[Other] Improve examples and readme for Ascend deployment (#1052)

* Add Huawei Ascend NPU deploy through PaddleLite CANN

* Add NNAdapter interface for paddlelite

* Modify Huawei Ascend Cmake

* Update way for compiling Huawei Ascend NPU deployment

* remove UseLiteBackend in UseCANN

* Support compile python whlee

* Change names of nnadapter API

* Add nnadapter pybind and remove useless API

* Support Python deployment on Huawei Ascend NPU

* Add models suppor for ascend

* Add PPOCR rec reszie for ascend

* fix conflict for ascend

* Rename CANN to Ascend

* Rename CANN to Ascend

* Improve ascend

* fix ascend bug

* improve ascend docs

* improve ascend docs

* improve ascend docs

* Improve Ascend

* Improve Ascend

* Move ascend python demo

* Imporve ascend

* Improve ascend

* Improve ascend

* Improve ascend

* Improve ascend

* Imporve ascend

* Imporve ascend

* Improve ascend

* acc eval script

* acc eval

* remove acc_eval from branch huawei

* Add detection and segmentation examples for Ascend deployment

* Add detection and segmentation examples for Ascend deployment

* Add PPOCR example for ascend deploy

* Imporve paddle lite compiliation

* Add FlyCV doc

* Add FlyCV doc

* Add FlyCV doc

* Imporve Ascend docs

* Imporve Ascend docs

* Improve PPOCR example

examples/vision/detection/paddledetection/cpp/README.md

@@ -41,6 +41,9 @@ tar xvf ppyoloe_crn_l_300e_coco.tgz
 以上命令只适用于Linux或MacOS, Windows下SDK的使用方式请参考:  
 - [如何在Windows中使用FastDeploy C++ SDK](../../../../../docs/cn/faq/use_sdk_on_windows.md)
 
+如果用户使用华为昇腾NPU部署, 请参考以下方式在部署前初始化部署环境:
+- [如何使用华为昇腾NPU部署](../../../../../docs/cn/faq/use_sdk_on_ascend.md)
+
 ## PaddleDetection C++接口
 
 ### 模型类

examples/vision/detection/yolov5/cpp/README.md

@@ -55,6 +55,9 @@ wget https://gitee.com/paddlepaddle/PaddleDetection/raw/release/2.4/demo/0000000
 以上命令只适用于Linux或MacOS, Windows下SDK的使用方式请参考:  
 - [如何在Windows中使用FastDeploy C++ SDK](../../../../../docs/cn/faq/use_sdk_on_windows.md)
 
+如果用户使用华为昇腾NPU部署, 请参考以下方式在部署前初始化部署环境:
+- [如何使用华为昇腾NPU部署](../../../../../docs/cn/faq/use_sdk_on_ascend.md)
+
 ## YOLOv5 C++接口
 
 ### YOLOv5类

examples/vision/detection/yolov6/cpp/README.md

@@ -33,6 +33,10 @@ wget https://gitee.com/paddlepaddle/PaddleDetection/raw/release/2.4/demo/0000000
 ./infer_paddle_demo yolov6s_infer 000000014439.jpg 3
 ```
 
+如果用户使用华为昇腾NPU部署, 请参考以下方式在部署前初始化部署环境:
+- [如何使用华为昇腾NPU部署](../../../../../docs/cn/faq/use_sdk_on_ascend.md)
+
+
 如果想要验证ONNX模型的推理，可以参考如下命令：
 ```bash
 #下载官方转换好的YOLOv6 ONNX模型文件和测试图片

examples/vision/detection/yolov7/cpp/README.md

@@ -31,6 +31,10 @@ wget https://gitee.com/paddlepaddle/PaddleDetection/raw/release/2.4/demo/0000000
 # 华为昇腾推理
 ./infer_paddle_model_demo yolov7_infer 000000014439.jpg 3
 ```
+
+如果用户使用华为昇腾NPU部署, 请参考以下方式在部署前初始化部署环境:
+- [如何使用华为昇腾NPU部署](../../../../../docs/cn/faq/use_sdk_on_ascend.md)
+
 如果想要验证ONNX模型的推理，可以参考如下命令：
 ```bash
 #下载官方转换好的yolov7 ONNX模型文件和测试图片

examples/vision/ocr/PP-OCRv2/cpp/CMakeLists.txt

@@ -12,3 +12,7 @@ include_directories(${FASTDEPLOY_INCS})
 add_executable(infer_demo ${PROJECT_SOURCE_DIR}/infer.cc)
 # 添加FastDeploy库依赖
 target_link_libraries(infer_demo ${FASTDEPLOY_LIBS})
+
+add_executable(infer_static_shape_demo ${PROJECT_SOURCE_DIR}/infer_static_shape.cc)
+# 添加FastDeploy库依赖
+target_link_libraries(infer_static_shape_demo ${FASTDEPLOY_LIBS})

