Compare equivalent FP32 head GEMM layouts 7eca78d68fa8fdd23b53e286d859d8d6e2f752dc
Browse files
experiments/README.md
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Equivalent FP32 head experiments
|
| 2 |
+
Keep RTMW3D-L trained weights, image resolution and all 133 3D joints while comparing explicit im2col plus MatMul against the original convolution kernels. The split8 candidate increases the batch of independent partial products and sums their FP32 results. Neither candidate is assumed faster before GPU measurement.
|
| 3 |
+
Upstream model and implementation: https://github.com/open-mmlab/mmpose
|
| 4 |
+
Only the two measured final head convolutions are rewritten. Original FP32 weights are reordered without quantization; all other model parameters are preserved. Both candidates pass small public-fixture CPU conversion checks, including independent hips, knees, ankles and foot gates. This is not target-video accuracy or native WebGPU validation.
|
| 5 |
+
Use the browser benchmark to measure synchronized inference, output drift and per-kernel GPU time on the target device.
|
experiments/catalog.json
ADDED
|
@@ -0,0 +1,602 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"schema": "rtmw3d-head-experiments/v1",
|
| 3 |
+
"sourceCommit": "7eca78d68fa8fdd23b53e286d859d8d6e2f752dc",
|
| 4 |
+
"keypointCount": 133,
|
| 5 |
+
"variants": [
|
| 6 |
+
{
|
| 7 |
+
"id": "rtmw3d-l-fp32-head-gemm-split8",
|
| 8 |
+
"model": "rtmw3d-l",
|
| 9 |
+
"policy": "fp32-head-gemm-split8",
|
| 10 |
+
"file": "experiments/rtmw3d-l-fp32-head-gemm-split8.onnx",
|
| 11 |
+
"sizeBytes": 230139419,
|
| 12 |
+
"sha256": "a2767b1c20120f8351852f1c87617794c6852ace4de76bf12c7c483fa8ab76ed",
|
| 13 |
+
"precision": "fp32",
|
| 14 |
+
"experimental": true,
|
| 15 |
+
"qualityPassed": true,
|
| 16 |
+
"browserGPUValidated": false,
|
| 17 |
+
"sourceRevision": "88ee5cb728cb2ed3efaf43011c9dd4908a6504b7",
|
| 18 |
+
"sourceModelSha256": "4e37e5e9043817c5ff387b61c44bb48af94ad42e4978c4e43780761c65582abc",
|
| 19 |
+
"upstream": "https://github.com/open-mmlab/mmpose",
|
| 20 |
+
"weightsUnchangedFP32": true,
|
| 21 |
+
"input": {
|
| 22 |
+
"name": "input",
|
| 23 |
+
"type": "float32",
|
| 24 |
+
"shape": [
|
| 25 |
+
1,
|
| 26 |
+
3,
|
| 27 |
+
384,
|
| 28 |
+
288
|
| 29 |
+
]
|
| 30 |
+
},
|
| 31 |
+
"outputs": [
|
| 32 |
+
{
|
| 33 |
+
"name": "keypoints",
|
| 34 |
+
"type": "float32",
|
| 35 |
+
"shape": [
|
| 36 |
+
1,
|
| 37 |
+
133,
|
| 38 |
+
3
|
| 39 |
+
],
|
| 40 |
+
"units": [
|
| 41 |
+
"crop-pixels",
|
| 42 |
+
"crop-pixels",
|
| 43 |
+
"root-relative-metres"
|
| 44 |
+
]
|
| 45 |
+
},
|
| 46 |
+
{
|
| 47 |
+
"name": "scores",
|
| 48 |
+
"type": "float32",
|
| 49 |
+
"shape": [
|
| 50 |
+
1,
|
| 51 |
+
133
|
| 52 |
+
]
|
| 53 |
+
}
|
| 54 |
+
],
|
| 55 |
+
"outputBytesPerFrame": 2128,
|
| 56 |
+
"changedConvolutions": [
|
| 57 |
+
{
|
| 58 |
+
"node": "/head/final_layer/conv/Conv",
|
| 59 |
+
"inputShape": [
|
| 60 |
+
1,
|
| 61 |
+
1024,
|
| 62 |
+
12,
|
| 63 |
+
9
|
| 64 |
+
],
|
| 65 |
+
"weightShape": [
|
| 66 |
+
133,
|
| 67 |
+
1024,
|
| 68 |
+
7,
|
| 69 |
+
7
|
| 70 |
+
],
|
| 71 |
+
"outputShape": [
|
| 72 |
+
1,
|
| 73 |
+
133,
|
| 74 |
+
12,
|
| 75 |
+
9
|
| 76 |
+
],
|
| 77 |
+
"kernel": [
|
| 78 |
+
7,
|
| 79 |
+
7
|
| 80 |
+
],
|
| 81 |
+
"pads": [
|
| 82 |
+
3,
|
| 83 |
+
3,
|
| 84 |
+
3,
|
| 85 |
+
3
|
| 86 |
+
],
|
| 87 |
+
"strides": [
|
| 88 |
+
1,
|
| 89 |
+
1
|
| 90 |
+
],
|
| 91 |
+
"dilations": [
|
| 92 |
+
1,
|
| 93 |
+
1
|
| 94 |
+
],
|
| 95 |
+
"splits": 8,
|
| 96 |
+
"convolutionFLOPs": 1441456128,
|
| 97 |
+
"columnBytes": 21676032
|
| 98 |
+
},
|
| 99 |
+
{
|
| 100 |
+
"node": "/head/final_layer2/conv/Conv",
|
| 101 |
+
"inputShape": [
|
| 102 |
+
1,
|
| 103 |
+
768,
|
| 104 |
+
24,
|
| 105 |
+
18
|
| 106 |
+
],
|
| 107 |
+
"weightShape": [
|
| 108 |
+
133,
|
| 109 |
+
768,
|
| 110 |
+
7,
|
| 111 |
+
7
|
| 112 |
+
],
|
| 113 |
+
"outputShape": [
|
| 114 |
+
1,
|
| 115 |
+
133,
|
| 116 |
+
24,
|
| 117 |
+
18
|
| 118 |
+
],
|
| 119 |
+
"kernel": [
|
| 120 |
+
7,
|
| 121 |
+
7
|
| 122 |
+
],
|
| 123 |
+
"pads": [
|
| 124 |
+
3,
|
| 125 |
+
3,
|
| 126 |
+
3,
|
| 127 |
+
3
|
| 128 |
+
],
|
| 129 |
+
"strides": [
|
| 130 |
+
1,
|
| 131 |
+
1
|
| 132 |
+
],
|
| 133 |
+
"dilations": [
|
| 134 |
+
1,
|
| 135 |
+
1
|
| 136 |
+
],
|
| 137 |
+
"splits": 8,
|
| 138 |
+
"convolutionFLOPs": 4324368384,
|
| 139 |
+
"columnBytes": 65028096
|
| 140 |
+
}
|
| 141 |
+
],
|
