Commit ·
f7eff56
1
Parent(s): 5e4a805
QNN: DeepLabV3-MobileNetV3 a16w8, one context binary per Hexagon version
Browse files66.55% mIoU on VOC2012 val (200 images) against 66.56% for the XNNPACK int8
file, at 3.20 ms/inference on an S26 Ultra. Returns a class index per pixel,
(1,520,520) int32, instead of (1,21,520,520) logits.
- README.md +20 -0
- qnn/config.json +172 -0
- qnn/deeplab_v3_mobilenet_v3_large_qnn_a16w8_v69.pte +3 -0
- qnn/deeplab_v3_mobilenet_v3_large_qnn_a16w8_v73.pte +3 -0
- qnn/deeplab_v3_mobilenet_v3_large_qnn_a16w8_v75.pte +3 -0
- qnn/deeplab_v3_mobilenet_v3_large_qnn_a16w8_v79.pte +3 -0
- qnn/deeplab_v3_mobilenet_v3_large_qnn_a16w8_v81.pte +3 -0
README.md
CHANGED
|
@@ -25,6 +25,26 @@ Upstream model: [DeepLabV3](https://docs.pytorch.org/vision/main/models/deeplabv
|
|
| 25 |
| `xnnpack/deeplab_v3_resnet50_xnnpack_int8.pte` | xnnpack | int8 |
|
| 26 |
| `xnnpack/deeplab_v3_resnet101_xnnpack_fp32.pte` | xnnpack | fp32 |
|
| 27 |
| `xnnpack/deeplab_v3_resnet101_xnnpack_int8.pte` | xnnpack | int8 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
|
| 29 |
## Unshipped backends
|
| 30 |
|
|
|
|
| 25 |
| `xnnpack/deeplab_v3_resnet50_xnnpack_int8.pte` | xnnpack | int8 |
|
| 26 |
| `xnnpack/deeplab_v3_resnet101_xnnpack_fp32.pte` | xnnpack | fp32 |
|
| 27 |
| `xnnpack/deeplab_v3_resnet101_xnnpack_int8.pte` | xnnpack | int8 |
|
| 28 |
+
| `qnn/deeplab_v3_mobilenet_v3_large_qnn_a16w8_v69.pte` | qnn | a16w8 |
|
| 29 |
+
| `qnn/deeplab_v3_mobilenet_v3_large_qnn_a16w8_v73.pte` | qnn | a16w8 |
|
| 30 |
+
| `qnn/deeplab_v3_mobilenet_v3_large_qnn_a16w8_v75.pte` | qnn | a16w8 |
|
| 31 |
+
| `qnn/deeplab_v3_mobilenet_v3_large_qnn_a16w8_v79.pte` | qnn | a16w8 |
|
| 32 |
+
| `qnn/deeplab_v3_mobilenet_v3_large_qnn_a16w8_v81.pte` | qnn | a16w8 |
|
| 33 |
+
|
| 34 |
+
## QNN variants
|
| 35 |
+
|
| 36 |
+
A QNN `.pte` embeds an HTP context binary compiled for one Hexagon version, so
|
| 37 |
+
there is one file per version and the runtime picks the one matching the device.
|
| 38 |
+
These return a class index per pixel, `(1, 520, 520)` int32, rather than
|
| 39 |
+
`(1, 21, 520, 520)` logits.
