[GSoC] Add block quantized models (#270)

DaniAffCH · web-flow · commit 25f423d0e04c · 2024-11-06T20:54:36.000+08:00
* Gemm and MatMul block quantization support

* refactoring

* fix indentation

* node name independent

* Block quantization tool:
- constant weight category supported
- add data type saturation
- handled the case in which all the elements within a block are the same

benchmark script modified to support block quantized models

block quantized some models

* add missing block quantized models

* formatting

* add blocked models to eval script. Evaluation yunet

* Add sface and pphumanseg evaluation, block quantization tool fix, handpose blocked model fix, removed blocked CRNN EN,

* changed evaluation metric in block_quantize script and add verbose mode

* Add evaluation for PP-ResNet and Mobilenet

* changed file suffix and update readmes

* renamed int8bq
diff --git a/benchmark/README.md b/benchmark/README.md
@@ -26,7 +26,7 @@ python benchmark.py --cfg ./config/face_detection_yunet.yaml
 # All configs
 python benchmark.py --all
 
-# All configs but only fp32 models (--fp32, --fp16, --int8 are available for now)
+# All configs but only fp32 models (--fp32, --fp16, --int8 --int8bq are available for now)
 python benchmark.py --all --fp32
 
 # All configs but exclude some of them (fill with config name keywords, not sensitive to upper/lower case, seperate with colons)
diff --git a/benchmark/benchmark.py b/benchmark/benchmark.py
@@ -46,6 +46,7 @@
 parser.add_argument("--fp32", action="store_true", help="Benchmark models of float32 precision only.")
 parser.add_argument("--fp16", action="store_true", help="Benchmark models of float16 precision only.")
 parser.add_argument("--int8", action="store_true", help="Benchmark models of int8 precision only.")
+parser.add_argument("--int8bq", action="store_true", help="Benchmark models of blocked int8 precision only.")
 parser.add_argument("--all", action="store_true", help="Benchmark all models")
 args = parser.parse_args()
 
@@ -194,15 +195,17 @@ def printResults(self, model_name, model_path):
         model_handler, model_paths = MODELS.get(model_config.pop('name'))
 
         _model_paths = []
-        if args.fp32 or args.fp16 or args.int8:
+        if args.fp32 or args.fp16 or args.int8 or args.int8bq:
             if args.fp32:
                 _model_paths += model_paths['fp32']
             if args.fp16:
                 _model_paths += model_paths['fp16']
             if args.int8:
                 _model_paths += model_paths['int8']
+            if args.int8bq:
+                _model_paths += model_paths['int8bq']
         else:
-            _model_paths = model_paths['fp32'] + model_paths['fp16'] + model_paths['int8']
+            _model_paths = model_paths['fp32'] + model_paths['fp16'] + model_paths['int8'] + model_paths["int8bq"]
         # filter out excluded models
         excludes = []
         if args.model_exclude is not None:
diff --git a/models/__init__.py b/models/__init__.py
@@ -46,6 +46,7 @@ def register(self, item):
         fp32_model_paths = []
         fp16_model_paths = []
         int8_model_paths = []
+        int8bq_model_paths = []
         # onnx
         ret_onnx = sorted(glob.glob(os.path.join(model_dir, "*.onnx")))
         if "object_tracking" in item.__module__:
@@ -57,6 +58,8 @@ def register(self, item):
                     int8_model_paths.append([r])
                 elif "fp16" in r: # exclude fp16 for now
                     fp16_model_paths.append([r])
+                elif "blocked" in r:
+                    int8bq_model_paths.append([r])
                 else:
                     fp32_model_paths.append([r])
         # caffe
@@ -72,6 +75,7 @@ def register(self, item):
             fp32=fp32_model_paths,
             fp16=fp16_model_paths,
             int8=int8_model_paths,
+            int8bq=int8bq_model_paths
         )
 