| 142 |
+
"comparisons": [
|
| 143 |
+
{
|
| 144 |
+
"publicFixtureImageId": 40083,
|
| 145 |
+
"parts": {
|
| 146 |
+
"all": {
|
| 147 |
+
"joints": 111,
|
| 148 |
+
"xyP95_px": 0.0,
|
| 149 |
+
"zP95_m": 0.0,
|
| 150 |
+
"xyMax_px": 0.0,
|
| 151 |
+
"zMax_m": 0.0,
|
| 152 |
+
"confidenceLossCount": 0
|
| 153 |
+
},
|
| 154 |
+
"body": {
|
| 155 |
+
"joints": 16,
|
| 156 |
+
"xyP95_px": 0.0,
|
| 157 |
+
"zP95_m": 0.0,
|
| 158 |
+
"xyMax_px": 0.0,
|
| 159 |
+
"zMax_m": 0.0,
|
| 160 |
+
"confidenceLossCount": 0
|
| 161 |
+
},
|
| 162 |
+
"body-feet": {
|
| 163 |
+
"joints": 22,
|
| 164 |
+
"xyP95_px": 0.0,
|
| 165 |
+
"zP95_m": 0.0,
|
| 166 |
+
"xyMax_px": 0.0,
|
| 167 |
+
"zMax_m": 0.0,
|
| 168 |
+
"confidenceLossCount": 0
|
| 169 |
+
},
|
| 170 |
+
"hips": {
|
| 171 |
+
"joints": 2,
|
| 172 |
+
"xyP95_px": 0.0,
|
| 173 |
+
"zP95_m": 0.0,
|
| 174 |
+
"xyMax_px": 0.0,
|
| 175 |
+
"zMax_m": 0.0,
|
| 176 |
+
"confidenceLossCount": 0
|
| 177 |
+
},
|
| 178 |
+
"knees": {
|
| 179 |
+
"joints": 2,
|
| 180 |
+
"xyP95_px": 0.0,
|
| 181 |
+
"zP95_m": 0.0,
|
| 182 |
+
"xyMax_px": 0.0,
|
| 183 |
+
"zMax_m": 0.0,
|
| 184 |
+
"confidenceLossCount": 0
|
| 185 |
+
},
|
| 186 |
+
"ankles": {
|
| 187 |
+
"joints": 2,
|
| 188 |
+
"xyP95_px": 0.0,
|
| 189 |
+
"zP95_m": 0.0,
|
| 190 |
+
"xyMax_px": 0.0,
|
| 191 |
+
"zMax_m": 0.0,
|
| 192 |
+
"confidenceLossCount": 0
|
| 193 |
+
},
|
| 194 |
+
"feet": {
|
| 195 |
+
"joints": 6,
|
| 196 |
+
"xyP95_px": 0.0,
|
| 197 |
+
"zP95_m": 0.0,
|
| 198 |
+
"xyMax_px": 0.0,
|
| 199 |
+
"zMax_m": 0.0,
|
| 200 |
+
"confidenceLossCount": 0
|
| 201 |
+
},
|
| 202 |
+
"face": {
|
| 203 |
+
"joints": 68,
|
| 204 |
+
"xyP95_px": 0.0,
|
| 205 |
+
"zP95_m": 0.0,
|
| 206 |
+
"xyMax_px": 0.0,
|
| 207 |
+
"zMax_m": 0.0,
|
| 208 |
+
"confidenceLossCount": 0
|
| 209 |
+
},
|
| 210 |
+
"hands": {
|
| 211 |
+
"joints": 21,
|
| 212 |
+
"xyP95_px": 0.0,
|
| 213 |
+
"zP95_m": 0.0,
|
| 214 |
+
"xyMax_px": 0.0,
|
| 215 |
+
"zMax_m": 0.0,
|
| 216 |
+
"confidenceLossCount": 0
|
| 217 |
+
}
|
| 218 |
+
},
|
| 219 |
+
"scoreMaxAbsoluteError": 5.960464477539062e-07
|
| 220 |
+
},
|
| 221 |
+
{
|
| 222 |
+
"publicFixtureImageId": 197388,
|
| 223 |
+
"parts": {
|
| 224 |
+
"all": {
|
| 225 |
+
"joints": 133,
|
| 226 |
+
"xyP95_px": 0.0,
|
| 227 |
+
"zP95_m": 0.0,
|
| 228 |
+
"xyMax_px": 0.0,
|
| 229 |
+
"zMax_m": 0.0,
|
| 230 |
+
"confidenceLossCount": 0
|
| 231 |
+
},
|
| 232 |
+
"body": {
|
| 233 |
+
"joints": 17,
|
| 234 |
+
"xyP95_px": 0.0,
|
| 235 |
+
"zP95_m": 0.0,
|
| 236 |
+
"xyMax_px": 0.0,
|
| 237 |
+
"zMax_m": 0.0,
|
| 238 |
+
"confidenceLossCount": 0
|
| 239 |
+
},
|
| 240 |
+
"body-feet": {
|
| 241 |
+
"joints": 23,
|
| 242 |
+
"xyP95_px": 0.0,
|
| 243 |
+
"zP95_m": 0.0,
|
| 244 |
+
"xyMax_px": 0.0,
|
| 245 |
+
"zMax_m": 0.0,
|
| 246 |
+
"confidenceLossCount": 0
|
| 247 |
+
},
|
| 248 |
+
"hips": {
|
| 249 |
+
"joints": 2,
|
| 250 |
+
"xyP95_px": 0.0,
|
| 251 |
+
"zP95_m": 0.0,
|
| 252 |
+
"xyMax_px": 0.0,
|
| 253 |
+
"zMax_m": 0.0,
|
| 254 |
+
"confidenceLossCount": 0
|
| 255 |
+
},
|
| 256 |
+
"knees": {
|
| 257 |
+
"joints": 2,
|
| 258 |
+
"xyP95_px": 0.0,
|
| 259 |
+
"zP95_m": 0.0,
|
| 260 |
+
"xyMax_px": 0.0,
|
| 261 |
+
"zMax_m": 0.0,
|
| 262 |
+
"confidenceLossCount": 0
|
| 263 |
+
},
|
| 264 |
+
"ankles": {
|
| 265 |
+
"joints": 2,
|
| 266 |
+
"xyP95_px": 0.0,
|
| 267 |
+
"zP95_m": 0.0,
|
| 268 |
+
"xyMax_px": 0.0,
|
| 269 |
+
"zMax_m": 0.0,
|
| 270 |
+
"confidenceLossCount": 0
|
| 271 |
+
},
|
| 272 |
+
"feet": {
|
| 273 |
+
"joints": 6,
|
| 274 |
+
"xyP95_px": 0.0,
|
| 275 |
+
"zP95_m": 0.0,
|
| 276 |
+
"xyMax_px": 0.0,
|
| 277 |
+
"zMax_m": 0.0,
|
| 278 |
+
"confidenceLossCount": 0
|
| 279 |
+
},
|
| 280 |
+
"face": {
|
| 281 |
+
"joints": 68,
|
| 282 |
+
"xyP95_px": 0.0,
|
| 283 |
+
"zP95_m": 0.0,
|
| 284 |
+
"xyMax_px": 0.0,
|
| 285 |
+
"zMax_m": 0.0,
|
| 286 |
+
"confidenceLossCount": 0
|
| 287 |
+
},
|
| 288 |
+
"hands": {
|
| 289 |
+
"joints": 42,
|
| 290 |
+
"xyP95_px": 0.0,
|
| 291 |
+
"zP95_m": 0.0,
|
| 292 |
+
"xyMax_px": 0.0,
|
| 293 |
+
"zMax_m": 0.0,
|
| 294 |
+
"confidenceLossCount": 0
|
| 295 |
+
}
|
| 296 |
+
},
|
| 297 |
+
"scoreMaxAbsoluteError": 4.172325134277344e-07
|
| 298 |
+
}
|
| 299 |
+
],
|
| 300 |
+
"referenceLatencyMs": [],
|
| 301 |
+
"note": "Same trained weights and 133-point 3D contract; explicit im2col and optional batched split-K reduction. CPU conversion tests are not WebGPU performance or target-video accuracy measurements."
