add int8 efficientnet, inceptionv1, ssdmbv1, ssd, fcn_rn50 (#495)

* add int8 efficientnet, inceptionv1, ssdmbv1, ssd, fcn_rn50

Signed-off-by: mengniwa <[email protected]>

* fix model path

Signed-off-by: mengniwa <[email protected]>

Author: mengniwang95

vision/classification/efficientnet-lite4/README.md

@@ -10,13 +10,19 @@ EfficientNet-Lite 4 is the largest variant and most accurate of the set of Effic
 
 ## Model
 
- |Model        |Download | Download (with sample test data)|ONNX version|Opset version|
-|-------------|:--------------|:--------------|:--------------|:--------------|
-|EfficientNet-Lite4     | [51.9 MB](model/efficientnet-lite4-11.onnx)	  | [48.6 MB](model/efficientnet-lite4-11.tar.gz)|1.7.0|11|
-
+ |Model        |Download | Download (with sample test data)|ONNX version|Opset version|Top-1 accuracy (%)|
+|-------------|:--------------|:--------------|:--------------|:--------------|:--------------|
+|EfficientNet-Lite4     | [51.9 MB](model/efficientnet-lite4-11.onnx)	  | [48.6 MB](model/efficientnet-lite4-11.tar.gz)|1.7.0|11|80.4|
+|EfficientNet-Lite4-int8     | [13.0 MB](model/efficientnet-lite4-11-int8.onnx)	  | [12.2 MB](model/efficientnet-lite4-11-int8.tar.gz)|1.9.0|11|77.56|
+> The fp32 Top-1 accuracy got by [Intel® Neural Compressor](https://github.com/intel/neural-compressor) is 77.70%, and compared with this value, int8 EfficientNet-Lite4's Top-1 accuracy drop ratio is 0.18% and performance improvement is 1.12x.
+>
+> **Note** 
+>
+> The performance depends on the test hardware. Performance data here is collected with Intel® Xeon® Platinum 8280 Processor, 1s 4c per instance, CentOS Linux 8.3, data batch size is 1.
 
 ### Source
 Tensorflow EfficientNet-Lite4 => ONNX EfficientNet-Lite4
+ONNX EfficientNet-Lite4 => Quantized ONNX EfficientNet-Lite4
 
 <hr>
 
@@ -132,14 +138,43 @@ The model was trained using [COCO 2017 Train Images, Val Images, and Train/Val a
 Refer to [efficientnet-lite4 conversion notebook](https://github.com/onnx/tensorflow-onnx/blob/master/tutorials/efficientnet-lite.ipynb) for details of how to use it and reproduce accuracy.
 <hr>
 
+## Quantization
+CaffeNet-int8 is obtained by quantizing fp32 CaffeNet 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/efficientnet/quantization/ptq/README.md) to understand how to use Intel® Neural Compressor for quantization.
+
+### Environment
+onnx: 1.9.0 
+onnxruntime: 1.8.0
+
+### Prepare model
+```shell
+wget https://github.com/onnx/models/raw/master/vision/classification/efficientnet-lite4/model/efficientnet-lite4-11.onnx
+```
+
+### Model quantize
+Make sure to specify the appropriate dataset path in the configuration file.
+```bash
+bash run_tuning.sh --input_model=path/to/model \  # model path as *.onnx
+                   --config=efficientnet.yaml \
+                   --output_model=path/to/save
+```
+<hr>
+
 ## References
-Tensorflow to Onnx conversion [tutorial](https://github.com/onnx/tensorflow-onnx/blob/master/tutorials/efficientnet-lite.ipynb). The Juypter Notebook references how to run an evaluation on the efficientnet-lite4 model and export it as a saved model. It also details how to convert the tensorflow model into onnx, and how to run its preprocessing and postprocessing code for the inputs and outputs.
+* Tensorflow to Onnx conversion [tutorial](https://github.com/onnx/tensorflow-onnx/blob/master/tutorials/efficientnet-lite.ipynb). The Juypter Notebook references how to run an evaluation on the efficientnet-lite4 model and export it as a saved model. It also details how to convert the tensorflow model into onnx, and how to run its preprocessing and postprocessing code for the inputs and outputs.
+
+* Refer to this [paper](https://arxiv.org/abs/1905.11946) for more details on the model.
+
+* [Intel® Neural Compressor](https://github.com/intel/neural-compressor)
 
-Refer to this [paper](https://arxiv.org/abs/1905.11946) for more details on the model.
 <hr>
 
 ## Contributors
- [Shirley Su](https://github.com/shirleysu8)
+* [Shirley Su](https://github.com/shirleysu8)
+* [mengniwang95](https://github.com/mengniwang95) (Intel)
+* [airMeng](https://github.com/airMeng) (Intel)
+* [ftian1](https://github.com/ftian1) (Intel)
+* [hshen14](https://github.com/hshen14) (Intel)
+
 <hr>
 