         self._dict[item.__name__] = (item, all_model_paths)
diff --git a/models/face_detection_yunet/README.md b/models/face_detection_yunet/README.md
@@ -8,15 +8,20 @@ Notes:
 - This model can detect **faces of pixels between around 10x10 to 300x300** due to the training scheme.
 - For details on training this model, please visit https://github.com/ShiqiYu/libfacedetection.train.
 - This ONNX model has fixed input shape, but OpenCV DNN infers on the exact shape of input image. See https://github.com/opencv/opencv_zoo/issues/44 for more information.
+- `face_detection_yunet_2023mar_int8bq.onnx` represents the block-quantized version in int8 precision and is generated using [block_quantize.py](../../tools/quantize/block_quantize.py) with `block_size=64`.
 
 Results of accuracy evaluation with [tools/eval](../../tools/eval).
 
 | Models      | Easy AP | Medium AP | Hard AP |
 | ----------- | ------- | --------- | ------- |
-| YuNet       | 0.8871  | 0.8710    | 0.7681  |
-| YuNet quant | 0.8838  | 0.8683    | 0.7676  |
+| YuNet       | 0.8844  | 0.8656    | 0.7503  |
+| YuNet block | 0.8845  | 0.8652    | 0.7504  |
+| YuNet quant | 0.8810  | 0.8629    | 0.7503  |
+
 
 \*: 'quant' stands for 'quantized'.
+\*\*: 'block' stands for 'blockwise quantized'.
+
 
 ## Demo
 
diff --git a/models/face_detection_yunet/face_detection_yunet_2023mar_int8bq.onnx b/models/face_detection_yunet/face_detection_yunet_2023mar_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:49f000ec501fef24739071fc7e68267d32209045b6822c0c72dce1da25726f10
+size 122489
diff --git a/models/face_recognition_sface/README.md b/models/face_recognition_sface/README.md
@@ -8,15 +8,18 @@ Note:
 - Model files encode MobileFaceNet instances trained on the SFace loss function, see the [SFace paper](https://arxiv.org/abs/2205.12010) for reference.
 - ONNX file conversions from [original code base](https://github.com/zhongyy/SFace) thanks to [Chengrui Wang](https://github.com/crywang).
 - (As of Sep 2021) Supporting 5-landmark warping for now, see below for details.
+- `face_recognition_sface_2021dec_int8bq.onnx` represents the block-quantized version in int8 precision and is generated using [block_quantize.py](../../tools/quantize/block_quantize.py) with `block_size=64`.
 
 Results of accuracy evaluation with [tools/eval](../../tools/eval).
 
 | Models      | Accuracy |
 | ----------- | -------- |
 | SFace       | 0.9940   |
+| SFace block | 0.9942   |
 | SFace quant | 0.9932   |
 
 \*: 'quant' stands for 'quantized'.
+\*\*: 'block' stands for 'blockwise quantized'.
 
 ## Demo
 
diff --git a/models/face_recognition_sface/face_recognition_sface_2021dec_int8bq.onnx b/models/face_recognition_sface/face_recognition_sface_2021dec_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fb143eea07838aa532d1c95df5f69899974ea0140e1fba05e94204be13ed74ee
+size 10667852
diff --git a/models/facial_expression_recognition/README.md b/models/facial_expression_recognition/README.md
@@ -7,6 +7,7 @@ Note:
 - Progressive Teacher is contributed by [Jing Jiang](https://scholar.google.com/citations?user=OCwcfAwAAAAJ&hl=zh-CN).
 -  [MobileFaceNet](https://link.springer.com/chapter/10.1007/978-3-319-97909-0_46) is used as the backbone and the model is able to classify seven basic facial expressions (angry, disgust, fearful, happy, neutral, sad, surprised).
 - [facial_expression_recognition_mobilefacenet_2022july.onnx](https://github.com/opencv/opencv_zoo/raw/master/models/facial_expression_recognition/facial_expression_recognition_mobilefacenet_2022july.onnx) is implemented thanks to [Chengrui Wang](https://github.com/crywang).
+- `facial_expression_recognition_mobilefacenet_2022july_int8bq.onnx` represents the block-quantized version in int8 precision and is generated using [block_quantize.py](../../tools/quantize/block_quantize.py) with `block_size=64`.
 
