Add XNNPACK int8 export
Browse files- README.md +14 -1
- xnnpack/config.json +42 -0
- xnnpack/rfdetr_keypoint_preview_xnnpack_int8.pte +3 -0
README.md
CHANGED
|
@@ -17,7 +17,17 @@ Upstream model: [RFDetr](https://github.com/roboflow/rf-detr)
|
|
| 17 |
| Path | Backend | Precision |
|
| 18 |
| --- | --- | --- |
|
| 19 |
| `coreml/rfdetr_keypoint_preview_coreml_fp16.pte` | coreml | fp16 |
|
|
|
|
| 20 |
| `xnnpack/rfdetr_keypoint_preview_xnnpack_fp32.pte` | xnnpack | fp32 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
|
| 22 |
## Unshipped backends
|
| 23 |
|
|
@@ -33,8 +43,11 @@ config.json 37 B
|
|
| 33 |
coreml/config.json 1.1 kB
|
| 34 |
coreml/rfdetr_keypoint_preview_coreml_fp16.pte 71.5 MB
|
| 35 |
mlx/NOTES.md 1.4 kB
|
| 36 |
-
|
|
|
|
|
|
|
| 37 |
xnnpack/rfdetr_keypoint_preview_xnnpack_fp32.pte 139 MB
|
|
|
|
| 38 |
```
|
| 39 |
|
| 40 |
## Compatibility
|
|
|
|
| 17 |
| Path | Backend | Precision |
|
| 18 |
| --- | --- | --- |
|
| 19 |
| `coreml/rfdetr_keypoint_preview_coreml_fp16.pte` | coreml | fp16 |
|
| 20 |
+
| `vulkan/rfdetr_keypoint_preview_vulkan_fp16.pte` | vulkan | fp16 |
|
| 21 |
| `xnnpack/rfdetr_keypoint_preview_xnnpack_fp32.pte` | xnnpack | fp32 |
|
| 22 |
+
| `xnnpack/rfdetr_keypoint_preview_xnnpack_int8.pte` | xnnpack | int8 |
|
| 23 |
+
|
| 24 |
+
## XNNPACK int8
|
| 25 |
+
|
| 26 |
+
`xnnpack/rfdetr_keypoint_preview_xnnpack_int8.pte` is 8da8w (int8 dynamic
|
| 27 |
+
activations, int8 per-channel weights) on the Linear layers, except the
|
| 28 |
+
deformable-attention offsets and the class, box and keypoint heads; convolutions stay
|
| 29 |
+
fp32. COCO keypoint (OKS) mAP50-95 on the first 500 val2017 images: fp32
|
| 30 |
+
70.37, int8 70.00.
|
| 31 |
|
| 32 |
## Unshipped backends
|
| 33 |
|
|
|
|
| 43 |
coreml/config.json 1.1 kB
|
| 44 |
coreml/rfdetr_keypoint_preview_coreml_fp16.pte 71.5 MB
|
| 45 |
mlx/NOTES.md 1.4 kB
|
| 46 |
+
vulkan/config.json 1.1 kB
|
| 47 |
+
vulkan/rfdetr_keypoint_preview_vulkan_fp16.pte 70.0 MB
|
| 48 |
+
xnnpack/config.json 1.9 kB
|
| 49 |
xnnpack/rfdetr_keypoint_preview_xnnpack_fp32.pte 139 MB
|
| 50 |
+
xnnpack/rfdetr_keypoint_preview_xnnpack_int8.pte 56.6 MB
|
| 51 |
```
|
| 52 |
|
| 53 |
## Compatibility
|
xnnpack/config.json
CHANGED
|
@@ -49,6 +49,48 @@
|
|
| 49 |
]
|
| 50 |
}
|
| 51 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 52 |
}
|
| 53 |
]
|
| 54 |
}
|
|
|
|
| 49 |
]
|
| 50 |
}
|
| 51 |
}
|
| 52 |
+
},
|
| 53 |
+
{
|
| 54 |
+
"file": "rfdetr_keypoint_preview_xnnpack_int8.pte",
|
| 55 |
+
"precision": "int8",
|
| 56 |
+
"methods": {
|
| 57 |
+
"forward": {
|
| 58 |
+
"inputs": [
|
| 59 |
+
{
|
| 60 |
+
"shape": [
|
| 61 |
+
1,
|
| 62 |
+
3,
|
| 63 |
+
576,
|
| 64 |
+
576
|
| 65 |
+
],
|
| 66 |
+
"dtype": "float32"
|
| 67 |
+
}
|
| 68 |
+
],
|
| 69 |
+
"outputs": [
|
| 70 |
+
{
|
| 71 |
+
"shape": [
|
| 72 |
+
100,
|
| 73 |
+
4
|
| 74 |
+
],
|
| 75 |
+
"dtype": "float32"
|
| 76 |
+
},
|
| 77 |
+
{
|
| 78 |
+
"shape": [
|
| 79 |
+
100
|
| 80 |
+
],
|
| 81 |
+
"dtype": "float32"
|
| 82 |
+
},
|
| 83 |
+
{
|
| 84 |
+
"shape": [
|
| 85 |
+
100,
|
| 86 |
+
17,
|
| 87 |
+
3
|
| 88 |
+
],
|
| 89 |
+
"dtype": "float32"
|
| 90 |
+
}
|
| 91 |
+
]
|
| 92 |
+
}
|
| 93 |
+
}
|
| 94 |
}
|
| 95 |
]
|
| 96 |
}
|
xnnpack/rfdetr_keypoint_preview_xnnpack_int8.pte
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d66d6d57709904eb52ed088689fc932ba55e268a9f2e0fe24829c9a989a36b71
|
| 3 |
+
size 59381136
|