Skip to content

Commit 3d125fa

Browse files
yuwenzhojcwchen
yuwenzho
and
jcwchen
authored
add vgg16, zfnet512 and fcn qdq models (#586)
* add vgg16, zfnet512 and fcn qdq models Signed-off-by: yuwenzho <[email protected]> * reupload tar.gz file Signed-off-by: yuwenzho <[email protected]> --------- Signed-off-by: yuwenzho <[email protected]> Co-authored-by: Chun-Wei Chen <[email protected]>
1 parent da635bc commit 3d125fa

File tree

10 files changed

+165
-3
lines changed

10 files changed

+165
-3
lines changed

‎ONNX_HUB_MANIFEST.json‎

Lines changed: 138 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -5624,6 +5624,48 @@
56245624
"model_with_data_bytes": 106112057
56255625
}
56265626
},
5627+
{
5628+
"model": "VGG 16-qdq",
5629+
"model_path": "vision/classification/vgg/model/vgg16-12-qdq.onnx",
5630+
"onnx_version": "1.9.0",
5631+
"opset_version": 12,
5632+
"metadata": {
5633+
"model_sha": "74ae0fa7e9e5b16782581bf17b5582080d1c06b69ed59410c7b1f190d9328bb1",
5634+
"model_bytes": 138424439,
5635+
"tags": [
5636+
"vision",
5637+
"classification",
5638+
"vgg"
5639+
],
5640+
"io_ports": {
5641+
"inputs": [
5642+
{
5643+
"name": "data",
5644+
"shape": [
5645+
1,
5646+
3,
5647+
224,
5648+
224
5649+
],
5650+
"type": "tensor(float)"
5651+
}
5652+
],
5653+
"outputs": [
5654+
{
5655+
"name": "vgg0_dense2_fwd",
5656+
"shape": [
5657+
1,
5658+
1000
5659+
],
5660+
"type": "tensor(float)"
5661+
}
5662+
]
5663+
},
5664+
"model_with_data_path": "vision/classification/vgg/model/vgg16-12-qdq.tar.gz",
5665+
"model_with_data_sha": "ba8f1a789554deab15d28a3c4e6a1508595c326d36e5d744e8d5acf2729f7621",
5666+
"model_with_data_bytes": 103626016
5667+
}
5668+
},
56275669
{
56285670
"model": "VGG 16-fp32",
56295671
"model_path": "vision/classification/vgg/model/vgg16-12.onnx",
@@ -6115,6 +6157,48 @@
61156157
"model_with_data_bytes": 50270897
61166158
}
61176159
},
6160+
{
6161+
"model": "ZFNet-512-qdq",
6162+
"model_path": "vision/classification/zfnet-512/model/zfnet512-12-qdq.onnx",
6163+
"onnx_version": "1.9",
6164+
"opset_version": 12,
6165+
"metadata": {
6166+
"model_sha": "d29f9fe228a3b1ea4333e69f156faee1075e610721b46730c7ebeeb9ede36727",
6167+
"model_bytes": 87288258,
6168+
"tags": [
6169+
"vision",
6170+
"classification",
6171+
"zfnet-512"
6172+
],
6173+
"io_ports": {
6174+
"inputs": [
6175+
{
6176+
"name": "gpu_0/data_0",
6177+
"shape": [
6178+
1,
6179+
3,
6180+
224,
6181+
224
6182+
],
6183+
"type": "tensor(float)"
6184+
}
6185+
],
6186+
"outputs": [
6187+
{
6188+
"name": "gpu_0/softmax_1",
6189+
"shape": [
6190+
1,
6191+
1000
6192+
],
6193+
"type": "tensor(float)"
6194+
}
6195+
]
6196+
},
6197+
"model_with_data_path": "vision/classification/zfnet-512/model/zfnet512-12-qdq.tar.gz",
6198+
"model_with_data_sha": "b65d9e14088acabfd11efccc7601cad1697b5ab48fe6074c41f15c9f7b5fed8c",
6199+
"model_with_data_bytes": 58038148
6200+
}
6201+
},
61186202
{
61196203
"model": "ZFNet-512",
61206204
"model_path": "vision/classification/zfnet-512/model/zfnet512-12.onnx",
@@ -6874,6 +6958,60 @@
68746958
"model_with_data_bytes": 30117697
68756959
}
68766960
},
6961+
{
6962+
"model": "FCN ResNet-50-qdq",
6963+
"model_path": "vision/object_detection_segmentation/fcn/model/fcn-resnet50-12-qdq.onnx",
6964+
"onnx_version": "1.8.0",
6965+
"opset_version": 12,
6966+
"metadata": {
6967+
"model_sha": "0a6aef19ef5401364aacb6883303401b4b7ccd3e3cd5eb60180b467dae88dcf5",
6968+
"model_bytes": 35440011,
6969+
"tags": [
6970+
"vision",
6971+
"object detection segmentation",
6972+
"fcn"
6973+
],
6974+
"io_ports": {
6975+
"inputs": [
6976+
{
6977+
"name": "input",
6978+
"shape": [
6979+
"batch",
6980+
3,
6981+
"height",
6982+
"width"
6983+
],
6984+
"type": "tensor(float)"
6985+
}
6986+
],
6987+
"outputs": [
6988+
{
6989+
"name": "out",
6990+
"shape": [
6991+
"batch",
6992+
21,
6993+
"height",
6994+
"width"
6995+
],
6996+
"type": "tensor(float)"
6997+
},
6998+
{
6999+
"name": "aux",
7000+
"shape": [
7001+
"batch",
7002+
21,
7003+
"height",
7004+
"width"
7005+
],
7006+
"type": "tensor(float)"
7007+
}
7008+
]
7009+
},
7010+
"model_with_data_path": "vision/object_detection_segmentation/fcn/model/fcn-resnet50-12-qdq.tar.gz",
7011+
"model_with_data_sha": "0feb90c5e17f1fac9c21e17bd536e43aeb56c04adcf87d5eb66acc1ef57fbafe",
7012+
"model_with_data_bytes": 21876642
7013+
}
7014+
},
68777015
{
68787016
"model": "FCN ResNet-50",
68797017
"model_path": "vision/object_detection_segmentation/fcn/model/fcn-resnet50-12.onnx",