 Results of accuracy evaluation on [RAF-DB](http://whdeng.cn/RAF/model1.html).
 
diff --git a/models/facial_expression_recognition/facial_expression_recognition_mobilefacenet_2022july_int8bq.onnx b/models/facial_expression_recognition/facial_expression_recognition_mobilefacenet_2022july_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c3eaf9d0a7d442c0aa3beb3234243e1cdff9ad8871fb3cec346e90874caf57d
+size 1376702
diff --git a/models/handpose_estimation_mediapipe/README.md b/models/handpose_estimation_mediapipe/README.md
@@ -14,6 +14,7 @@ This model is converted from TFlite to ONNX using following tools:
 **Note**:
 - The int8-quantized model may produce invalid results due to a significant drop of accuracy.
 - Visit https://github.com/google/mediapipe/blob/master/docs/solutions/models.md#hands for models of larger scale.
+- `handpose_estimation_mediapipe_2023feb_int8bq.onnx` represents the block-quantized version in int8 precision and is generated using [block_quantize.py](../../tools/quantize/block_quantize.py) with `block_size=64`.
 
 ## Demo
 
diff --git a/models/handpose_estimation_mediapipe/handpose_estimation_mediapipe_2023feb_int8bq.onnx b/models/handpose_estimation_mediapipe/handpose_estimation_mediapipe_2023feb_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d7e63ed33e39c8b532a04a5466ac68c8680981637df288bcf93286bec08befbd
+size 1222348
diff --git a/models/human_segmentation_pphumanseg/README.md b/models/human_segmentation_pphumanseg/README.md
@@ -2,6 +2,9 @@
 
 This model is ported from [PaddleHub](https://github.com/PaddlePaddle/PaddleHub) using [this script from OpenCV](https://github.com/opencv/opencv/blob/master/samples/dnn/dnn_model_runner/dnn_conversion/paddlepaddle/paddle_humanseg.py).
 
+**Note**:
+- `human_segmentation_pphumanseg_2023mar_int8bq.onnx` represents the block-quantized version in int8 precision and is generated using [block_quantize.py](../../tools/quantize/block_quantize.py) with `block_size=64`.
+
 ## Demo
 
 ### Python
@@ -46,11 +49,13 @@ Results of accuracy evaluation with [tools/eval](../../tools/eval).
 
 | Models             | Accuracy       | mIoU          |
 | ------------------ | -------------- | ------------- |
-| PPHumanSeg         | 0.9581         | 0.8996        |
-| PPHumanSeg quant   | 0.4365         | 0.2788        |
+| PPHumanSeg         | 0.9656         | 0.9164        |
+| PPHumanSeg block   | 0.9655         | 0.9162        |
+| PPHumanSeg quant   | 0.7285         | 0.3642        |
 
 
 \*: 'quant' stands for 'quantized'.
+\*\*: 'block' stands for 'blockwise quantized'.
 
 ---
 ## License
diff --git a/models/human_segmentation_pphumanseg/human_segmentation_pphumanseg_2023mar_int8bq.onnx b/models/human_segmentation_pphumanseg/human_segmentation_pphumanseg_2023mar_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1da023e95e3adbeef16ac2862767b53b86a743ff26a34692e0833d7e088f0231
+size 1734724
diff --git a/models/image_classification_mobilenet/README.md b/models/image_classification_mobilenet/README.md
@@ -4,16 +4,22 @@ MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applicatio
 
 MobileNetV2: Inverted Residuals and Linear Bottlenecks
 
+**Note**:
+- `image_classification_mobilenetvX_2022apr_int8bq.onnx` represents the block-quantized version in int8 precision and is generated using [block_quantize.py](../../tools/quantize/block_quantize.py) with `block_size=64`.
+
 Results of accuracy evaluation with [tools/eval](../../tools/eval).
 
 | Models             | Top-1 Accuracy | Top-5 Accuracy |
 | ------------------ | -------------- | -------------- |
 | MobileNet V1       | 67.64          | 87.97          |
+| MobileNet V1 block | 67.21          | 87.62          |
 | MobileNet V1 quant | 55.53          | 78.74          |
 | MobileNet V2       | 69.44          | 89.23          |
+| MobileNet V2 block | 68.66          | 88.90          |
 | MobileNet V2 quant | 68.37          | 88.56          |
 
 \*: 'quant' stands for 'quantized'.
+\*\*: 'block' stands for 'blockwise quantized'.
 