|
| 302 |
+
},
|
| 303 |
+
{
|
| 304 |
+
"id": "rtmw3d-l-fp32-head-gemm",
|
| 305 |
+
"model": "rtmw3d-l",
|
| 306 |
+
"policy": "fp32-head-gemm",
|
| 307 |
+
"file": "experiments/rtmw3d-l-fp32-head-gemm.onnx",
|
| 308 |
+
"sizeBytes": 230137728,
|
| 309 |
+
"sha256": "b1a60dccb44759a9af3de6d54691c6dc0c9318909722f7baaf2cf781697ccefe",
|
| 310 |
+
"precision": "fp32",
|
| 311 |
+
"experimental": true,
|
| 312 |
+
"qualityPassed": true,
|
| 313 |
+
"browserGPUValidated": false,
|
| 314 |
+
"sourceRevision": "88ee5cb728cb2ed3efaf43011c9dd4908a6504b7",
|
| 315 |
+
"sourceModelSha256": "4e37e5e9043817c5ff387b61c44bb48af94ad42e4978c4e43780761c65582abc",
|
| 316 |
+
"upstream": "https://github.com/open-mmlab/mmpose",
|
| 317 |
+
"weightsUnchangedFP32": true,
|
| 318 |
+
"input": {
|
| 319 |
+
"name": "input",
|
| 320 |
+
"type": "float32",
|
| 321 |
+
"shape": [
|
| 322 |
+
1,
|
| 323 |
+
3,
|
| 324 |
+
384,
|
| 325 |
+
288
|
| 326 |
+
]
|
| 327 |
+
},
|
| 328 |
+
"outputs": [
|
| 329 |
+
{
|
| 330 |
+
"name": "keypoints",
|
| 331 |
+
"type": "float32",
|
| 332 |
+
"shape": [
|
| 333 |
+
1,
|
| 334 |
+
133,
|
| 335 |
+
3
|
| 336 |
+
],
|
| 337 |
+
"units": [
|
| 338 |
+
"crop-pixels",
|
| 339 |
+
"crop-pixels",
|
| 340 |
+
"root-relative-metres"
|
| 341 |
+
]
|
| 342 |
+
},
|
| 343 |
+
{
|
| 344 |
+
"name": "scores",
|
| 345 |
+
"type": "float32",
|
| 346 |
+
"shape": [
|
| 347 |
+
1,
|
| 348 |
+
133
|
| 349 |
+
]
|
| 350 |
+
}
|
| 351 |
+
],
|
| 352 |
+
"outputBytesPerFrame": 2128,
|
| 353 |
+
"changedConvolutions": [
|
| 354 |
+
{
|
| 355 |
+
"node": "/head/final_layer/conv/Conv",
|
| 356 |
+
"inputShape": [
|
| 357 |
+
1,
|
| 358 |
+
1024,
|
| 359 |
+
12,
|
| 360 |
+
9
|
| 361 |
+
],
|
| 362 |
+
"weightShape": [
|
| 363 |
+
133,
|
| 364 |
+
1024,
|
| 365 |
+
7,
|
| 366 |
+
7
|
| 367 |
+
],
|
| 368 |
+
"outputShape": [
|
| 369 |
+
1,
|
| 370 |
+
133,
|
| 371 |
+
12,
|
| 372 |
+
9
|
| 373 |
+
],
|
| 374 |
+
"kernel": [
|
| 375 |
+
7,
|
| 376 |
+
7
|
| 377 |
+
],
|
| 378 |
+
"pads": [
|
| 379 |
+
3,
|
| 380 |
+
3,
|
| 381 |
+
3,
|
| 382 |
+
3
|
| 383 |
+
],
|
| 384 |
+
"strides": [
|
| 385 |
+
1,
|
| 386 |
+
1
|
| 387 |
+
],
|
| 388 |
+
"dilations": [
|
| 389 |
+
1,
|
| 390 |
+
1
|
| 391 |
+
],
|
| 392 |
+
"splits": 1,
|
| 393 |
+
"convolutionFLOPs": 1441456128,
|
| 394 |
+
"columnBytes": 21676032
|
| 395 |
+
},
|
| 396 |
+
{
|
| 397 |
+
"node": "/head/final_layer2/conv/Conv",
|
| 398 |
+
"inputShape": [
|
| 399 |
+
1,
|
| 400 |
+
768,
|
| 401 |
+
24,
|
| 402 |
+
18
|
| 403 |
+
],
|
| 404 |
+
"weightShape": [
|
| 405 |
+
133,
|
| 406 |
+
768,
|
| 407 |
+
7,
|
| 408 |
+
7
|
| 409 |
+
],
|
| 410 |
+
"outputShape": [
|
| 411 |
+
1,
|
| 412 |
+
133,
|
| 413 |
+
24,
|
| 414 |
+
18
|
| 415 |
+
],
|
| 416 |
+
"kernel": [
|
| 417 |
+
7,
|
| 418 |
+
7
|
| 419 |
+
],
|
| 420 |
+
"pads": [
|
| 421 |
+
3,
|
| 422 |
+
3,
|
| 423 |
+
3,
|
| 424 |
+
3
|
| 425 |
+
],
|
| 426 |
+
"strides": [
|
| 427 |
+
1,
|
| 428 |
+
1
|
| 429 |
+
],
|
| 430 |
+
"dilations": [
|
| 431 |
+
1,
|
| 432 |
+
1
|
| 433 |
+
],
|
| 434 |
+
"splits": 1,
|
| 435 |
+
"convolutionFLOPs": 4324368384,
|
| 436 |
+
"columnBytes": 65028096
|
| 437 |
+
}
|
| 438 |
+
],
|
| 439 |
+
"comparisons": [
|
| 440 |
+
{
|
| 441 |
+
"publicFixtureImageId": 40083,
|
| 442 |
+
"parts": {
|
| 443 |
+
"all": {
|
| 444 |
+
"joints": 111,
|
| 445 |
+
"xyP95_px": 0.0,
|
| 446 |
+
"zP95_m": 0.0,
|
| 447 |
+
"xyMax_px": 0.0,
|
| 448 |
+
"zMax_m": 0.0,
|
| 449 |
+
"confidenceLossCount": 0
|
| 450 |
+
},
|
| 451 |
+
"body": {
|
| 452 |
+
"joints": 16,
|
| 453 |
+
"xyP95_px": 0.0,
|
| 454 |
+
"zP95_m": 0.0,
|
| 455 |
+
"xyMax_px": 0.0,
|
| 456 |
+
"zMax_m": 0.0,
|
| 457 |
+
"confidenceLossCount": 0
|
| 458 |
+
},
|
| 459 |
+
"body-feet": {
|
| 460 |
+
"joints": 22,
|
| 461 |
+
"xyP95_px": 0.0,
|
| 462 |
+
"zP95_m": 0.0,
