.9 MB, fp32, XNNPACK-delegated)
|
| 15 |
+
|
| 16 |
+
- **Source**: PekingU/rtdetr_v2_r18vd
|
| 17 |
+
- **License**: Apache-2.0
|
| 18 |
+
- **Input**: [[1, 3, 640, 640]] — RGB/255 only (no mean/std norm), 640x640
|
| 19 |
+
- **Output**: logits [1,300,80] (sigmoid -> per-class score), boxes [1,300,4] cxcywh normalized 0..1; postprocess = sigmoid + top-k, NO NMS
|
| 20 |
+
|
| 21 |
+
## Verification (Mac arm64, executorch 1.4.0, torch 2.13.0)
|
| 22 |
+
|
| 23 |
+
Parity vs torch fp32 eager on random input:
|
| 24 |
+
|
| 25 |
+
| output | shape | max_abs_diff | corr |
|
| 26 |
+
|--------|-------|--------------|------|
|
| 27 |
+
| 0 | [1, 300, 80] | 3.958e-05 | 1.000000 |
|
| 28 |
+
| 1 | [1, 300, 4] | 1.138e-05 | 1.000000 |
|
| 29 |
+
|
| 30 |
+
Median latency over 10 runs (single Mac process, reference only — device numbers to follow):
|
| 31 |
+
ExecuTorch 55.5 ms vs torch eager 64.8 ms.
|
| 32 |
+
|
| 33 |
+
## Conversion
|
| 34 |
+
|
| 35 |
+
torch.export -> to_edge_transform_and_lower(XnnpackPartitioner) -> .pte
|
| 36 |
+
(conversion scripts: `convert/` in the conversion repo)
|