 ## Demo
 
diff --git a/models/image_classification_mobilenet/image_classification_mobilenetv1_2022apr_int8bq.onnx b/models/image_classification_mobilenet/image_classification_mobilenetv1_2022apr_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4437385b4908011f6e3019ae253031272c19522ae1cbdff8374bceffe262a5ee
+size 4599388
diff --git a/models/image_classification_mobilenet/image_classification_mobilenetv2_2022apr_int8bq.onnx b/models/image_classification_mobilenet/image_classification_mobilenetv2_2022apr_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3fe457eda49d71b664918ae87349bdb3e3815a56bb82cdb7f477bc83597f4313
+size 3872948
diff --git a/models/image_classification_ppresnet/README.md b/models/image_classification_ppresnet/README.md
@@ -4,14 +4,19 @@ Deep Residual Learning for Image Recognition
 
 This model is ported from [PaddleHub](https://github.com/PaddlePaddle/PaddleHub) using [this script from OpenCV](https://github.com/opencv/opencv/blob/master/samples/dnn/dnn_model_runner/dnn_conversion/paddlepaddle/paddle_resnet50.py).
 
+**Note**:
+- `image_classification_ppresnet50_2022jan_int8bq.onnx` represents the block-quantized version in int8 precision and is generated using [block_quantize.py](../../tools/quantize/block_quantize.py) with `block_size=64`.
+
 Results of accuracy evaluation with [tools/eval](../../tools/eval).
 
 | Models          | Top-1 Accuracy | Top-5 Accuracy |
 | --------------- | -------------- | -------------- |
 | PP-ResNet       | 82.28          | 96.15          |
+| PP-ResNet block | 82.27          | 96.15          |
 | PP-ResNet quant | 0.22           | 0.96           |
 
 \*: 'quant' stands for 'quantized'.
+\*\*: 'block' stands for 'blockwise quantized'.
 
 ## Demo
 
diff --git a/models/image_classification_ppresnet/image_classification_ppresnet50_2022jan_int8bq.onnx b/models/image_classification_ppresnet/image_classification_ppresnet50_2022jan_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:92c4ba8e363cc5114279ca61f62838600f3121481b74b73c744086b64c694003
+size 28093644
diff --git a/models/license_plate_detection_yunet/README.md b/models/license_plate_detection_yunet/README.md
@@ -4,6 +4,9 @@ This model is contributed by Dong Xu (徐栋) from [watrix.ai](watrix.ai) (银
 
 Please note that the model is trained with Chinese license plates, so the detection results of other license plates with this model may be limited.
 
+**Note**:
+- `license_plate_detection_lpd_yunet_2023mar_int8bq.onnx` represents the block-quantized version in int8 precision and is generated using [block_quantize.py](../../tools/quantize/block_quantize.py) with `block_size=64`.
+
 ## Demo
 
 Run the following command to try the demo:
diff --git a/models/license_plate_detection_yunet/license_plate_detection_lpd_yunet_2023mar_int8bq.onnx b/models/license_plate_detection_yunet/license_plate_detection_lpd_yunet_2023mar_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a346e9db6a085a79848903a95cb902b2ab01d47972057f1cf71ede095410b49
+size 1185866
diff --git a/models/object_detection_nanodet/README.md b/models/object_detection_nanodet/README.md
@@ -2,8 +2,10 @@
 
 Nanodet: NanoDet is a FCOS-style one-stage anchor-free object detection model which using Generalized Focal Loss as classification and regression loss.In NanoDet-Plus, we propose a novel label assignment strategy with a simple assign guidance module (AGM) and a dynamic soft label assigner (DSLA) to solve the optimal label assignment problem in lightweight model training.
 