|
| 463 |
+
"xyMax_px": 0.0,
|
| 464 |
+
"zMax_m": 0.0,
|
| 465 |
+
"confidenceLossCount": 0
|
| 466 |
+
},
|
| 467 |
+
"hips": {
|
| 468 |
+
"joints": 2,
|
| 469 |
+
"xyP95_px": 0.0,
|
| 470 |
+
"zP95_m": 0.0,
|
| 471 |
+
"xyMax_px": 0.0,
|
| 472 |
+
"zMax_m": 0.0,
|
| 473 |
+
"confidenceLossCount": 0
|
| 474 |
+
},
|
| 475 |
+
"knees": {
|
| 476 |
+
"joints": 2,
|
| 477 |
+
"xyP95_px": 0.0,
|
| 478 |
+
"zP95_m": 0.0,
|
| 479 |
+
"xyMax_px": 0.0,
|
| 480 |
+
"zMax_m": 0.0,
|
| 481 |
+
"confidenceLossCount": 0
|
| 482 |
+
},
|
| 483 |
+
"ankles": {
|
| 484 |
+
"joints": 2,
|
| 485 |
+
"xyP95_px": 0.0,
|
| 486 |
+
"zP95_m": 0.0,
|
| 487 |
+
"xyMax_px": 0.0,
|
| 488 |
+
"zMax_m": 0.0,
|
| 489 |
+
"confidenceLossCount": 0
|
| 490 |
+
},
|
| 491 |
+
"feet": {
|
| 492 |
+
"joints": 6,
|
| 493 |
+
"xyP95_px": 0.0,
|
| 494 |
+
"zP95_m": 0.0,
|
| 495 |
+
"xyMax_px": 0.0,
|
| 496 |
+
"zMax_m": 0.0,
|
| 497 |
+
"confidenceLossCount": 0
|
| 498 |
+
},
|
| 499 |
+
"face": {
|
| 500 |
+
"joints": 68,
|
| 501 |
+
"xyP95_px": 0.0,
|
| 502 |
+
"zP95_m": 0.0,
|
| 503 |
+
"xyMax_px": 0.0,
|
| 504 |
+
"zMax_m": 0.0,
|
| 505 |
+
"confidenceLossCount": 0
|
| 506 |
+
},
|
| 507 |
+
"hands": {
|
| 508 |
+
"joints": 21,
|
| 509 |
+
"xyP95_px": 0.0,
|
| 510 |
+
"zP95_m": 0.0,
|
| 511 |
+
"xyMax_px": 0.0,
|
| 512 |
+
"zMax_m": 0.0,
|
| 513 |
+
"confidenceLossCount": 0
|
| 514 |
+
}
|
| 515 |
+
},
|
| 516 |
+
"scoreMaxAbsoluteError": 4.76837158203125e-07
|
| 517 |
+
},
|
| 518 |
+
{
|
| 519 |
+
"publicFixtureImageId": 197388,
|
| 520 |
+
"parts": {
|
| 521 |
+
"all": {
|
| 522 |
+
"joints": 133,
|
| 523 |
+
"xyP95_px": 0.0,
|
| 524 |
+
"zP95_m": 0.0,
|
| 525 |
+
"xyMax_px": 0.0,
|
| 526 |
+
"zMax_m": 0.0,
|
| 527 |
+
"confidenceLossCount": 0
|
| 528 |
+
},
|
| 529 |
+
"body": {
|
| 530 |
+
"joints": 17,
|
| 531 |
+
"xyP95_px": 0.0,
|
| 532 |
+
"zP95_m": 0.0,
|
| 533 |
+
"xyMax_px": 0.0,
|
| 534 |
+
"zMax_m": 0.0,
|
| 535 |
+
"confidenceLossCount": 0
|
| 536 |
+
},
|
| 537 |
+
"body-feet": {
|
| 538 |
+
"joints": 23,
|
| 539 |
+
"xyP95_px": 0.0,
|
| 540 |
+
"zP95_m": 0.0,
|
| 541 |
+
"xyMax_px": 0.0,
|
| 542 |
+
"zMax_m": 0.0,
|
| 543 |
+
"confidenceLossCount": 0
|
| 544 |
+
},
|
| 545 |
+
"hips": {
|
| 546 |
+
"joints": 2,
|
| 547 |
+
"xyP95_px": 0.0,
|
| 548 |
+
"zP95_m": 0.0,
|
| 549 |
+
"xyMax_px": 0.0,
|
| 550 |
+
"zMax_m": 0.0,
|
| 551 |
+
"confidenceLossCount": 0
|
| 552 |
+
},
|
| 553 |
+
"knees": {
|
| 554 |
+
"joints": 2,
|
| 555 |
+
"xyP95_px": 0.0,
|
| 556 |
+
"zP95_m": 0.0,
|
| 557 |
+
"xyMax_px": 0.0,
|
| 558 |
+
"zMax_m": 0.0,
|
| 559 |
+
"confidenceLossCount": 0
|
| 560 |
+
},
|
| 561 |
+
"ankles": {
|
| 562 |
+
"joints": 2,
|
| 563 |
+
"xyP95_px": 0.0,
|
| 564 |
+
"zP95_m": 0.0,
|
| 565 |
+
"xyMax_px": 0.0,
|
| 566 |
+
"zMax_m": 0.0,
|
| 567 |
+
"confidenceLossCount": 0
|
| 568 |
+
},
|
| 569 |
+
"feet": {
|
| 570 |
+
"joints": 6,
|
| 571 |
+
"xyP95_px": 0.0,
|
| 572 |
+
"zP95_m": 0.0,
|
| 573 |
+
"xyMax_px": 0.0,
|
| 574 |
+
"zMax_m": 0.0,
|
| 575 |
+
"confidenceLossCount": 0
|
| 576 |
+
},
|
| 577 |
+
"face": {
|
| 578 |
+
"joints": 68,
|
| 579 |
+
"xyP95_px": 0.0,
|
| 580 |
+
"zP95_m": 0.0,
|
| 581 |
+
"xyMax_px": 0.0,
|
| 582 |
+
"zMax_m": 0.0,
|
| 583 |
+
"confidenceLossCount": 0
|
| 584 |
+
},
|
| 585 |
+
"hands": {
|
| 586 |
+
"joints": 42,
|
| 587 |
+
"xyP95_px": 0.0,
|
| 588 |
+
"zP95_m": 0.0,
|
| 589 |
+
"xyMax_px": 0.0,
|
| 590 |
+
"zMax_m": 0.0,
|
| 591 |
+
"confidenceLossCount": 0
|
| 592 |
+
}
|
| 593 |
+
},
|
| 594 |
+
"scoreMaxAbsoluteError": 4.76837158203125e-07
|
| 595 |
+
}
|
| 596 |
+
],
|
| 597 |
+
"referenceLatencyMs": [],
|
| 598 |
+
"note": "Same trained weights and 133-point 3D contract; explicit im2col and optional batched split-K reduction. CPU conversion tests are not WebGPU performance or target-video accuracy measurements."