‎vision/classification/vgg/README.md‎

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -27,6 +27,7 @@ The models below are variant of same network with different number of layers and
2727
|VGG 19-bn| [548.1 MB](model/vgg19-bn-7.onnx) |[508.6 MB](model/vgg19-bn-7.tar.gz) | 1.2.1 |7 | 73.83 | 91.79 |
2828
|VGG 16-fp32| [527.8 MB](model/vgg16-12.onnx) |[488.2 MB](model/vgg16-12.tar.gz)| 1.9.0 | 12 | 72.38 | 91.00 |
2929
|VGG 16-int8| [132.0 MB](model/vgg16-12-int8.onnx) |[101.1 MB](model/vgg16-12-int8.tar.gz)| 1.9.0 | 12 | 72.32 | 90.97 |
30+
|VGG 16-qdq| [133.0 MB](model/vgg16-12-qdq.onnx) |[99 MB](model/vgg16-12-qdq.tar.gz)| 1.9.0 | 12 | 72.35 | 91.02 |
3031
> Compared with the fp32 VGG 16, int8 VGG 16's Top-1 accuracy drop ratio is 0.06%, Top-5 accuracy drop ratio is 0.03% and performance improvement is 2.31x.
3132
>
3233
> Note the performance depends on the test hardware.
@@ -81,7 +82,7 @@ We used MXNet as framework with gluon APIs to perform training. View the [traini
8182
We used MXNet as framework with gluon APIs to perform validation. Use the notebook [imagenet_validation](../imagenet_validation.ipynb) to verify the accuracy of the model on the validation set. Make sure to specify the appropriate model name in the notebook.
8283

8384
## Quantization
84-
VGG 16-int8 is obtained by quantizing VGG 16-fp32 model. We use [Intel® Neural Compressor](https://github.com/intel/neural-compressor) with onnxruntime backend to perform quantization. View the [instructions](https://github.com/intel/neural-compressor/blob/master/examples/onnxrt/image_recognition/onnx_model_zoo/vgg16/quantization/ptq/README.md) to understand how to use Intel® Neural Compressor for quantization.
85+
VGG 16-int8 and VGG 16-qdq are obtained by quantizing VGG 16-fp32 model. We use [Intel® Neural Compressor](https://github.com/intel/neural-compressor) with onnxruntime backend to perform quantization. View the [instructions](https://github.com/intel/neural-compressor/blob/master/examples/onnxrt/image_recognition/onnx_model_zoo/vgg16/quantization/ptq/README.md) to understand how to use Intel® Neural Compressor for quantization.
8586