-Note:
+**Note**:
 - This version of nanodet: Nanodet-m-plus-1.5x_416
+- `object_detection_nanodet_2022nov_int8bq.onnx` represents the block-quantized version in int8 precision and is generated using [block_quantize.py](../../tools/quantize/block_quantize.py) with `block_size=64`.
+
 
 ## Demo
 
diff --git a/models/object_detection_nanodet/object_detection_nanodet_2022nov_int8bq.onnx b/models/object_detection_nanodet/object_detection_nanodet_2022nov_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a2c877cc6f09e7dfac7a9066e33ee5ae68de530b3b994f6ee9125cff6e34d3f
+size 1123958
diff --git a/models/object_detection_yolox/README.md b/models/object_detection_yolox/README.md
@@ -8,8 +8,10 @@ Key features of the YOLOX object detector
 - **SimOTA advanced label assignment strategy** reduces training time and avoids additional solver hyperparameters
 - **Strong data augmentations like MixUp and Mosiac** to boost YOLOX performance
 
-Note:
+**Note**:
 - This version of YoloX: YoloX_s
+- `object_detection_yolox_2022nov_int8bq.onnx` represents the block-quantized version in int8 precision and is generated using [block_quantize.py](../../tools/quantize/block_quantize.py) with `block_size=64`.
+
 
 ## Demo
 
diff --git a/models/object_detection_yolox/object_detection_yolox_2022nov_int8bq.onnx b/models/object_detection_yolox/object_detection_yolox_2022nov_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dcaae0aaa2fea4167f89235ee340eb869d3707b25712218d4c7ce921ac90e2ba
+size 9744418
diff --git a/models/object_tracking_vittrack/README.md b/models/object_tracking_vittrack/README.md
@@ -8,7 +8,10 @@ Video demo: https://youtu.be/MJiPnu1ZQRI
 
 This model is contributed by [Pengyu Liu](https://github.com/lpylpy0514) in GSoC 2023 project [**Realtime object tracking models**](https://github.com/opencv/opencv/wiki/GSoC_2023#idea-realtime-object-tracking-models)
 
-**NOTE: OpenCV > 4.8.0 is required. Build from source with instructions from https://opencv.org/get-started/.**
+**Note**:
+- OpenCV > 4.8.0 is required. Build from source with instructions from https://opencv.org/get-started/.**
+- `object_tracking_vittrack_2023sep_int8bq.onnx` represents the block-quantized version in int8 precision and is generated using [block_quantize.py](../../tools/quantize/block_quantize.py) with `block_size=64`.
+
 
 # Demo
 ## Python
diff --git a/models/object_tracking_vittrack/object_tracking_vittrack_2023sep_int8bq.onnx b/models/object_tracking_vittrack/object_tracking_vittrack_2023sep_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54e8d58892a49de71fadf6673ba10193f7899324a4c1b8fe8c2f2d8d5d661fb4
+size 271327
diff --git a/models/optical_flow_estimation_raft/README.md b/models/optical_flow_estimation_raft/README.md
@@ -1,6 +1,8 @@
 # RAFT
 This model is originally created by Zachary Teed and Jia Deng of Princeton University. The source code for the model is at [their repository on GitHub](https://github.com/princeton-vl/RAFT), and the original [research paper](https://arxiv.org/abs/2003.12039) is published on [Arxiv](https://arxiv.org/abs/2003.12039). The model was converted to ONNX by [PINTO0309](https://github.com/PINTO0309) in his [model zoo](https://github.com/PINTO0309/PINTO_model_zoo/tree/main/252_RAFT). The ONNX model has several variations depending on the training dataset and input dimesnions. The model used in this demo is trained on Sintel dataset with input size of 360 $\times$ 480.
 