examples/vision/ocr/PP-OCRv2/cpp/README.md

@@ -43,13 +43,16 @@ wget https://gitee.com/paddlepaddle/PaddleOCR/raw/release/2.6/ppocr/utils/ppocr_
 ./infer_demo ./ch_PP-OCRv2_det_infer ./ch_ppocr_mobile_v2.0_cls_infer ./ch_PP-OCRv2_rec_infer ./ppocr_keys_v1.txt ./12.jpg 3
 # 昆仑芯XPU推理
 ./infer_demo ./ch_PP-OCRv2_det_infer ./ch_ppocr_mobile_v2.0_cls_infer ./ch_PP-OCRv2_rec_infer ./ppocr_keys_v1.txt ./12.jpg 4
-# 华为昇腾推理
+# 华为昇腾推理, 需要使用静态shape的demo, 若用户需要连续地预测图片, 输入图片尺寸需要准备为统一尺寸
-./infer_demo ./ch_PP-OCRv2_det_infer ./ch_ppocr_mobile_v2.0_cls_infer ./ch_PP-OCRv2_rec_infer ./ppocr_keys_v1.txt ./12.jpg 5
+./infer_static_shape_demo ./ch_PP-OCRv2_det_infer ./ch_ppocr_mobile_v2.0_cls_infer ./ch_PP-OCRv2_rec_infer ./ppocr_keys_v1.txt ./12.jpg 1
 ```
 
 以上命令只适用于Linux或MacOS, Windows下SDK的使用方式请参考:  
 - [如何在Windows中使用FastDeploy C++ SDK](../../../../../docs/cn/faq/use_sdk_on_windows.md)
 
+如果用户使用华为昇腾NPU部署, 请参考以下方式在部署前初始化部署环境:
+- [如何使用华为昇腾NPU部署](../../../../../docs/cn/faq/use_sdk_on_ascend.md)
+
 运行完成可视化结果如下图所示
 
 <img width="640" src="https://user-images.githubusercontent.com/109218879/185826024-f7593a0c-1bd2-4a60-b76c-15588484fa08.jpg">

examples/vision/ocr/PP-OCRv2/cpp/infer.cc

@@ -55,10 +55,6 @@ void InitAndInfer(const std::string& det_model_dir, const std::string& cls_model
   auto cls_model = fastdeploy::vision::ocr::Classifier(cls_model_file, cls_params_file, cls_option);
   auto rec_model = fastdeploy::vision::ocr::Recognizer(rec_model_file, rec_params_file, rec_label_file, rec_option);
 
-  // Users could enable static shape infer for rec model when deploy PP-OCR on hardware 
-  // which can not support dynamic shape infer well, like Huawei Ascend series. 
-  // rec_model.GetPreprocessor().SetStaticShapeInfer(true);
-
   assert(det_model.Initialized());
   assert(cls_model.Initialized());
   assert(rec_model.Initialized());
@@ -70,9 +66,6 @@ void InitAndInfer(const std::string& det_model_dir, const std::string& cls_model
   // Set inference batch size for cls model and rec model, the value could be -1 and 1 to positive infinity.
   // When inference batch size is set to -1, it means that the inference batch size 
   // of the cls and rec models will be the same as the number of boxes detected by the det model. 
-  // When users enable static shape infer for rec model, the batch size of cls and rec model needs to be set to 1.
-  // ppocr_v2.SetClsBatchSize(1);
-  // ppocr_v2.SetRecBatchSize(1); 
   ppocr_v2.SetClsBatchSize(cls_batch_size);
   ppocr_v2.SetRecBatchSize(rec_batch_size); 
 
@@ -129,8 +122,6 @@ int main(int argc, char* argv[]) {
     option.EnablePaddleToTrt();
   } else if (flag == 4) {
     option.UseKunlunXin();
-  } else if (flag == 5) {
-    option.UseAscend();
   }
 
   std::string det_model_dir = argv[1];

examples/vision/ocr/PP-OCRv2/cpp/infer_static_shape.cc

@@ -0,0 +1,107 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "fastdeploy/vision.h"
+#ifdef WIN32
+const char sep = '\\';
+#else
+const char sep = '/';
+#endif
+
+void InitAndInfer(const std::string& det_model_dir, const std::string& cls_model_dir, const std::string& rec_model_dir, const std::string& rec_label_file, const std::string& image_file, const fastdeploy::RuntimeOption& option) {
+  auto det_model_file = det_model_dir + sep + "inference.pdmodel";
+  auto det_params_file = det_model_dir + sep + "inference.pdiparams";
+
+  auto cls_model_file = cls_model_dir + sep + "inference.pdmodel";
+  auto cls_params_file = cls_model_dir + sep + "inference.pdiparams";
+
+  auto rec_model_file = rec_model_dir + sep + "inference.pdmodel";
+  auto rec_params_file = rec_model_dir + sep + "inference.pdiparams";
+
+  auto det_option = option;
+  auto cls_option = option;
+  auto rec_option = option;
+
+  auto det_model = fastdeploy::vision::ocr::DBDetector(det_model_file, det_params_file, det_option);
+  auto cls_model = fastdeploy::vision::ocr::Classifier(cls_model_file, cls_params_file, cls_option);
+  auto rec_model = fastdeploy::vision::ocr::Recognizer(rec_model_file, rec_params_file, rec_label_file, rec_option);
+
+  // Users could enable static shape infer for rec model when deploy PP-OCR on hardware 
+  // which can not support dynamic shape infer well, like Huawei Ascend series. 
+  rec_model.GetPreprocessor().SetStaticShapeInfer(true);
+
+  assert(det_model.Initialized());
+  assert(cls_model.Initialized());
+  assert(rec_model.Initialized());
+
+  // The classification model is optional, so the PP-OCR can also be connected in series as follows
+  // auto ppocr_v2 = fastdeploy::pipeline::PPOCRv2(&det_model, &rec_model);
+  auto ppocr_v2 = fastdeploy::pipeline::PPOCRv2(&det_model, &cls_model, &rec_model);
+
+  // When users enable static shape infer for rec model, the batch size of cls and rec model must to be set to 1.
+  ppocr_v2.SetClsBatchSize(1);
+  ppocr_v2.SetRecBatchSize(1); 
+
+  if(!ppocr_v2.Initialized()){
+    std::cerr << "Failed to initialize PP-OCR." << std::endl;
+    return;
+  }
+
+  auto im = cv::imread(image_file);
+  
+  fastdeploy::vision::OCRResult result;
+  if (!ppocr_v2.Predict(im, &result)) {
+    std::cerr << "Failed to predict." << std::endl;
+    return;
+  }
+
+  std::cout << result.Str() << std::endl;
+
+  auto vis_im = fastdeploy::vision::VisOcr(im, result);
+  cv::imwrite("vis_result.jpg", vis_im);
+  std::cout << "Visualized result saved in ./vis_result.jpg" << std::endl;
+}
+
+int main(int argc, char* argv[]) {
+  if (argc < 7) {
+    std::cout << "Usage: infer_demo path/to/det_model path/to/cls_model "
+                 "path/to/rec_model path/to/rec_label_file path/to/image "
+                 "run_option, "
+                 "e.g ./infer_demo ./ch_PP-OCRv2_det_infer "
+                 "./ch_ppocr_mobile_v2.0_cls_infer ./ch_PP-OCRv2_rec_infer "
+                 "./ppocr_keys_v1.txt ./12.jpg 0"
+              << std::endl;
+    std::cout << "The data type of run_option is int, 0: run with cpu; 1: run "
+                 "with ascend."
+              << std::endl;
+    return -1;
+  }
+
+  fastdeploy::RuntimeOption option;
+  int flag = std::atoi(argv[6]);
+
+  if (flag == 0) {
+    option.UseCpu(); 
+  } else if (flag == 1) {
+    option.UseAscend();
+  }
+
+  std::string det_model_dir = argv[1];
+  std::string cls_model_dir = argv[2];
+  std::string rec_model_dir = argv[3];
+  std::string rec_label_file = argv[4];
+  std::string test_image = argv[5];
+  InitAndInfer(det_model_dir, cls_model_dir, rec_model_dir, rec_label_file, test_image, option);
+  return 0;
+}

examples/vision/ocr/PP-OCRv2/python/README.md

@@ -36,8 +36,8 @@ python infer.py --det_model ch_PP-OCRv2_det_infer --cls_model ch_ppocr_mobile_v2
 python infer.py --det_model ch_PP-OCRv2_det_infer --cls_model ch_ppocr_mobile_v2.0_cls_infer --rec_model ch_PP-OCRv2_rec_infer --rec_label_file ppocr_keys_v1.txt --image 12.jpg --device gpu --backend trt
 # 昆仑芯XPU推理
 python infer.py --det_model ch_PP-OCRv2_det_infer --cls_model ch_ppocr_mobile_v2.0_cls_infer --rec_model ch_PP-OCRv2_rec_infer --rec_label_file ppocr_keys_v1.txt --image 12.jpg --device kunlunxin
-# 华为昇腾推理
+# 华为昇腾推理,需要使用静态shape脚本, 若用户需要连续地预测图片, 输入图片尺寸需要准备为统一尺寸
-python infer.py --det_model ch_PP-OCRv2_det_infer --cls_model ch_ppocr_mobile_v2.0_cls_infer --rec_model ch_PP-OCRv2_rec_infer --rec_label_file ppocr_keys_v1.txt --image 12.jpg --device ascend
+python infer_static_shape.py --det_model ch_PP-OCRv2_det_infer --cls_model ch_ppocr_mobile_v2.0_cls_infer --rec_model ch_PP-OCRv2_rec_infer --rec_label_file ppocr_keys_v1.txt --image 12.jpg --device ascend
 ```
 