8687
### Environment
8788
onnx: 1.9.0
@@ -116,6 +117,7 @@ We use onnxruntime to perform VGG 16-fp32 and VGG 16-int8 inference. View the no
116117
* [abhinavs95](https://github.com/abhinavs95) (Amazon AI)
117118
* [ankkhedia](https://github.com/ankkhedia) (Amazon AI)
118119
* [mengniwang95](https://github.com/mengniwang95) (Intel)
120+
* [yuwenzho](https://github.com/yuwenzho) (Intel)
119121
* [airMeng](https://github.com/airMeng) (Intel)
120122
* [ftian1](https://github.com/ftian1) (Intel)
121123
* [hshen14](https://github.com/hshen14) (Intel)

‎vision/classification/vgg/model/vgg16-12-qdq.onnx‎

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,3 @@
1+
version https://git-lfs.github.com/spec/v1
2+
oid sha256:74ae0fa7e9e5b16782581bf17b5582080d1c06b69ed59410c7b1f190d9328bb1
3+
size 138424439

‎vision/classification/vgg/model/vgg16-12-qdq.tar.gz‎

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,3 @@
1+
version https://git-lfs.github.com/spec/v1
2+
oid sha256:ba8f1a789554deab15d28a3c4e6a1508595c326d36e5d744e8d5acf2729f7621
3+
size 103626016

‎vision/classification/zfnet-512/README.md‎

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -11,6 +11,7 @@
1111
|ZFNet-512| [341 MB](model/zfnet512-9.onnx) | [318 MB](model/zfnet512-9.tar.gz) | 1.4 | 9| | |
1212
|ZFNet-512| [333 MB](model/zfnet512-12.onnx) | [309 MB](model/zfnet512-12.tar.gz) | 1.9 | 12|55.97|79.41|
1313
|ZFNet-512-int8| [83 MB](model/zfnet512-12-int8.onnx) | [48 MB](model/zfnet512-12-int8.tar.gz) | 1.9 | 12|55.84|79.33|
14+
|ZFNet-512-qdq| [84 MB](model/zfnet512-12-qdq.onnx) | [56 MB](model/zfnet512-12-qdq.tar.gz) | 1.9 | 12|55.83|79.42|
1415
> Compared with the fp32 ZFNet-512, int8 ZFNet-512's Top-1 accuracy drop ratio is 0.23%, Top-5 accuracy drop ratio is 0.10% and performance improvement is 1.78x.
1516
>
1617
> **Note**
@@ -52,7 +53,7 @@ random generated sampe test data:
5253
## Results/accuracy on test set
5354

5455
## Quantization
55-
ZFNet-512-int8 is obtained by quantizing fp32 ZFNet-512 model. We use [Intel® Neural Compressor](https://github.com/intel/neural-compressor) with onnxruntime backend to perform quantization. View the [instructions](https://github.com/intel/neural-compressor/blob/master/examples/onnxrt/image_recognition/onnx_model_zoo/zfnet/quantization/ptq/README.md) to understand how to use Intel® Neural Compressor for quantization.
56+
ZFNet-512-int8 and ZFNet-512-qdq are obtained by quantizing fp32 ZFNet-512 model. We use [Intel® Neural Compressor](https://github.com/intel/neural-compressor) with onnxruntime backend to perform quantization. View the [instructions](https://github.com/intel/neural-compressor/blob/master/examples/onnxrt/image_recognition/onnx_model_zoo/zfnet/quantization/ptq/README.md) to understand how to use Intel® Neural Compressor for quantization.
5657

5758
### Environment
5859
onnx: 1.9.0
@@ -80,6 +81,7 @@ bash run_tuning.sh --input_model=path/to/model \ # model path as *.onnx
8081

8182
## Contributors
8283
* [mengniwang95](https://github.com/mengniwang95) (Intel)
84+
* [yuwenzho](https://github.com/yuwenzho) (Intel)
8385
* [airMeng](https://github.com/airMeng) (Intel)
8486
* [ftian1](https://github.com/ftian1) (Intel)
8587
* [hshen14](https://github.com/hshen14) (Intel)