|
| 599 |
+
}
|
| 600 |
+
],
|
| 601 |
+
"browserGPUValidated": false
|
| 602 |
+
}
|
experiments/rtmw3d-l-fp32-head-gemm-split8.json
ADDED
|
@@ -0,0 +1,297 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"id": "rtmw3d-l-fp32-head-gemm-split8",
|
| 3 |
+
"model": "rtmw3d-l",
|
| 4 |
+
"policy": "fp32-head-gemm-split8",
|
| 5 |
+
"file": "experiments/rtmw3d-l-fp32-head-gemm-split8.onnx",
|
| 6 |
+
"sizeBytes": 230139419,
|
| 7 |
+
"sha256": "a2767b1c20120f8351852f1c87617794c6852ace4de76bf12c7c483fa8ab76ed",
|
| 8 |
+
"precision": "fp32",
|
| 9 |
+
"experimental": true,
|
| 10 |
+
"qualityPassed": true,
|
| 11 |
+
"browserGPUValidated": false,
|
| 12 |
+
"sourceRevision": "88ee5cb728cb2ed3efaf43011c9dd4908a6504b7",
|
| 13 |
+
"sourceModelSha256": "4e37e5e9043817c5ff387b61c44bb48af94ad42e4978c4e43780761c65582abc",
|
| 14 |
+
"upstream": "https://github.com/open-mmlab/mmpose",
|
| 15 |
+
"weightsUnchangedFP32": true,
|
| 16 |
+
"input": {
|
| 17 |
+
"name": "input",
|
| 18 |
+
"type": "float32",
|
| 19 |
+
"shape": [
|
| 20 |
+
1,
|
| 21 |
+
3,
|
| 22 |
+
384,
|
| 23 |
+
288
|
| 24 |
+
]
|
| 25 |
+
},
|
| 26 |
+
"outputs": [
|
| 27 |
+
{
|
| 28 |
+
"name": "keypoints",
|
| 29 |
+
"type": "float32",
|
| 30 |
+
"shape": [
|
| 31 |
+
1,
|
| 32 |
+
133,
|
| 33 |
+
3
|
| 34 |
+
],
|
| 35 |
+
"units": [
|
| 36 |
+
"crop-pixels",
|
| 37 |
+
"crop-pixels",
|
| 38 |
+
"root-relative-metres"
|
| 39 |
+
]
|
| 40 |
+
},
|
| 41 |
+
{
|
| 42 |
+
"name": "scores",
|
| 43 |
+
"type": "float32",
|
| 44 |
+
"shape": [
|
| 45 |
+
1,
|
| 46 |
+
133
|
| 47 |
+
]
|
| 48 |
+
}
|
| 49 |
+
],
|
| 50 |
+
"outputBytesPerFrame": 2128,
|
| 51 |
+
"changedConvolutions": [
|
| 52 |
+
{
|
| 53 |
+
"node": "/head/final_layer/conv/Conv",
|
| 54 |
+
"inputShape": [
|
| 55 |
+
1,
|
| 56 |
+
1024,
|
| 57 |
+
12,
|
| 58 |
+
9
|
| 59 |
+
],
|
| 60 |
+
"weightShape": [
|
| 61 |
+
133,
|
| 62 |
+
1024,
|
| 63 |
+
7,
|
| 64 |
+
7
|
| 65 |
+
],
|
| 66 |
+
"outputShape": [
|
| 67 |
+
1,
|
| 68 |
+
133,
|
| 69 |
+
12,
|
| 70 |
+
9
|
| 71 |
+
],
|
| 72 |
+
"kernel": [
|
| 73 |
+
7,
|
| 74 |
+
7
|
| 75 |
+
],
|
| 76 |
+
"pads": [
|
| 77 |
+
3,
|
| 78 |
+
3,
|
| 79 |
+
3,
|
| 80 |
+
3
|
| 81 |
+
],
|
| 82 |
+
"strides": [
|
| 83 |
+
1,
|
| 84 |
+
1
|
| 85 |
+
],
|
| 86 |
+
"dilations": [
|
| 87 |
+
1,
|
| 88 |
+
1
|
| 89 |
+
],
|
| 90 |
+
"splits": 8,
|
| 91 |
+
"convolutionFLOPs": 1441456128,
|
| 92 |
+
"columnBytes": 21676032
|
| 93 |
+
},
|
| 94 |
+
{
|
| 95 |
+
"node": "/head/final_layer2/conv/Conv",
|
| 96 |
+
"inputShape": [
|
| 97 |
+
1,
|
| 98 |
+
768,
|
| 99 |
+
24,
|
| 100 |
+
18
|
| 101 |
+
],
|
| 102 |
+
"weightShape": [
|
| 103 |
+
133,
|
| 104 |
+
768,
|
| 105 |
+
7,
|
| 106 |
+
7
|
| 107 |
+
],
|
| 108 |
+
"outputShape": [
|
| 109 |
+
1,
|
| 110 |
+
133,
|
| 111 |
+
24,
|
| 112 |
+
18
|
| 113 |
+
],
|
| 114 |
+
"kernel": [
|
| 115 |
+
7,
|
| 116 |
+
7
|
| 117 |
+
],
|
| 118 |
+
"pads": [
|
| 119 |
+
3,
|
| 120 |
+
3,
|
| 121 |
+
3,
|
| 122 |
+
3
|
| 123 |
+
],
|
| 124 |
+
"strides": [
|
| 125 |
+
1,
|
| 126 |
+
1
|
| 127 |
+
],
|
| 128 |
+
"dilations": [
|
| 129 |
+
1,
|
| 130 |
+
1
|
| 131 |
+
],
|
| 132 |
+
"splits": 8,
|
| 133 |
+
"convolutionFLOPs": 4324368384,
|
| 134 |
+
"columnBytes": 65028096
|
| 135 |
+
}
|
| 136 |
+
],
|
| 137 |
+
"comparisons": [
|
| 138 |
+
{
|
| 139 |
+
"publicFixtureImageId": 40083,
|
| 140 |
+
"parts": {
|
| 141 |
+
"all": {
|
| 142 |
+
"joints": 111,
|
| 143 |
+
"xyP95_px": 0.0,
|
| 144 |
+
"zP95_m": 0.0,
|
| 145 |
+
"xyMax_px": 0.0,
|
| 146 |
+
"zMax_m": 0.0,
|
| 147 |
+
"confidenceLossCount": 0
|
| 148 |
+
},
|
| 149 |
+
"body": {
|
| 150 |
+
"joints": 16,
|
| 151 |
+
"xyP95_px": 0.0,
|
| 152 |
+
"zP95_m": 0.0,
|
| 153 |
+
"xyMax_px": 0.0,
|
| 154 |
+
"zMax_m": 0.0,
|
| 155 |
+
"confidenceLossCount": 0
|
| 156 |
+
},
|
| 157 |
+
"body-feet": {
|
| 158 |
+
"joints": 22,
|
| 159 |
+