 运行完成可视化结果如下图所示

examples/vision/ocr/PP-OCRv2/python/infer.py

@@ -58,43 +58,113 @@ def parse_arguments():
         type=int,
         default=9,
         help="Number of threads while inference on CPU.")
+    parser.add_argument(
+        "--cls_bs",
+        type=int,
+        default=1,
+        help="Classification model inference batch size.")
+    parser.add_argument(
+        "--rec_bs",
+        type=int,
+        default=6,
+        help="Recognition model inference batch size")
     return parser.parse_args()
 
 
 def build_option(args):
-    option = fd.RuntimeOption()
-    if args.device.lower() == "gpu":
-        option.use_gpu(0)
 
-    option.set_cpu_thread_num(args.cpu_thread_num)
+    det_option = fd.RuntimeOption()
+    cls_option = fd.RuntimeOption()
+    rec_option = fd.RuntimeOption()
+
+    det_option.set_cpu_thread_num(args.cpu_thread_num)
+    cls_option.set_cpu_thread_num(args.cpu_thread_num)
+    rec_option.set_cpu_thread_num(args.cpu_thread_num)
+
+    if args.device.lower() == "gpu":
+        det_option.use_gpu(args.device_id)
+        cls_option.use_gpu(args.device_id)
+        rec_option.use_gpu(args.device_id)
 
     if args.device.lower() == "kunlunxin":
-        option.use_kunlunxin()
+        det_option.use_kunlunxin()
-        return option
+        cls_option.use_kunlunxin()
+        rec_option.use_kunlunxin()
 