|
| 40 |
+
|
| 41 |
+
| Hexagon | Example SoC |
|
| 42 |
+
| --- | --- |
|
| 43 |
+
| v69 | Snapdragon 8 Gen 1 |
|
| 44 |
+
| v73 | Snapdragon 8 Gen 2 |
|
| 45 |
+
| v75 | Snapdragon 8 Gen 3 |
|
| 46 |
+
| v79 | Snapdragon 8 Elite |
|
| 47 |
+
| v81 | Snapdragon 8 Elite Gen 5 |
|
| 48 |
|
| 49 |
## Unshipped backends
|
| 50 |
|
qnn/config.json
ADDED
|
@@ -0,0 +1,172 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"$schema": "https://huggingface.co/software-mansion/react-native-executorch-spec/resolve/main/config.schema.json",
|
| 3 |
+
"model": "deeplab_v3",
|
| 4 |
+
"family": "deeplab",
|
| 5 |
+
"capabilities": [
|
| 6 |
+
"semantic-segmentation"
|
| 7 |
+
],
|
| 8 |
+
"backend": "qnn",
|
| 9 |
+
"license": "bsd-3-clause",
|
| 10 |
+
"variants": [
|
| 11 |
+
{
|
| 12 |
+
"file": "deeplab_v3_mobilenet_v3_large_qnn_a16w8_v69.pte",
|
| 13 |
+
"precision": "a16w8",
|
| 14 |
+
"quantized": true,
|
| 15 |
+
"default": false,
|
| 16 |
+
"size": "mobilenet_v3_large",
|
| 17 |
+
"methods": {
|
| 18 |
+
"forward": {
|
| 19 |
+
"inputs": [
|
| 20 |
+
{
|
| 21 |
+
"shape": [
|
| 22 |
+
1,
|
| 23 |
+
3,
|
| 24 |
+
520,
|
| 25 |
+
520
|
| 26 |
+
],
|
| 27 |
+
"dtype": "float32"
|
| 28 |
+
}
|
| 29 |
+
],
|
| 30 |
+
"outputs": [
|
| 31 |
+
{
|
| 32 |
+
"shape": [
|
| 33 |
+
1,
|
| 34 |
+
520,
|
| 35 |
+
520
|
| 36 |
+
],
|
| 37 |
+
"dtype": "int32"
|
| 38 |
+
}
|
| 39 |
+
]
|
| 40 |
+
}
|
| 41 |
+
}
|
| 42 |
+
},
|
| 43 |
+
{
|
| 44 |
+
"file": "deeplab_v3_mobilenet_v3_large_qnn_a16w8_v73.pte",
|
| 45 |
+
"precision": "a16w8",
|
| 46 |
+
"quantized": true,
|
| 47 |
+
"default": false,
|
| 48 |
+
"size": "mobilenet_v3_large",
|
| 49 |
+
"methods": {
|
| 50 |
+
"forward": {
|
| 51 |
+
"inputs": [
|
| 52 |
+
{
|
| 53 |
+
"shape": [
|
| 54 |
+
1,
|
| 55 |
+
3,
|
| 56 |
+
520,
|
| 57 |
+
520
|
| 58 |
+
],
|
| 59 |
+
"dtype": "float32"
|
| 60 |
+
}
|
| 61 |
+
],
|
| 62 |
+
"outputs": [
|
| 63 |
+
{
|
| 64 |
+
"shape": [
|
| 65 |
+
1,
|
| 66 |
+
520,
|
| 67 |
+
520
|
| 68 |
+
],
|
| 69 |
+
"dtype": "int32"
|
| 70 |
+
}
|
| 71 |
+
]
|
| 72 |
+
}
|
| 73 |
+
}
|
| 74 |
+
},
|
| 75 |
+
{
|
| 76 |
+
"file": "deeplab_v3_mobilenet_v3_large_qnn_a16w8_v75.pte",
|
| 77 |
+
"precision": "a16w8",
|
| 78 |
+
"quantized": true,
|
| 79 |
+
"default": false,
|
| 80 |
+
"size": "mobilenet_v3_large",
|
| 81 |
+
"methods": {
|
| 82 |
+
"forward": {
|
| 83 |
+
"inputs": [
|
| 84 |
+
{
|
| 85 |
+
"shape": [
|
| 86 |
+
1,
|
| 87 |
+
3,
|
| 88 |
+
520,
|
| 89 |
+
520
|
| 90 |
+
],
|
| 91 |
+
"dtype": "float32"
|
| 92 |
+
}
|
| 93 |
+
],
|
| 94 |
+
"outputs": [
|
| 95 |
+
{