"xyP95_px": 0.0,
|
| 160 |
+
"zP95_m": 0.0,
|
| 161 |
+
"xyMax_px": 0.0,
|
| 162 |
+
"zMax_m": 0.0,
|
| 163 |
+
"confidenceLossCount": 0
|
| 164 |
+
},
|
| 165 |
+
"hips": {
|
| 166 |
+
"joints": 2,
|
| 167 |
+
"xyP95_px": 0.0,
|
| 168 |
+
"zP95_m": 0.0,
|
| 169 |
+
"xyMax_px": 0.0,
|
| 170 |
+
"zMax_m": 0.0,
|
| 171 |
+
"confidenceLossCount": 0
|
| 172 |
+
},
|
| 173 |
+
"knees": {
|
| 174 |
+
"joints": 2,
|
| 175 |
+
"xyP95_px": 0.0,
|
| 176 |
+
"zP95_m": 0.0,
|
| 177 |
+
"xyMax_px": 0.0,
|
| 178 |
+
"zMax_m": 0.0,
|
| 179 |
+
"confidenceLossCount": 0
|
| 180 |
+
},
|
| 181 |
+
"ankles": {
|
| 182 |
+
"joints": 2,
|
| 183 |
+
"xyP95_px": 0.0,
|
| 184 |
+
"zP95_m": 0.0,
|
| 185 |
+
"xyMax_px": 0.0,
|
| 186 |
+
"zMax_m": 0.0,
|
| 187 |
+
"confidenceLossCount": 0
|
| 188 |
+
},
|
| 189 |
+
"feet": {
|
| 190 |
+
"joints": 6,
|
| 191 |
+
"xyP95_px": 0.0,
|
| 192 |
+
"zP95_m": 0.0,
|
| 193 |
+
"xyMax_px": 0.0,
|
| 194 |
+
"zMax_m": 0.0,
|
| 195 |
+
"confidenceLossCount": 0
|
| 196 |
+
},
|
| 197 |
+
"face": {
|
| 198 |
+
"joints": 68,
|
| 199 |
+
"xyP95_px": 0.0,
|
| 200 |
+
"zP95_m": 0.0,
|
| 201 |
+
"xyMax_px": 0.0,
|
| 202 |
+
"zMax_m": 0.0,
|
| 203 |
+
"confidenceLossCount": 0
|
| 204 |
+
},
|
| 205 |
+
"hands": {
|
| 206 |
+
"joints": 21,
|
| 207 |
+
"xyP95_px": 0.0,
|
| 208 |
+
"zP95_m": 0.0,
|
| 209 |
+
"xyMax_px": 0.0,
|
| 210 |
+
"zMax_m": 0.0,
|
| 211 |
+
"confidenceLossCount": 0
|
| 212 |
+
}
|
| 213 |
+
},
|
| 214 |
+
"scoreMaxAbsoluteError": 5.960464477539062e-07
|
| 215 |
+
},
|
| 216 |
+
{
|
| 217 |
+
"publicFixtureImageId": 197388,
|
| 218 |
+
"parts": {
|
| 219 |
+
"all": {
|
| 220 |
+
"joints": 133,
|
| 221 |
+
"xyP95_px": 0.0,
|
| 222 |
+
"zP95_m": 0.0,
|
| 223 |
+
"xyMax_px": 0.0,
|
| 224 |
+
"zMax_m": 0.0,
|
| 225 |
+
"confidenceLossCount": 0
|
| 226 |
+
},
|
| 227 |
+
"body": {
|
| 228 |
+
"joints": 17,
|
| 229 |
+
"xyP95_px": 0.0,
|
| 230 |
+
"zP95_m": 0.0,
|
| 231 |
+
"xyMax_px": 0.0,
|
| 232 |
+
"zMax_m": 0.0,
|
| 233 |
+
"confidenceLossCount": 0
|
| 234 |
+
},
|
| 235 |
+
"body-feet": {
|
| 236 |
+
"joints": 23,
|
| 237 |
+
"xyP95_px": 0.0,
|
| 238 |
+
"zP95_m": 0.0,
|
| 239 |
+
"xyMax_px": 0.0,
|
| 240 |
+
"zMax_m": 0.0,
|
| 241 |
+
"confidenceLossCount": 0
|
| 242 |
+
},
|
| 243 |
+
"hips": {
|
| 244 |
+
"joints": 2,
|
| 245 |
+
"xyP95_px": 0.0,
|
| 246 |
+
"zP95_m": 0.0,
|
| 247 |
+
"xyMax_px": 0.0,
|
| 248 |
+
"zMax_m": 0.0,
|
| 249 |
+
"confidenceLossCount": 0
|
| 250 |
+
},
|
| 251 |
+
"knees": {
|
| 252 |
+
"joints": 2,
|
| 253 |
+
"xyP95_px": 0.0,
|
| 254 |
+
"zP95_m": 0.0,
|
| 255 |
+
"xyMax_px": 0.0,
|
| 256 |
+
"zMax_m": 0.0,
|
| 257 |
+
"confidenceLossCount": 0
|
| 258 |
+
},
|
| 259 |
+
"ankles": {
|
| 260 |
+
"joints": 2,
|
| 261 |
+
"xyP95_px": 0.0,
|
| 262 |
+
"zP95_m": 0.0,
|
| 263 |
+
"xyMax_px": 0.0,
|
| 264 |
+
"zMax_m": 0.0,
|
| 265 |
+
"confidenceLossCount": 0
|
| 266 |
+
},
|
| 267 |
+
"feet": {
|
| 268 |
+
"joints": 6,
|
| 269 |
+
"xyP95_px": 0.0,
|
| 270 |
+
"zP95_m": 0.0,
|
| 271 |
+
"xyMax_px": 0.0,
|
| 272 |
+
"zMax_m": 0.0,
|
| 273 |
+
"confidenceLossCount": 0
|
| 274 |
+
},
|
| 275 |
+
"face": {
|
| 276 |
+
"joints": 68,
|
| 277 |
+
"xyP95_px": 0.0,
|
| 278 |
+
"zP95_m": 0.0,
|
| 279 |
+
"xyMax_px": 0.0,
|
| 280 |
+
"zMax_m": 0.0,
|
| 281 |
+
"confidenceLossCount": 0
|
| 282 |
+
},
|
| 283 |
+
"hands": {
|
| 284 |
+
"joints": 42,
|
| 285 |
+
"xyP95_px": 0.0,
|
| 286 |
+
"zP95_m": 0.0,
|
| 287 |
+
"xyMax_px": 0.0,
|
| 288 |
+
"zMax_m": 0.0,
|
| 289 |
+
"confidenceLossCount": 0
|
| 290 |
+
}
|
| 291 |
+
},
|
| 292 |
+
"scoreMaxAbsoluteError": 4.172325134277344e-07
|
| 293 |
+
}
|
| 294 |
+
],
|
| 295 |
+
"referenceLatencyMs": [],
|
| 296 |
+
"note": "Same trained weights and 133-point 3D contract; explicit im2col and optional batched split-K reduction. CPU conversion tests are not WebGPU performance or target-video accuracy measurements."