-    if args.device.lower() == "ascend":
+        return det_option, cls_option, rec_option
-        option.use_ascend()
-        return option
 
     if args.backend.lower() == "trt":
         assert args.device.lower(
         ) == "gpu", "TensorRT backend require inference on device GPU."
-        option.use_trt_backend()
+        det_option.use_trt_backend()
+        cls_option.use_trt_backend()
+        rec_option.use_trt_backend()
+
+        # 设置trt input shape
+        # 如果用户想要自己改动检测模型的输入shape, 我们建议用户把检测模型的长和高设置为32的倍数.
+        det_option.set_trt_input_shape("x", [1, 3, 64, 64], [1, 3, 640, 640],
+                                       [1, 3, 960, 960])
+        cls_option.set_trt_input_shape("x", [1, 3, 48, 10],
+                                       [args.cls_bs, 3, 48, 320],
+                                       [args.cls_bs, 3, 48, 1024])
+        rec_option.set_trt_input_shape("x", [1, 3, 32, 10],
+                                       [args.rec_bs, 3, 32, 320],
+                                       [args.rec_bs, 3, 32, 2304])
+
+        # 用户可以把TRT引擎文件保存至本地
+        det_option.set_trt_cache_file(args.det_model + "/det_trt_cache.trt")
+        cls_option.set_trt_cache_file(args.cls_model + "/cls_trt_cache.trt")
+        rec_option.set_trt_cache_file(args.rec_model + "/rec_trt_cache.trt")
+
     elif args.backend.lower() == "pptrt":
         assert args.device.lower(
         ) == "gpu", "Paddle-TensorRT backend require inference on device GPU."
-        option.use_trt_backend()
+        det_option.use_trt_backend()
-        option.enable_paddle_trt_collect_shape()
+        det_option.enable_paddle_trt_collect_shape()
-        option.enable_paddle_to_trt()
+        det_option.enable_paddle_to_trt()
+
+        cls_option.use_trt_backend()
+        cls_option.enable_paddle_trt_collect_shape()
+        cls_option.enable_paddle_to_trt()
+
+        rec_option.use_trt_backend()
+        rec_option.enable_paddle_trt_collect_shape()
+        rec_option.enable_paddle_to_trt()
+
+        # 设置trt input shape
+        # 如果用户想要自己改动检测模型的输入shape, 我们建议用户把检测模型的长和高设置为32的倍数.
+        det_option.set_trt_input_shape("x", [1, 3, 64, 64], [1, 3, 640, 640],
+                                       [1, 3, 960, 960])
+        cls_option.set_trt_input_shape("x", [1, 3, 48, 10],
+                                       [args.cls_bs, 3, 48, 320],
+                                       [args.cls_bs, 3, 48, 1024])
+        rec_option.set_trt_input_shape("x", [1, 3, 32, 10],
+                                       [args.rec_bs, 3, 32, 320],
+                                       [args.rec_bs, 3, 32, 2304])
+
+        # 用户可以把TRT引擎文件保存至本地
+        det_option.set_trt_cache_file(args.det_model)
+        cls_option.set_trt_cache_file(args.cls_model)
+        rec_option.set_trt_cache_file(args.rec_model)
+
     elif args.backend.lower() == "ort":
-        option.use_ort_backend()
+        det_option.use_ort_backend()
+        cls_option.use_ort_backend()
+        rec_option.use_ort_backend()
+
     elif args.backend.lower() == "paddle":
-        option.use_paddle_infer_backend()
+        det_option.use_paddle_infer_backend()
+        cls_option.use_paddle_infer_backend()
+        rec_option.use_paddle_infer_backend()
+
     elif args.backend.lower() == "openvino":
         assert args.device.lower(
         ) == "cpu", "OpenVINO backend require inference on device CPU."
-        option.use_openvino_backend()
+        det_option.use_openvino_backend()
-    return option
+        cls_option.use_openvino_backend()
+        rec_option.use_openvino_backend()
+
+    return det_option, cls_option, rec_option
 
 
 args = parse_arguments()
@@ -111,49 +181,18 @@ rec_params_file = os.path.join(args.rec_model, "inference.pdiparams")
 rec_label_file = args.rec_label_file
 
 # 对于三个模型，均采用同样的部署配置
-# 用户也可根据自行需求分别配置
+# 用户也可根据自己的需求,个性化配置
-runtime_option = build_option(args)
+det_option, cls_option, rec_option = build_option(args)
 
-# PPOCR的cls和rec模型现在已经支持推理一个Batch的数据
-# 定义下面两个变量后, 可用于设置trt输入shape, 并在PPOCR模型初始化后, 完成Batch推理设置
-# 当用户要把PP-OCR部署在对动态shape推理支持有限的设备上时,(例如华为昇腾)
-# 需要把cls_batch_size和rec_batch_size都设置为1.
-cls_batch_size = 1
-rec_batch_size = 6
-
-# 当使用TRT时，分别给三个模型的runtime设置动态shape,并完成模型的创建.
-# 注意: 需要在检测模型创建完成后，再设置分类模型的动态输入并创建分类模型, 识别模型同理.
-# 如果用户想要自己改动检测模型的输入shape, 我们建议用户把检测模型的长和高设置为32的倍数.
-det_option = runtime_option
-det_option.set_trt_input_shape("x", [1, 3, 64, 64], [1, 3, 640, 640],
-                               [1, 3, 960, 960])
-# 用户可以把TRT引擎文件保存至本地
-# det_option.set_trt_cache_file(args.det_model  + "/det_trt_cache.trt")
 det_model = fd.vision.ocr.DBDetector(
     det_model_file, det_params_file, runtime_option=det_option)
 