+**Note**:
+- `optical_flow_estimation_raft_2023aug_int8bq.onnx` represents the block-quantized version in int8 precision and is generated using [block_quantize.py](../../tools/quantize/block_quantize.py) with `block_size=64`.
 
 ## Demo
 
diff --git a/models/optical_flow_estimation_raft/optical_flow_estimation_raft_2023aug_int8bq.onnx b/models/optical_flow_estimation_raft/optical_flow_estimation_raft_2023aug_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d54cf6e44be6da7efa8843b696b24cf24a8cf03e877dd5207ae5556cfa6f11ec
+size 48845109
diff --git a/models/palm_detection_mediapipe/README.md b/models/palm_detection_mediapipe/README.md
@@ -9,6 +9,7 @@ SSD Anchors are generated from [GenMediaPipePalmDectionSSDAnchors](https://githu
 
 **Note**:
 - Visit https://github.com/google/mediapipe/blob/master/docs/solutions/models.md#hands for models of larger scale.
+- `palm_detection_mediapipe_2023feb_int8bq.onnx` represents the block-quantized version in int8 precision and is generated using [block_quantize.py](../../tools/quantize/block_quantize.py) with `block_size=64`.
 
 ## Demo
 
diff --git a/models/palm_detection_mediapipe/palm_detection_mediapipe_2023feb_int8bq.onnx b/models/palm_detection_mediapipe/palm_detection_mediapipe_2023feb_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0096a81cf82349b00d0c4cb965973662a86967a8d44ccd8134da2a2f408ee5c
+size 1169351
diff --git a/models/person_detection_mediapipe/README.md b/models/person_detection_mediapipe/README.md
@@ -7,6 +7,9 @@ This model detects upper body and full body keypoints of a person, and is downlo
 
 SSD Anchors are generated from [GenMediaPipePalmDectionSSDAnchors](https://github.com/VimalMollyn/GenMediaPipePalmDectionSSDAnchors)
 
+**Note**:
+- `person_detection_mediapipe_2023mar_int8bq.onnx` represents the block-quantized version in int8 precision and is generated using [block_quantize.py](../../tools/quantize/block_quantize.py) with `block_size=64`.
+
 ## Demo
 
 ### Python
diff --git a/models/person_detection_mediapipe/person_detection_mediapipe_2023mar_int8bq.onnx b/models/person_detection_mediapipe/person_detection_mediapipe_2023mar_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5ed8c00c028b98e5d2c55b920a6e975af6c4cd538cfeea7c054f4fbbd8b9075
+size 3482053
diff --git a/models/person_reid_youtureid/README.md b/models/person_reid_youtureid/README.md
@@ -2,9 +2,9 @@
 
 This model is provided by Tencent Youtu Lab [[Credits]](https://github.com/opencv/opencv/blob/394e640909d5d8edf9c1f578f8216d513373698c/samples/dnn/person_reid.py#L6-L11).
 
-Note:
-
+**Note**:
 - Model source: https://github.com/ReID-Team/ReID_extra_testdata
+- `person_reid_youtu_2021nov_int8bq.onnx` represents the block-quantized version in int8 precision and is generated using [block_quantize.py](../../tools/quantize/block_quantize.py) with `block_size=64`.
 
 ## Demo
 
diff --git a/models/person_reid_youtureid/person_reid_youtu_2021nov_int8bq.onnx b/models/person_reid_youtureid/person_reid_youtu_2021nov_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b88597426335e6cd625119bdda090f9d3497bc80ba5b8a8910f65b8ccc09471
+size 29203236
diff --git a/models/pose_estimation_mediapipe/README.md b/models/pose_estimation_mediapipe/README.md
@@ -10,6 +10,8 @@ This model is converted from TFlite to ONNX using following tools:
 
 **Note**:
 - Visit https://github.com/google/mediapipe/blob/master/docs/solutions/models.md#pose for models of larger scale.
+- `pose_estimation_mediapipe_2023mar_int8bq.onnx` represents the block-quantized version in int8 precision and is generated using [block_quantize.py](../../tools/quantize/block_quantize.py) with `block_size=64`.
+
 ## Demo
 