 ## License

vision/classification/efficientnet-lite4/model/efficientnet-lite4-11-int8.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b3cbb5077262b20df565dacddecb3724c0976c35029a87e512d13aa4eff04a2
+size 13585963

vision/classification/efficientnet-lite4/model/efficientnet-lite4-11-int8.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3b03546fb343ba48e3243c893ea37e47cb1f337d80ed0aca562077d8d3da51b6
+size 12780764

vision/classification/inception_and_googlenet/inception_v1/README.md

@@ -2,25 +2,31 @@
 
 # Inception v1
 
-|Model        |Download  |Download (with sample test data)| ONNX version |Opset version|
-| ------------- | ------------- | ------------- | ------------- | ------------- |
-|Inception-1| [28 MB](model/inception-v1-3.onnx)  |  [29 MB](model/inception-v1-3.tar.gz) |  1.1 | 3|
-|Inception-1| [28 MB](model/inception-v1-6.onnx)  |  [29 MB](model/inception-v1-6.tar.gz) |  1.1.2 | 6|
-|Inception-1| [28 MB](model/inception-v1-7.onnx)  |  [29 MB](model/inception-v1-7.tar.gz) |  1.2 | 7|
-|Inception-1| [28 MB](model/inception-v1-8.onnx)  |  [29 MB](model/inception-v1-8.tar.gz) |  1.3 | 8|
-|Inception-1| [28 MB](model/inception-v1-9.onnx)  |  [29 MB](model/inception-v1-9.tar.gz) |  1.4 | 9|
+|Model        |Download  |Download (with sample test data)| ONNX version |Opset version| Top-1 accuracy (%)|
+| ------------- | ------------- | ------------- | ------------- | ------------- |------------- |
+|Inception-1| [28 MB](model/inception-v1-3.onnx)  |  [29 MB](model/inception-v1-3.tar.gz) |  1.1 | 3| |
+|Inception-1| [28 MB](model/inception-v1-6.onnx)  |  [29 MB](model/inception-v1-6.tar.gz) |  1.1.2 | 6| |
+|Inception-1| [28 MB](model/inception-v1-7.onnx)  |  [29 MB](model/inception-v1-7.tar.gz) |  1.2 | 7| |
+|Inception-1| [28 MB](model/inception-v1-8.onnx)  |  [29 MB](model/inception-v1-8.tar.gz) |  1.3 | 8| |
+|Inception-1| [28 MB](model/inception-v1-9.onnx)  |  [29 MB](model/inception-v1-9.tar.gz) |  1.4 | 9| |
+|Inception-1| [27 MB](model/inception-v1-12.onnx)  |  [25 MB](model/inception-v1-12.tar.gz) |  1.9 | 12| 67.23|
+|Inception-1-int8| [10 MB](model/inception-v1-12-int8.onnx)  |  [9 MB](model/inception-v1-12-int8.tar.gz) |  1.9 | 12| 67.24|
+> Compared with the fp32 Inception-1, int8 Inception-1's Top-1 accuracy drop ratio is -0.01% and performance improvement is 1.26x.
+>
+> **Note** 
+> 
+> The performance depends on the test hardware. Performance data here is collected with Intel® Xeon® Platinum 8280 Processor, 1s 4c per instance, CentOS Linux 8.3, data batch size is 1.
+
 
 ## Description
 Inception v1 is a reproduction of GoogLeNet.
 
-### Paper
-[Going deeper with convolutions](https://arxiv.org/abs/1409.4842)
-
 ### Dataset
 [ILSVRC2012](http://www.image-net.org/challenges/LSVRC/2012/)
 
 ## Source
 Caffe2 Inception v1 ==> ONNX Inception v1
+ONNX Inception v1 ==> Quantized ONNX Inception v1
 
 ## Model input and output
 ### Input
@@ -44,5 +50,39 @@ random generated sampe test data:
 
 ## Results/accuracy on test set
 
+## Quantization
+Inception-1-int8 is obtained by quantizing fp32 Inception-1 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/inception/quantization/ptq/README.md) to understand how to use Intel® Neural Compressor for quantization.
+
+### Environment
+onnx: 1.9.0 
+onnxruntime: 1.8.0
+
+### Prepare model
+```shell
+wget https://github.com/onnx/models/blob/master/vision/classification/inception_and_googlenet/inception_v1/model/inception-v1-12.onnx
+```
+
+### Model quantize
+Make sure to specify the appropriate dataset path in the configuration file.
+```bash
+bash run_tuning.sh --input_model=path/to/model \  # model path as *.onnx
+                   --config=inception_v1.yaml \
+                   --data_path=/path/to/imagenet \
+                   --label_path=/path/to/imagenet/label \
+                   --output_model=path/to/save
+```
+
+## References
+* [Going deeper with convolutions](https://arxiv.org/abs/1409.4842)
+
+* [Intel® Neural Compressor](https://github.com/intel/neural-compressor)
+
+
+## Contributors
+* [mengniwang95](https://github.com/mengniwang95) (Intel)
+* [airMeng](https://github.com/airMeng) (Intel)
+* [ftian1](https://github.com/ftian1) (Intel)
+* [hshen14](https://github.com/hshen14) (Intel)
+
 ## License
 MIT

vision/classification/inception_and_googlenet/inception_v1/model/inception-v1-12-int8.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d25f0bfa7a92e9a67d8049facd4454911ca4388a303fc61ee14664b1010c717
+size 10191535