|
| 96 |
+
"shape": [
|
| 97 |
+
1,
|
| 98 |
+
520,
|
| 99 |
+
520
|
| 100 |
+
],
|
| 101 |
+
"dtype": "int32"
|
| 102 |
+
}
|
| 103 |
+
]
|
| 104 |
+
}
|
| 105 |
+
}
|
| 106 |
+
},
|
| 107 |
+
{
|
| 108 |
+
"file": "deeplab_v3_mobilenet_v3_large_qnn_a16w8_v79.pte",
|
| 109 |
+
"precision": "a16w8",
|
| 110 |
+
"quantized": true,
|
| 111 |
+
"default": false,
|
| 112 |
+
"size": "mobilenet_v3_large",
|
| 113 |
+
"methods": {
|
| 114 |
+
"forward": {
|
| 115 |
+
"inputs": [
|
| 116 |
+
{
|
| 117 |
+
"shape": [
|
| 118 |
+
1,
|
| 119 |
+
3,
|
| 120 |
+
520,
|
| 121 |
+
520
|
| 122 |
+
],
|
| 123 |
+
"dtype": "float32"
|
| 124 |
+
}
|
| 125 |
+
],
|
| 126 |
+
"outputs": [
|
| 127 |
+
{
|
| 128 |
+
"shape": [
|
| 129 |
+
1,
|
| 130 |
+
520,
|
| 131 |
+
520
|
| 132 |
+
],
|
| 133 |
+
"dtype": "int32"
|
| 134 |
+
}
|
| 135 |
+
]
|
| 136 |
+
}
|
| 137 |
+
}
|
| 138 |
+
},
|
| 139 |
+
{
|
| 140 |
+
"file": "deeplab_v3_mobilenet_v3_large_qnn_a16w8_v81.pte",
|
| 141 |
+
"precision": "a16w8",
|
| 142 |
+
"quantized": true,
|
| 143 |
+
"default": true,
|
| 144 |
+
"size": "mobilenet_v3_large",
|
| 145 |
+
"methods": {
|
| 146 |
+
"forward": {
|
| 147 |
+
"inputs": [
|
| 148 |
+
{
|
| 149 |
+
"shape": [
|
| 150 |
+
1,
|
| 151 |
+
3,
|
| 152 |
+
520,
|
| 153 |
+
520
|
| 154 |
+
],
|
| 155 |
+
"dtype": "float32"
|
| 156 |
+
}
|
| 157 |
+
],
|
| 158 |
+
"outputs": [
|
| 159 |
+
{
|
| 160 |
+
"shape": [
|
| 161 |
+
1,
|
| 162 |
+
520,
|
| 163 |
+
520
|
| 164 |
+
],
|
| 165 |
+
"dtype": "int32"
|
| 166 |
+
}
|
| 167 |
+
]
|
| 168 |
+
}
|
| 169 |
+
}
|
| 170 |
+
}
|
| 171 |
+
]
|
| 172 |
+
}
|
qnn/deeplab_v3_mobilenet_v3_large_qnn_a16w8_v69.pte
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:30588189e6b35a229f651e237cb38a4a51a6413ec005abf99f3c72169e290d06
|
| 3 |
+
size 15001088
|
qnn/deeplab_v3_mobilenet_v3_large_qnn_a16w8_v73.pte
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ba3e3b923bcf0d9f2dc2c6de1c5da9eeae7677db5b42596097f1118466b86d27
|
| 3 |
+
size 12760576
|
qnn/deeplab_v3_mobilenet_v3_large_qnn_a16w8_v75.pte
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:4e4dff48666de9513194a49420d69a849e08a8f65408b07e2f991b2471ee5c54
|
| 3 |
+
size 12764672
|
qnn/deeplab_v3_mobilenet_v3_large_qnn_a16w8_v79.pte
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:dd0a2d84825370126f6634f70b36a5644b4652384a0718fd9d4a3ddbf947a1ac
|
| 3 |
+
size 12760576
|
qnn/deeplab_v3_mobilenet_v3_large_qnn_a16w8_v81.pte
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:36221ea5e11a9581d1b308fabc92a3bcaa3a2ffeeb6eadc00ce6d2587a416e46
|
| 3 |
+
size 12805632
|