-cls_option = runtime_option
-cls_option.set_trt_input_shape("x", [1, 3, 48, 10],
-                               [cls_batch_size, 3, 48, 320],
-                               [cls_batch_size, 3, 48, 1024])
-# 用户可以把TRT引擎文件保存至本地
-# cls_option.set_trt_cache_file(args.cls_model  + "/cls_trt_cache.trt")
 cls_model = fd.vision.ocr.Classifier(
     cls_model_file, cls_params_file, runtime_option=cls_option)
 
-rec_option = runtime_option
-rec_option.set_trt_input_shape("x", [1, 3, 32, 10],
-                               [rec_batch_size, 3, 32, 320],
-                               [rec_batch_size, 3, 32, 2304])
-# 用户可以把TRT引擎文件保存至本地
-# rec_option.set_trt_cache_file(args.rec_model  + "/rec_trt_cache.trt")
 rec_model = fd.vision.ocr.Recognizer(
     rec_model_file, rec_params_file, rec_label_file, runtime_option=rec_option)
 
-# 当用户要把PP-OCR部署在对动态shape推理支持有限的设备上时,(例如华为昇腾)
-# 需要使用下行代码, 来启用rec模型的静态shape推理.
-# rec_model.preprocessor.static_shape_infer = True
-
 # 创建PP-OCR，串联3个模型，其中cls_model可选，如无需求，可设置为None
 ppocr_v2 = fd.vision.ocr.PPOCRv2(
     det_model=det_model, cls_model=cls_model, rec_model=rec_model)
@@ -161,8 +200,8 @@ ppocr_v2 = fd.vision.ocr.PPOCRv2(
 # 给cls和rec模型设置推理时的batch size
 # 此值能为-1, 和1到正无穷
 # 当此值为-1时, cls和rec模型的batch size将默认和det模型检测出的框的数量相同
-ppocr_v2.cls_batch_size = cls_batch_size
+ppocr_v2.cls_batch_size = args.cls_bs
-ppocr_v2.rec_batch_size = rec_batch_size
+ppocr_v2.rec_batch_size = args.rec_bs
 
 # 预测图片准备
 im = cv2.imread(args.image)

examples/vision/ocr/PP-OCRv2/python/infer_static_shape.py

@@ -0,0 +1,114 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import fastdeploy as fd
+import cv2
+import os
+
+
+def parse_arguments():
+    import argparse
+    import ast
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--det_model", required=True, help="Path of Detection model of PPOCR.")
+    parser.add_argument(
+        "--cls_model",
+        required=True,
+        help="Path of Classification model of PPOCR.")
+    parser.add_argument(
+        "--rec_model",
+        required=True,
+        help="Path of Recognization model of PPOCR.")
+    parser.add_argument(
+        "--rec_label_file",
+        required=True,
+        help="Path of Recognization model of PPOCR.")
+    parser.add_argument(
+        "--image", type=str, required=True, help="Path of test image file.")
+    parser.add_argument(
+        "--device",
+        type=str,
+        default='cpu',
+        help="Type of inference device, support 'cpu', 'kunlunxin' or 'gpu'.")
+    parser.add_argument(
+        "--cpu_thread_num",
+        type=int,
+        default=9,
+        help="Number of threads while inference on CPU.")
+    return parser.parse_args()
+
+
+def build_option(args):
+
+    det_option = fd.RuntimeOption()
+    cls_option = fd.RuntimeOption()
+    rec_option = fd.RuntimeOption()
+
+    # 当前需要对PP-OCR启用静态shape推理的硬件只有昇腾.
+    if args.device.lower() == "ascend":
+        det_option.use_ascend()
+        cls_option.use_ascend()
+        rec_option.use_ascend()
+
+    return det_option, cls_option, rec_option
+
+
+args = parse_arguments()
+
+# Detection模型, 检测文字框
+det_model_file = os.path.join(args.det_model, "inference.pdmodel")
+det_params_file = os.path.join(args.det_model, "inference.pdiparams")
+# Classification模型，方向分类，可选
+cls_model_file = os.path.join(args.cls_model, "inference.pdmodel")
+cls_params_file = os.path.join(args.cls_model, "inference.pdiparams")
+# Recognition模型，文字识别模型
+rec_model_file = os.path.join(args.rec_model, "inference.pdmodel")
+rec_params_file = os.path.join(args.rec_model, "inference.pdiparams")
+rec_label_file = args.rec_label_file
+
+det_option, cls_option, rec_option = build_option(args)
+
+det_model = fd.vision.ocr.DBDetector(
+    det_model_file, det_params_file, runtime_option=det_option)
+
+cls_model = fd.vision.ocr.Classifier(
+    cls_model_file, cls_params_file, runtime_option=cls_option)
+
+rec_model = fd.vision.ocr.Recognizer(
+    rec_model_file, rec_params_file, rec_label_file, runtime_option=rec_option)
+
+# Rec模型启用静态shape推理
+rec_model.preprocessor.static_shape_infer = True
+
+# 创建PP-OCR，串联3个模型，其中cls_model可选，如无需求，可设置为None
+ppocr_v2 = fd.vision.ocr.PPOCRv2(
+    det_model=det_model, cls_model=cls_model, rec_model=rec_model)
+
+# Cls模型和Rec模型的batch size 必须设置为1, 开启静态shape推理
+ppocr_v2.cls_batch_size = 1
+ppocr_v2.rec_batch_size = 1
+
+# 预测图片准备
+im = cv2.imread(args.image)
+
+#预测并打印结果
+result = ppocr_v2.predict(im)
+
+print(result)
+
+# 可视化结果
+vis_im = fd.vision.vis_ppocr(im, result)
+cv2.imwrite("visualized_result.jpg", vis_im)
+print("Visualized result save in ./visualized_result.jpg")