vision/classification/inception_and_googlenet/inception_v1/model/inception-v1-12-int8.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:296e49e0f1253877053d5aa48f4192f675b302b5828e70a821d6d94e0c8ee4e3
+size 9463738

vision/classification/inception_and_googlenet/inception_v1/model/inception-v1-12.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:814b31a6316873bd03d2d808415ce1f40a7bf48ad88eab5ecc3cdc8dc6cf38ec
+size 28021958

vision/classification/inception_and_googlenet/inception_v1/model/inception-v1-12.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4198ed3415d59dacab2f6fa90c81f1a3aa9f3623b66e2d9653f16d06a5cc68c
+size 26545542

vision/object_detection_segmentation/fcn/README.md

@@ -13,11 +13,14 @@ This specific model detects 20 different [classes](dependencies/voc_classes.txt)
 |----------------|:--------------------------------------|:----------------------------------------|:-------------|:--------------|:--------|
 | FCN ResNet-50  | [134 MB](model/fcn-resnet50-11.onnx)  | [213 MB](model/fcn-resnet50-11.tar.gz)  | 1.8.0        | 11 | 60.5% |
 | FCN ResNet-101 | [207 MB](model/fcn-resnet101-11.onnx) | [281 MB](model/fcn-resnet101-11.tar.gz) | 1.8.0        | 11 | 63.7% |
+| FCN ResNet-50  | [134 MB](model/fcn-resnet50-12.onnx)  | [125 MB](model/fcn-resnet50-12.tar.gz)  | 1.8.0        | 12 | 60.5% |
+| 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% |
 
 ### Source
 
  * PyTorch Torchvision FCN ResNet50 ==> ONNX FCN ResNet50
  * PyTorch Torchvision FCN ResNet101 ==> ONNX FCN ResNet101
+ * ONNX FCN ResNet50 ==> Quantized ONNX FCN ResNet50
 
 ## Inference
 
@@ -129,18 +132,53 @@ If you have the [COCO val2017 dataset](https://cocodataset.org/#download) downlo
 | Model          | mean IoU | global pixelwise accuracy |
 |----------------|:---------|:--------------------------|
 | FCN ResNet 50  | 65.0     | 99.6                      |
+| FCN ResNet 50-int8 | 64.7 | 99.5                      |
 | FCN ResNet 101 | 66.7     | 99.6                      |
 
 The more conservative of the two estimates is used in the model files table.
+
+> Compared with the fp32 FCN ResNet 50, FCN ResNet 50-int8's mean IoU drop ratio is 0.46% global pixelwise accuracy drop ratio is 0.10% and performance improvement is 1.28x.
+>
+> **Note**
+>
+>The performance depends on the test hardware. Performance data here is collected with Intel® Xeon® Platinum 8280 Processor, 1s 4c per inst ance, CentOS Linux 8.3, data batch size is 1.
 <hr>
 
+## Quantization
+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.
+
+### Environment
+onnx: 1.9.0 
+onnxruntime: 1.8.0
+
+### Prepare model
+```shell
+wget https://github.com/onnx/models/raw/master/vision/object_detection_segmentation/fcn/model/fcn-resnet50-12.onnx
+```
+
+### Model quantize
+Make sure to specify the appropriate dataset path in the configuration file.
+```bash
+bash run_tuning.sh --input_model=path/to/model  \ # model path as *.onnx
+                   --config=fcn_rn50.yaml \ 
+                   --data_path=path/to/coco/val2017 \
+                   --label_path=path/to/coco/annotations/instances_val2017.json \
+                   --output_model=path/to/save
+```
+
 ## References
-Jonathan Long, Evan Shelhamer, Trevor Darrell; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 3431-3440
+* Jonathan Long, Evan Shelhamer, Trevor Darrell; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015, pp. 3431-3440
+
+* This model is converted from the [Torchvision Model Zoo](https://pytorch.org/docs/stable/torchvision/models.html), originally implemented by Francisco Moss [here](https://github.com/pytorch/vision/tree/master/torchvision/models/segmentation/fcn.py).
 
-This model is converted from the [Torchvision Model Zoo](https://pytorch.org/docs/stable/torchvision/models.html), originally implemented by Francisco Moss [here](https://github.com/pytorch/vision/tree/master/torchvision/models/segmentation/fcn.py).
+* [Intel® Neural Compressor](https://github.com/intel/neural-compressor)
 
 ## Contributors
-[Jack Duvall](https://github.com/duvallj)
+* [Jack Duvall](https://github.com/duvallj)
+* [mengniwang95](https://github.com/mengniwang95) (Intel)
+* [airMeng](https://github.com/airMeng) (Intel)
+* [ftian1](https://github.com/ftian1) (Intel)
+* [hshen14](https://github.com/hshen14) (Intel)
 
 ## License
 MIT License

vision/object_detection_segmentation/fcn/model/fcn-resnet50-12-int8.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:67fcba8ebd435c2fb6b297d2e7c084f3c3b9ae6bc7ae41453b748ae4ed8c0331
+size 35545058