|
| 297 |
+
}
|
experiments/rtmw3d-l-fp32-head-gemm-split8.onnx
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a2767b1c20120f8351852f1c87617794c6852ace4de76bf12c7c483fa8ab76ed
|
| 3 |
+
size 230139419
|
experiments/rtmw3d-l-fp32-head-gemm.json
ADDED
|
@@ -0,0 +1,297 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"id": "rtmw3d-l-fp32-head-gemm",
|
| 3 |
+
"model": "rtmw3d-l",
|
| 4 |
+
"policy": "fp32-head-gemm",
|
| 5 |
+
"file": "experiments/rtmw3d-l-fp32-head-gemm.onnx",
|
| 6 |
+
"sizeBytes": 230137728,
|
| 7 |
+
"sha256": "b1a60dccb44759a9af3de6d54691c6dc0c9318909722f7baaf2cf781697ccefe",
|
| 8 |
+
"precision": "fp32",
|
| 9 |
+
"experimental": true,
|
| 10 |
+
"qualityPassed": true,
|
| 11 |
+
"browserGPUValidated": false,
|
| 12 |
+
"sourceRevision": "88ee5cb728cb2ed3efaf43011c9dd4908a6504b7",
|
| 13 |
+
"sourceModelSha256": "4e37e5e9043817c5ff387b61c44bb48af94ad42e4978c4e43780761c65582abc",
|
| 14 |
+
"upstream": "https://github.com/open-mmlab/mmpose",
|
| 15 |
+
"weightsUnchangedFP32": true,
|
| 16 |
+
"input": {
|
| 17 |
+
"name": "input",
|
| 18 |
+
"type": "float32",
|
| 19 |
+
"shape": [
|
| 20 |
+
1,
|
| 21 |
+
3,
|
| 22 |
+
384,
|
| 23 |
+
288
|
| 24 |
+
]
|
| 25 |
+
},
|
| 26 |
+
"outputs": [
|
| 27 |
+
{
|
| 28 |
+
"name": "keypoints",
|
| 29 |
+
"type": "float32",
|
| 30 |
+
"shape": [
|
| 31 |
+
1,
|
| 32 |
+
133,
|
| 33 |
+
3
|
| 34 |
+
],
|
| 35 |
+
"units": [
|
| 36 |
+
"crop-pixels",
|
| 37 |
+
"crop-pixels",
|
| 38 |
+
"root-relative-metres"
|
| 39 |
+
]
|
| 40 |
+
},
|
| 41 |
+
{
|
| 42 |
+
"name": "scores",
|
| 43 |
+
"type": "float32",
|
| 44 |
+
"shape": [
|
| 45 |
+
1,
|
| 46 |
+
133
|
| 47 |
+
]
|
| 48 |
+
}
|
| 49 |
+
],
|
| 50 |
+
"outputBytesPerFrame": 2128,
|
| 51 |
+
"changedConvolutions": [
|
| 52 |
+
{
|
| 53 |
+
"node": "/head/final_layer/conv/Conv",
|
| 54 |
+
"inputShape": [
|
| 55 |
+
1,
|
| 56 |
+
1024,
|
| 57 |
+
12,
|
| 58 |
+
9
|
| 59 |
+
],
|
| 60 |
+
"weightShape": [
|
| 61 |
+
133,
|
| 62 |
+
1024,
|
| 63 |
+
7,
|
| 64 |
+
7
|
| 65 |
+
],
|
| 66 |
+
"outputShape": [
|
| 67 |
+
1,
|
| 68 |
+
133,
|
| 69 |
+
12,
|
| 70 |
+
9
|
| 71 |
+
],
|
| 72 |
+
"kernel": [
|
| 73 |
+
7,
|
| 74 |
+
7
|
| 75 |
+
],
|
| 76 |
+
"pads": [
|
| 77 |
+
3,
|
| 78 |
+
3,
|
| 79 |
+
3,
|
| 80 |
+
3
|
| 81 |
+
],
|
| 82 |
+
"strides": [
|
| 83 |
+
1,
|
| 84 |
+
1
|
| 85 |
+
],
|
| 86 |
+
"dilations": [
|
| 87 |
+
1,
|
| 88 |
+
1
|
| 89 |
+
],
|
| 90 |
+
"splits": 1,
|
| 91 |
+
"convolutionFLOPs": 1441456128,
|
| 92 |
+
"columnBytes": 21676032
|
| 93 |
+
},
|
| 94 |
+
{
|
| 95 |
+
"node": "/head/final_layer2/conv/Conv",
|
| 96 |
+
"inputShape": [
|
| 97 |
+
1,
|
| 98 |
+
768,
|
| 99 |
+
24,
|
| 100 |
+
18
|
| 101 |
+
],
|
| 102 |
+
"weightShape": [
|
| 103 |
+
133,
|
| 104 |
+
768,
|
| 105 |
+
7,
|
| 106 |
+
7
|
| 107 |
+
],
|
| 108 |
+
"outputShape": [
|
| 109 |
+
1,
|
| 110 |
+
133,
|
| 111 |
+
24,
|
| 112 |
+
18
|
| 113 |
+
],
|
| 114 |
+
"kernel": [
|
| 115 |
+
7,
|
| 116 |
+
7
|
| 117 |
+
],
|
| 118 |
+
"pads": [
|
| 119 |
+
3,
|
| 120 |
+
3,
|
| 121 |
+
3,
|
| 122 |
+
3
|
| 123 |
+
],
|
| 124 |
+
"strides": [
|
| 125 |
+
1,
|
| 126 |
+
1
|
| 127 |
+
],
|
| 128 |
+
"dilations": [
|
| 129 |
+
1,
|
| 130 |
+
1
|
| 131 |
+
],
|
| 132 |
+
"splits": 1,
|
| 133 |
+
"convolutionFLOPs": 4324368384,
|
| 134 |
+
"columnBytes": 65028096
|
| 135 |
+
}
|
| 136 |
+
],
|
| 137 |
+
"comparisons": [
|
| 138 |
+
{
|
| 139 |
+
"publicFixtureImageId": 40083,
|
| 140 |
+
"parts": {
|
| 141 |
+
"all": {
|
| 142 |
+
"joints": 111,
|
| 143 |
+
"xyP95_px": 0.0,
|
| 144 |
+
"zP95_m": 0.0,
|
| 145 |
+
"xyMax_px": 0.0,
|
| 146 |
+
"zMax_m": 0.0,
|
| 147 |
+
"confidenceLossCount": 0
|
| 148 |
+
},
|
| 149 |
+
"body": {
|
| 150 |
+
"joints": 16,
|
| 151 |
+
"xyP95_px": 0.0,
|
| 152 |
+
"zP95_m": 0.0,
|
| 153 |
+
"xyMax_px": 0.0,
|
| 154 |
+
"zMax_m": 0.0,
|
| 155 |
+
"confidenceLossCount": 0
|
| 156 |
+
},
|
| 157 |
+
"body-feet": {
|
| 158 |
+
"joints": 22,
|
| 159 |
+
"xyP95_px": 0.0,
|
| 160 |
+
"zP95_m": 0.0,
|
| 161 |
