msluszniak commited on
Commit
4ca25fa
·
1 Parent(s): ceb6a91

QNN: re-export a16w8 with MinMaxObserver activations

Browse files

The previous files used make_quantizer's MovingAverageMinMaxObserver default,
whose moving average trails the real activation range. Measured on an S26 Ultra
over 500 ImageNetV2 images: 70.00% top-1 before, 71.80% after, against 72.20%
for fp32. Quantization loss drops from 2.20 points to 0.40 at the same size.

Only the v81 file is device-validated; the other four are the same code path and
calibration compiled for Hexagon versions we have no hardware for.

qnn/config.json CHANGED
@@ -34,7 +34,9 @@
34
  }
35
  ]
36
  }
37
- }
 
 
38
  },
39
  {
40
  "file": "efficientnet_v2_s_qnn_a16w8_v73.pte",
@@ -62,7 +64,9 @@
62
  }
63
  ]
64
  }
65
- }
 
 
66
  },
67
  {
68
  "file": "efficientnet_v2_s_qnn_a16w8_v75.pte",
@@ -90,7 +94,9 @@
90
  }
91
  ]
92
  }
93
- }
 
 
94
  },
95
  {
96
  "file": "efficientnet_v2_s_qnn_a16w8_v79.pte",
@@ -118,7 +124,9 @@
118
  }
119
  ]
120
  }
121
- }
 
 
122
  },
123
  {
124
  "file": "efficientnet_v2_s_qnn_a16w8_v81.pte",
@@ -146,7 +154,9 @@
146
  }
147
  ]
148
  }
149
- }
 
 
150
  }
151
  ]
152
  }
 
34
  }
35
  ]
36
  }
37
+ },
38
+ "quantized": true,
39
+ "default": false
40
  },
41
  {
42
  "file": "efficientnet_v2_s_qnn_a16w8_v73.pte",
 
64
  }
65
  ]
66
  }
67
+ },
68
+ "quantized": true,
69
+ "default": false
70
  },
71
  {
72
  "file": "efficientnet_v2_s_qnn_a16w8_v75.pte",
 
94
  }
95
  ]
96
  }
97
+ },
98
+ "quantized": true,
99
+ "default": false
100
  },
101
  {
102
  "file": "efficientnet_v2_s_qnn_a16w8_v79.pte",
 
124
  }
125
  ]
126
  }
127
+ },
128
+ "quantized": true,
129
+ "default": false
130
  },
131
  {
132
  "file": "efficientnet_v2_s_qnn_a16w8_v81.pte",
 
154
  }
155
  ]
156
  }
157
+ },
158
+ "quantized": true,
159
+ "default": true
160
  }
161
  ]
162
  }
qnn/efficientnet_v2_s_qnn_a16w8_v69.pte CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d449e33e881d41f18ffc132c9c689f890141aa77742c32b7b8e918645b5b7ea4
3
  size 32839168
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:96552e464af2c400dc6e816be977afa3e81001ed8eb353e83d8e4b651be4f50c
3
  size 32839168
qnn/efficientnet_v2_s_qnn_a16w8_v73.pte CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:ac2c1e554c2f178051a96d07978520ef5165b1f0fcf67e29940de92d21f3d700
3
  size 25556480
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e044a5f69bcbd7c438524d476288fe083371c05e48b9f973d40fe5c7e592e127
3
  size 25556480
qnn/efficientnet_v2_s_qnn_a16w8_v75.pte CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:fd0e960b3547f7e1bac16e4ed59b72d117ffcb235b19d5fc79ac2dc97c4ec8a0
3
- size 25556480
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e0b1b7521626f263e9775b00a1cdd55a41ae08b4838678d2dd6305959f2d38f5
3
+ size 25552384
qnn/efficientnet_v2_s_qnn_a16w8_v79.pte CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:36cd725ef001a4cbee99120db0b271512e593383ef67f4d3236eb5b67af12680
3
- size 25646592
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e476c80bf2ccbc69cac20f3ae889e814de9809300a8a59e7f93f169b34a3cec7
3
+ size 25642496
qnn/efficientnet_v2_s_qnn_a16w8_v81.pte CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:f192c8acb92a63d6403149d90c0288ee1d34fd9f4b80a46b3cc90d1bca0b4d80
3
  size 25712128
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1ffb1d663490f59a5f14611a6c335dcc284eac7e6cf65ab47fb473b06fc08005
3
  size 25712128