 ### python
diff --git a/models/pose_estimation_mediapipe/pose_estimation_mediapipe_2023mar_int8bq.onnx b/models/pose_estimation_mediapipe/pose_estimation_mediapipe_2023mar_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1848ea80f657f4620e0de160594f598ed8107c9e7785d0be74f65cf15aa6deb1
+size 1694896
diff --git a/models/text_detection_ppocr/README.md b/models/text_detection_ppocr/README.md
@@ -2,14 +2,15 @@
 
 PP-OCRv3: More Attempts for the Improvement of Ultra Lightweight OCR System.
 
-Note:
+**Note**:
 
 - The int8 quantization model may produce unstable results due to some loss of accuracy.
 - Original Paddle Models source of English: [here](https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_det_infer.tar).
 - Original Paddle Models source of Chinese: [here](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar).
 - `IC15` in the filename means the model is trained on [IC15 dataset](https://rrc.cvc.uab.es/?ch=4&com=introduction), which can detect English text instances only.
 - `TD500` in the filename means the model is trained on [TD500 dataset](http://www.iapr-tc11.org/mediawiki/index.php/MSRA_Text_Detection_500_Database_(MSRA-TD500)), which can detect both English & Chinese instances.
 - Visit https://docs.opencv.org/master/d4/d43/tutorial_dnn_text_spotting.html for more information.
+- `text_detection_xx_ppocrv3_2023may_int8bq.onnx` represents the block-quantized version in int8 precision and is generated using [block_quantize.py](../../tools/quantize/block_quantize.py) with `block_size=64`.
 
 ## Demo
 
diff --git a/models/text_detection_ppocr/text_detection_cn_ppocrv3_2023may_int8bq.onnx b/models/text_detection_ppocr/text_detection_cn_ppocrv3_2023may_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f4638708dde26fc77b3cd84aed01019d281268276933ca0e13b0ade5220875f
+size 855375
diff --git a/models/text_detection_ppocr/text_detection_en_ppocrv3_2023may_int8bq.onnx b/models/text_detection_ppocr/text_detection_en_ppocrv3_2023may_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f4638708dde26fc77b3cd84aed01019d281268276933ca0e13b0ade5220875f
+size 855375
diff --git a/models/text_recognition_crnn/README.md b/models/text_recognition_crnn/README.md
@@ -15,7 +15,7 @@ Results of accuracy evaluation with [tools/eval](../../tools/eval) at different
 
 \*: 'FP16' or 'INT8' stands for 'model quantized into FP16' or 'model quantized into int8'
 
-Note:
+**Note**:
 
 - Model source:
   - `text_recognition_CRNN_EN_2021sep.onnx`: https://docs.opencv.org/4.5.2/d9/d1e/tutorial_dnn_OCR.html (CRNN_VGG_BiLSTM_CTC.onnx)
@@ -25,6 +25,7 @@ Note:
 - `text_recognition_CRNN_CH_2021sep.onnx` can detect digits (0\~9), upper/lower-case letters (a\~z and A\~Z), and some special characters (see `CHARSET_CH_94` for details in `crnn.py`).
 - `text_recognition_CRNN_CN_2021nov.onnx` can detect digits (0\~9), upper/lower-case letters (a\~z and A\~Z), some Chinese characters and some special characters (see `CHARSET_CN_3944` for details in `crnn.py`).
 - For details on training this model series, please visit https://github.com/zihaomu/deep-text-recognition-benchmark.
+- `text_recognition_CRNN_XX_2021xxx_int8bq.onnx` represents the block-quantized version in int8 precision and is generated using [block_quantize.py](../../tools/quantize/block_quantize.py) with `block_size=64`.
 