+
"xyMax_px": 0.0,
|
| 162 |
+
"zMax_m": 0.0,
|
| 163 |
+
"confidenceLossCount": 0
|
| 164 |
+
},
|
| 165 |
+
"hips": {
|
| 166 |
+
"joints": 2,
|
| 167 |
+
"xyP95_px": 0.0,
|
| 168 |
+
"zP95_m": 0.0,
|
| 169 |
+
"xyMax_px": 0.0,
|
| 170 |
+
"zMax_m": 0.0,
|
| 171 |
+
"confidenceLossCount": 0
|
| 172 |
+
},
|
| 173 |
+
"knees": {
|
| 174 |
+
"joints": 2,
|
| 175 |
+
"xyP95_px": 0.0,
|
| 176 |
+
"zP95_m": 0.0,
|
| 177 |
+
"xyMax_px": 0.0,
|
| 178 |
+
"zMax_m": 0.0,
|
| 179 |
+
"confidenceLossCount": 0
|
| 180 |
+
},
|
| 181 |
+
"ankles": {
|
| 182 |
+
"joints": 2,
|
| 183 |
+
"xyP95_px": 0.0,
|
| 184 |
+
"zP95_m": 0.0,
|
| 185 |
+
"xyMax_px": 0.0,
|
| 186 |
+
"zMax_m": 0.0,
|
| 187 |
+
"confidenceLossCount": 0
|
| 188 |
+
},
|
| 189 |
+
"feet": {
|
| 190 |
+
"joints": 6,
|
| 191 |
+
"xyP95_px": 0.0,
|
| 192 |
+
"zP95_m": 0.0,
|
| 193 |
+
"xyMax_px": 0.0,
|
| 194 |
+
"zMax_m": 0.0,
|
| 195 |
+
"confidenceLossCount": 0
|
| 196 |
+
},
|
| 197 |
+
"face": {
|
| 198 |
+
"joints": 68,
|
| 199 |
+
"xyP95_px": 0.0,
|
| 200 |
+
"zP95_m": 0.0,
|
| 201 |
+
"xyMax_px": 0.0,
|
| 202 |
+
"zMax_m": 0.0,
|
| 203 |
+
"confidenceLossCount": 0
|
| 204 |
+
},
|
| 205 |
+
"hands": {
|
| 206 |
+
"joints": 21,
|
| 207 |
+
"xyP95_px": 0.0,
|
| 208 |
+
"zP95_m": 0.0,
|
| 209 |
+
"xyMax_px": 0.0,
|
| 210 |
+
"zMax_m": 0.0,
|
| 211 |
+
"confidenceLossCount": 0
|
| 212 |
+
}
|
| 213 |
+
},
|
| 214 |
+
"scoreMaxAbsoluteError": 4.76837158203125e-07
|
| 215 |
+
},
|
| 216 |
+
{
|
| 217 |
+
"publicFixtureImageId": 197388,
|
| 218 |
+
"parts": {
|
| 219 |
+
"all": {
|
| 220 |
+
"joints": 133,
|
| 221 |
+
"xyP95_px": 0.0,
|
| 222 |
+
"zP95_m": 0.0,
|
| 223 |
+
"xyMax_px": 0.0,
|
| 224 |
+
"zMax_m": 0.0,
|
| 225 |
+
"confidenceLossCount": 0
|
| 226 |
+
},
|
| 227 |
+
"body": {
|
| 228 |
+
"joints": 17,
|
| 229 |
+
"xyP95_px": 0.0,
|
| 230 |
+
"zP95_m": 0.0,
|
| 231 |
+
"xyMax_px": 0.0,
|
| 232 |
+
"zMax_m": 0.0,
|
| 233 |
+
"confidenceLossCount": 0
|
| 234 |
+
},
|
| 235 |
+
"body-feet": {
|
| 236 |
+
"joints": 23,
|
| 237 |
+
"xyP95_px": 0.0,
|
| 238 |
+
"zP95_m": 0.0,
|
| 239 |
+
"xyMax_px": 0.0,
|
| 240 |
+
"zMax_m": 0.0,
|
| 241 |
+
"confidenceLossCount": 0
|
| 242 |
+
},
|
| 243 |
+
"hips": {
|
| 244 |
+
"joints": 2,
|
| 245 |
+
"xyP95_px": 0.0,
|
| 246 |
+
"zP95_m": 0.0,
|
| 247 |
+
"xyMax_px": 0.0,
|
| 248 |
+
"zMax_m": 0.0,
|
| 249 |
+
"confidenceLossCount": 0
|
| 250 |
+
},
|
| 251 |
+
"knees": {
|
| 252 |
+
"joints": 2,
|
| 253 |
+
"xyP95_px": 0.0,
|
| 254 |
+
"zP95_m": 0.0,
|
| 255 |
+
"xyMax_px": 0.0,
|
| 256 |
+
"zMax_m": 0.0,
|
| 257 |
+
"confidenceLossCount": 0
|
| 258 |
+
},
|
| 259 |
+
"ankles": {
|
| 260 |
+
"joints": 2,
|
| 261 |
+
"xyP95_px": 0.0,
|
| 262 |
+
"zP95_m": 0.0,
|
| 263 |
+
"xyMax_px": 0.0,
|
| 264 |
+
"zMax_m": 0.0,
|
| 265 |
+
"confidenceLossCount": 0
|
| 266 |
+
},
|
| 267 |
+
"feet": {
|
| 268 |
+
"joints": 6,
|
| 269 |
+
"xyP95_px": 0.0,
|
| 270 |
+
"zP95_m": 0.0,
|
| 271 |
+
"xyMax_px": 0.0,
|
| 272 |
+
"zMax_m": 0.0,
|
| 273 |
+
"confidenceLossCount": 0
|
| 274 |
+
},
|
| 275 |
+
"face": {
|
| 276 |
+
"joints": 68,
|
| 277 |
+
"xyP95_px": 0.0,
|
| 278 |
+
"zP95_m": 0.0,
|
| 279 |
+
"xyMax_px": 0.0,
|
| 280 |
+
"zMax_m": 0.0,
|
| 281 |
+
"confidenceLossCount": 0
|
| 282 |
+
},
|
| 283 |
+
"hands": {
|
| 284 |
+
"joints": 42,
|
| 285 |
+
"xyP95_px": 0.0,
|
| 286 |
+
"zP95_m": 0.0,
|
| 287 |
+
"xyMax_px": 0.0,
|
| 288 |
+
"zMax_m": 0.0,
|
| 289 |
+
"confidenceLossCount": 0
|
| 290 |
+
}
|
| 291 |
+
},
|
| 292 |
+
"scoreMaxAbsoluteError": 4.76837158203125e-07
|
| 293 |
+
}
|
| 294 |
+
],
|
| 295 |
+
"referenceLatencyMs": [],
|
| 296 |
+
"note": "Same trained weights and 133-point 3D contract; explicit im2col and optional batched split-K reduction. CPU conversion tests are not WebGPU performance or target-video accuracy measurements."
|
| 297 |
+
}
|
experiments/rtmw3d-l-fp32-head-gemm.onnx
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b1a60dccb44759a9af3de6d54691c6dc0c9318909722f7baaf2cf781697ccefe
|
| 3 |
+
size 230137728
|