examples/vision/ocr/PP-OCRv3/cpp/CMakeLists.txt

@@ -12,3 +12,7 @@ include_directories(${FASTDEPLOY_INCS})
 add_executable(infer_demo ${PROJECT_SOURCE_DIR}/infer.cc)
 # 添加FastDeploy库依赖
 target_link_libraries(infer_demo ${FASTDEPLOY_LIBS})
+
+add_executable(infer_static_shape_demo ${PROJECT_SOURCE_DIR}/infer_static_shape.cc)
+# 添加FastDeploy库依赖
+target_link_libraries(infer_static_shape_demo ${FASTDEPLOY_LIBS})

examples/vision/ocr/PP-OCRv3/cpp/README.md

@@ -43,13 +43,16 @@ wget https://gitee.com/paddlepaddle/PaddleOCR/raw/release/2.6/ppocr/utils/ppocr_
 ./infer_demo ./ch_PP-OCRv3_det_infer ./ch_ppocr_mobile_v2.0_cls_infer ./ch_PP-OCRv3_rec_infer ./ppocr_keys_v1.txt ./12.jpg 3
 # 昆仑芯XPU推理
 ./infer_demo ./ch_PP-OCRv3_det_infer ./ch_ppocr_mobile_v2.0_cls_infer ./ch_PP-OCRv3_rec_infer ./ppocr_keys_v1.txt ./12.jpg 4
-# 华为昇腾推理, 请用户在代码里正确开启Rec模型的静态shape推理，并设置分类模型和识别模型的推理batch size为1.
+# 华为昇腾推理,需要使用静态shape的demo, 若用户需要连续地预测图片, 输入图片尺寸需要准备为统一尺寸
-./infer_demo ./ch_PP-OCRv3_det_infer ./ch_ppocr_mobile_v2.0_cls_infer ./ch_PP-OCRv3_rec_infer ./ppocr_keys_v1.txt ./12.jpg 5
+./infer_static_shape_demo ./ch_PP-OCRv3_det_infer ./ch_ppocr_mobile_v2.0_cls_infer ./ch_PP-OCRv3_rec_infer ./ppocr_keys_v1.txt ./12.jpg 1
 ```
 
 以上命令只适用于Linux或MacOS, Windows下SDK的使用方式请参考:  
 - [如何在Windows中使用FastDeploy C++ SDK](../../../../../docs/cn/faq/use_sdk_on_windows.md)
 
+如果用户使用华为昇腾NPU部署, 请参考以下方式在部署前初始化部署环境:
+- [如何使用华为昇腾NPU部署](../../../../../docs/cn/faq/use_sdk_on_ascend.md)
+
 运行完成可视化结果如下图所示
 
 <img width="640" src="https://user-images.githubusercontent.com/109218879/185826024-f7593a0c-1bd2-4a60-b76c-15588484fa08.jpg">

examples/vision/ocr/PP-OCRv3/cpp/infer.cc

@@ -56,10 +56,6 @@ void InitAndInfer(const std::string& det_model_dir, const std::string& cls_model
   auto cls_model = fastdeploy::vision::ocr::Classifier(cls_model_file, cls_params_file, cls_option);
   auto rec_model = fastdeploy::vision::ocr::Recognizer(rec_model_file, rec_params_file, rec_label_file, rec_option);
 
-  // Users could enable static shape infer for rec model when deploy PP-OCR on hardware 
-  // which can not support dynamic shape infer well, like Huawei Ascend series. 
-  // rec_model.GetPreprocessor().SetStaticShapeInfer(true);
-
   assert(det_model.Initialized());
   assert(cls_model.Initialized());
   assert(rec_model.Initialized());
@@ -71,9 +67,6 @@ void InitAndInfer(const std::string& det_model_dir, const std::string& cls_model
   // Set inference batch size for cls model and rec model, the value could be -1 and 1 to positive infinity.
   // When inference batch size is set to -1, it means that the inference batch size 
   // of the cls and rec models will be the same as the number of boxes detected by the det model. 
-  // When users enable static shape infer for rec model, the batch size of cls and rec model needs to be set to 1.
-  // ppocr_v3.SetClsBatchSize(1);
-  // ppocr_v3.SetRecBatchSize(1); 
   ppocr_v3.SetClsBatchSize(cls_batch_size);
   ppocr_v3.SetRecBatchSize(rec_batch_size); 
 
@@ -130,8 +123,6 @@ int main(int argc, char* argv[]) {
     option.EnablePaddleToTrt();
   } else if (flag == 4) {
     option.UseKunlunXin();
-  } else if (flag == 5) {
-    option.UseAscend();
   }
 
   std::string det_model_dir = argv[1];

examples/vision/ocr/PP-OCRv3/cpp/infer_static_shape.cc

@@ -0,0 +1,107 @@
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "fastdeploy/vision.h"
+#ifdef WIN32
+const char sep = '\\';
+#else
+const char sep = '/';
+#endif
+
+void InitAndInfer(const std::string& det_model_dir, const std::string& cls_model_dir, const std::string& rec_model_dir, const std::string& rec_label_file, const std::string& image_file, const fastdeploy::RuntimeOption& option) {
+  auto det_model_file = det_model_dir + sep + "inference.pdmodel";
+  auto det_params_file = det_model_dir + sep + "inference.pdiparams";
+
+  auto cls_model_file = cls_model_dir + sep + "inference.pdmodel";
+  auto cls_params_file = cls_model_dir + sep + "inference.pdiparams";
+