‎vision/classification/zfnet-512/model/zfnet512-12-qdq.onnx‎

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,3 @@
1+
version https://git-lfs.github.com/spec/v1
2+
oid sha256:d29f9fe228a3b1ea4333e69f156faee1075e610721b46730c7ebeeb9ede36727
3+
size 87288258

‎vision/classification/zfnet-512/model/zfnet512-12-qdq.tar.gz‎

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,3 @@
1+
version https://git-lfs.github.com/spec/v1
2+
oid sha256:b65d9e14088acabfd11efccc7601cad1697b5ab48fe6074c41f15c9f7b5fed8c
3+
size 58038148

‎vision/object_detection_segmentation/fcn/README.md‎

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -15,6 +15,7 @@ This specific model detects 20 different [classes](dependencies/voc_classes.txt)
1515
| FCN ResNet-101 | [207 MB](model/fcn-resnet101-11.onnx) | [281 MB](model/fcn-resnet101-11.tar.gz) | 1.8.0 | 11 | 63.7% |
1616
| FCN ResNet-50 | [134 MB](model/fcn-resnet50-12.onnx) | [125 MB](model/fcn-resnet50-12.tar.gz) | 1.8.0 | 12 | 65.0% |
1717
| FCN ResNet-50-int8 | [34 MB](model/fcn-resnet50-12-int8.onnx) | [29 MB](model/fcn-resnet50-12-int8.tar.gz) | 1.8.0 | 12 | 64.7% |
18+
| FCN ResNet-50-qdq | [34 MB](model/fcn-resnet50-12-qdq.onnx) | [21 MB](model/fcn-resnet50-12-qdq.tar.gz) | 1.8.0 | 12 | 64.4% |
1819

1920
### Source
2021

@@ -145,7 +146,7 @@ The more conservative of the two estimates is used in the model files table.
145146
<hr>
146147
147148
## Quantization
148-
FCN ResNet 50-int8 is obtained by quantizing fp32 FCN ResNet 50 model. We use [Intel® Neural Compressor](https://github.com/intel/neural-compressor) with onnxruntime backend to perform quantization. View the [instructions](https://github.com/intel/neural-compressor/blob/master/examples/onnxrt/image_recognition/onnx_model_zoo/fcn/quantization/ptq/README.md) to understand how to use Intel® Neural Compressor for quantization.
149+
FCN ResNet 50-int8 and FCN ResNet-50-qdq are obtained by quantizing fp32 FCN ResNet 50 model. We use [Intel® Neural Compressor](https://github.com/intel/neural-compressor) with onnxruntime backend to perform quantization. View the [instructions](https://github.com/intel/neural-compressor/blob/master/examples/onnxrt/image_recognition/onnx_model_zoo/fcn/quantization/ptq/README.md) to understand how to use Intel® Neural Compressor for quantization.
149150

150151
### Environment
151152
onnx: 1.9.0
@@ -176,6 +177,7 @@ bash run_tuning.sh --input_model=path/to/model \ # model path as *.onnx
176177
## Contributors
177178
* [Jack Duvall](https://github.com/duvallj)
178179
* [mengniwang95](https://github.com/mengniwang95) (Intel)
180+
* [yuwenzho](https://github.com/yuwenzho) (Intel)
179181
* [airMeng](https://github.com/airMeng) (Intel)
180182
* [ftian1](https://github.com/ftian1) (Intel)
181183
* [hshen14](https://github.com/hshen14) (Intel)

‎vision/object_detection_segmentation/fcn/model/fcn-resnet50-12-qdq.onnx‎

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,3 @@
1+
version https://git-lfs.github.com/spec/v1
2+
oid sha256:0a6aef19ef5401364aacb6883303401b4b7ccd3e3cd5eb60180b467dae88dcf5
3+
size 35440011

‎vision/object_detection_segmentation/fcn/model/fcn-resnet50-12-qdq.tar.gz‎

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,3 @@
1+
version https://git-lfs.github.com/spec/v1
2+
oid sha256:0feb90c5e17f1fac9c21e17bd536e43aeb56c04adcf87d5eb66acc1ef57fbafe
3+
size 21876642

0 commit comments

Comments
 (0)