 ## Demo
 
diff --git a/models/text_recognition_crnn/text_recognition_CRNN_CH_2021sep_int8bq.onnx b/models/text_recognition_crnn/text_recognition_CRNN_CH_2021sep_int8bq.onnx
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c205e1a1a3bc5a1d585ea380b55d2801ba655c473bd5d41730fc14d6341f2e16
+size 26887550
diff --git a/models/text_recognition_crnn/text_recognition_CRNN_CN_2021nov_int8bq.onnx b/models/text_recognition_crnn/text_recognition_CRNN_CN_2021nov_int8bq.onnx
diff --git a/tools/eval/README.md b/tools/eval/README.md
diff --git a/tools/eval/eval.py b/tools/eval/eval.py
diff --git a/tools/quantize/README.md b/tools/quantize/README.md
diff --git a/tools/quantize/block_quantize.py b/tools/quantize/block_quantize.py

Original file line number	Diff line number	Diff line change
`@@ -0,0 +1,3 @@`
	`1`	`+version https://git-lfs.github.com/spec/v1`
	`2`	`+oid sha256:49f000ec501fef24739071fc7e68267d32209045b6822c0c72dce1da25726f10`
	`3`	`+size 122489`
Original file line number	Diff line number	Diff line change
`@@ -0,0 +1,3 @@`
	`1`	`+version https://git-lfs.github.com/spec/v1`
	`2`	`+oid sha256:fb143eea07838aa532d1c95df5f69899974ea0140e1fba05e94204be13ed74ee`
	`3`	`+size 10667852`
Original file line number	Diff line number	Diff line change
`@@ -0,0 +1,3 @@`
	`1`	`+version https://git-lfs.github.com/spec/v1`
	`2`	`+oid sha256:0c3eaf9d0a7d442c0aa3beb3234243e1cdff9ad8871fb3cec346e90874caf57d`
	`3`	`+size 1376702`
Original file line number	Diff line number	Diff line change
`@@ -0,0 +1,3 @@`
	`1`	`+version https://git-lfs.github.com/spec/v1`
	`2`	`+oid sha256:d7e63ed33e39c8b532a04a5466ac68c8680981637df288bcf93286bec08befbd`
	`3`	`+size 1222348`
Original file line number	Diff line number	Diff line change
`@@ -0,0 +1,3 @@`
	`1`	`+version https://git-lfs.github.com/spec/v1`
	`2`	`+oid sha256:1da023e95e3adbeef16ac2862767b53b86a743ff26a34692e0833d7e088f0231`
	`3`	`+size 1734724`
Original file line number	Diff line number	Diff line change
`@@ -0,0 +1,3 @@`
	`1`	`+version https://git-lfs.github.com/spec/v1`
	`2`	`+oid sha256:4437385b4908011f6e3019ae253031272c19522ae1cbdff8374bceffe262a5ee`
	`3`	`+size 4599388`
Original file line number	Diff line number	Diff line change
`@@ -0,0 +1,3 @@`
	`1`	`+version https://git-lfs.github.com/spec/v1`
	`2`	`+oid sha256:3fe457eda49d71b664918ae87349bdb3e3815a56bb82cdb7f477bc83597f4313`
	`3`	`+size 3872948`
Original file line number	Diff line number	Diff line change
`@@ -0,0 +1,3 @@`
	`1`	`+version https://git-lfs.github.com/spec/v1`
	`2`	`+oid sha256:92c4ba8e363cc5114279ca61f62838600f3121481b74b73c744086b64c694003`
	`3`	`+size 28093644`
Original file line number	Diff line number	Diff line change
`@@ -0,0 +1,3 @@`
	`1`	`+version https://git-lfs.github.com/spec/v1`
	`2`	`+oid sha256:8a346e9db6a085a79848903a95cb902b2ab01d47972057f1cf71ede095410b49`
	`3`	`+size 1185866`
Original file line number	Diff line number	Diff line change
`@@ -0,0 +1,3 @@`
	`1`	`+version https://git-lfs.github.com/spec/v1`
	`2`	`+oid sha256:8a2c877cc6f09e7dfac7a9066e33ee5ae68de530b3b994f6ee9125cff6e34d3f`
	`3`	`+size 1123958`