+  auto rec_model_file = rec_model_dir + sep + "inference.pdmodel";
+  auto rec_params_file = rec_model_dir + sep + "inference.pdiparams";
+
+  auto det_option = option;
+  auto cls_option = option;
+  auto rec_option = option;
+
+  auto det_model = fastdeploy::vision::ocr::DBDetector(det_model_file, det_params_file, det_option);
+  auto cls_model = fastdeploy::vision::ocr::Classifier(cls_model_file, cls_params_file, cls_option);
+  auto rec_model = fastdeploy::vision::ocr::Recognizer(rec_model_file, rec_params_file, rec_label_file, rec_option);
+
+  // Users could enable static shape infer for rec model when deploy PP-OCR on hardware 
+  // which can not support dynamic shape infer well, like Huawei Ascend series. 
+  rec_model.GetPreprocessor().SetStaticShapeInfer(true);
+
+  assert(det_model.Initialized());
+  assert(cls_model.Initialized());
+  assert(rec_model.Initialized());
+
+  // The classification model is optional, so the PP-OCR can also be connected in series as follows
+  // auto ppocr_v3 = fastdeploy::pipeline::PPOCRv3(&det_model, &rec_model);
+  auto ppocr_v3 = fastdeploy::pipeline::PPOCRv3(&det_model, &cls_model, &rec_model);
+
+  // When users enable static shape infer for rec model, the batch size of cls and rec model must to be set to 1.
+  ppocr_v3.SetClsBatchSize(1);
+  ppocr_v3.SetRecBatchSize(1); 
+
+  if(!ppocr_v3.Initialized()){
+    std::cerr << "Failed to initialize PP-OCR." << std::endl;
+    return;
+  }
+
+  auto im = cv::imread(image_file);
+  
+  fastdeploy::vision::OCRResult result;
+  if (!ppocr_v3.Predict(im, &result)) {
+    std::cerr << "Failed to predict." << std::endl;
+    return;
+  }
+
+  std::cout << result.Str() << std::endl;
+
+  auto vis_im = fastdeploy::vision::VisOcr(im, result);
+  cv::imwrite("vis_result.jpg", vis_im);
+  std::cout << "Visualized result saved in ./vis_result.jpg" << std::endl;
+}
+
+int main(int argc, char* argv[]) {
+  if (argc < 7) {
+    std::cout << "Usage: infer_demo path/to/det_model path/to/cls_model "
+                 "path/to/rec_model path/to/rec_label_file path/to/image "
+                 "run_option, "
+                 "e.g ./infer_demo ./ch_PP-OCRv3_det_infer "
+                 "./ch_ppocr_mobile_v2.0_cls_infer ./ch_PP-OCRv3_rec_infer "
+                 "./ppocr_keys_v1.txt ./12.jpg 0"
+              << std::endl;
+    std::cout << "The data type of run_option is int, 0: run with cpu; 1: run "
+                 "with ascend."
+              << std::endl;
+    return -1;
+  }
+
+  fastdeploy::RuntimeOption option;
+  int flag = std::atoi(argv[6]);
+
+  if (flag == 0) {
+    option.UseCpu(); 
+  } else if (flag == 1) {
+    option.UseAscend();
+  }
+
+  std::string det_model_dir = argv[1];
+  std::string cls_model_dir = argv[2];
+  std::string rec_model_dir = argv[3];
+  std::string rec_label_file = argv[4];
+  std::string test_image = argv[5];
+  InitAndInfer(det_model_dir, cls_model_dir, rec_model_dir, rec_label_file, test_image, option);
+  return 0;
+}

examples/vision/ocr/PP-OCRv3/python/README.md

@@ -35,8 +35,8 @@ python infer.py --det_model ch_PP-OCRv3_det_infer --cls_model ch_ppocr_mobile_v2
 python infer.py --det_model ch_PP-OCRv3_det_infer --cls_model ch_ppocr_mobile_v2.0_cls_infer --rec_model ch_PP-OCRv3_rec_infer --rec_label_file ppocr_keys_v1.txt --image 12.jpg --device gpu --backend trt
 # 昆仑芯XPU推理
 python infer.py --det_model ch_PP-OCRv3_det_infer --cls_model ch_ppocr_mobile_v2.0_cls_infer --rec_model ch_PP-OCRv3_rec_infer --rec_label_file ppocr_keys_v1.txt --image 12.jpg --device kunlunxin
-# 华为昇腾推理
+# 华为昇腾推理,需要使用静态shape脚本, 若用户需要连续地预测图片, 输入图片尺寸需要准备为统一尺寸
-python infer.py --det_model ch_PP-OCRv3_det_infer --cls_model ch_ppocr_mobile_v2.0_cls_infer --rec_model ch_PP-OCRv3_rec_infer --rec_label_file ppocr_keys_v1.txt --image 12.jpg --device ascend
+python infer_static_shape.py --det_model ch_PP-OCRv3_det_infer --cls_model ch_ppocr_mobile_v2.0_cls_infer --rec_model ch_PP-OCRv3_rec_infer --rec_label_file ppocr_keys_v1.txt --image 12.jpg --device ascend
 ```
 
 运行完成可视化结果如下图所示

examples/vision/ocr/PP-OCRv3/python/infer.py

@@ -58,43 +58,113 @@ def parse_arguments():
         type=int,
         default=9,
         help="Number of threads while inference on CPU.")
+    parser.add_argument(
+        "--cls_bs",
+        type=int,
+        default=1,
+        help="Classification model inference batch size.")
+    parser.add_argument(
+        "--rec_bs",
+        type=int,
+        default=6,
+        help="Recognition model inference batch size")
     return parser.parse_args()
 
 
 def build_option(args):
-    option = fd.RuntimeOption()
-    if args.device.lower() == "gpu":
-        option.use_gpu(0)
 
-    option.set_cpu_thread_num(args.cpu_thread_num)
+    det_option = fd.RuntimeOption()
+    cls_option = fd.RuntimeOption()
+    rec_option = fd.RuntimeOption()
+
+    det_option.set_cpu_thread_num(args.cpu_thread_num)
+    cls_option.set_cpu_thread_num(args.cpu_thread_num)
+    rec_option.set_cpu_thread_num(args.cpu_thread_num)
+
+    if args.device.lower() == "gpu":
+        det_option.use_gpu(args.device_id)
+        cls_option.use_gpu(args.device_id)
+        rec_option.use_gpu(args.device_id)
 
     if args.device.lower() == "kunlunxin":
-        option.use_kunlunxin()
+        det_option.use_kunlunxin()
-        return option
+        cls_option.use_kunlunxin()
+        rec_option.use_kunlunxin()
 
-    if args.device.lower() == "ascend":
+        return det_option, cls_option, rec_option
-        option.use_ascend()
-        return option
 
     if args.backend.lower() == "trt":
         assert args.device.lower(
         ) == "gpu", "TensorRT backend require inference on device GPU."
-        option.use_trt_backend()
+        det_option.use_trt_backend()
+        cls_option.use_trt_backend()
+        rec_option.use_trt_backend()
+
+        # 设置trt input shape
+        # 如果用户想要自己改动检测模型的输入shape, 我们建议用户把检测模型的长和高设置为32的倍数.
+        det_option.set_trt_input_shape("x", [1, 3, 64, 64], [1, 3, 640, 640],
+                                       [1, 3, 960, 960])
+        cls_option.set_trt_input_shape("x", [1, 3, 48, 10],
+                                       [args.cls_bs, 3, 48, 320],
+                                       [args.cls_bs, 3, 48, 1024])
+        rec_option.set_trt_input_shape("x", [1, 3, 48, 10],
+                                       [args.rec_bs, 3, 48, 320],
+                                       [args.rec_bs, 3, 48, 2304])
+
+        # 用户可以把TRT引擎文件保存至本地
+        det_option.set_trt_cache_file(args.det_model + "/det_trt_cache.trt")
+        cls_option.set_trt_cache_file(args.cls_model + "/cls_trt_cache.trt")
+        rec_option.set_trt_cache_file(args.rec_model + "/rec_trt_cache.trt")
+
     elif args.backend.lower() == "pptrt":
         assert args.device.lower(
         ) == "gpu", "Paddle-TensorRT backend require inference on device GPU."
-        option.use_trt_backend()
+        det_option.use_trt_backend()
-        option.enable_paddle_trt_collect_shape()
+        det_option.enable_paddle_trt_collect_shape()
-        option.enable_paddle_to_trt()
+        det_option.enable_paddle_to_trt()
+
+        cls_option.use_trt_backend()
+        cls_option.enable_paddle_trt_collect_shape()
+        cls_option.enable_paddle_to_trt()
+
+        rec_option.use_trt_backend()
+        rec_option.enable_paddle_trt_collect_shape()
+        rec_option.enable_paddle_to_trt()
+
+        # 设置trt input shape
+        # 如果用户想要自己改动检测模型的输入shape, 我们建议用户把检测模型的长和高设置为32的倍数.
+        det_option.set_trt_input_shape("x", [1, 3, 64, 64], [1, 3, 640, 640],
+                                       [1, 3, 960, 960])
+        cls_option.set_trt_input_shape("x", [1, 3, 48, 10],
+                                       [args.cls_bs, 3, 48, 320],
+                                       [args.cls_bs, 3, 48, 1024])
+        rec_option.set_trt_input_shape("x", [1, 3, 48, 10],
+                                       [args.rec_bs, 3, 48, 320],
+                                       [args.rec_bs, 3, 48, 2304])
+
+        # 用户可以把TRT引擎文件保存至本地
+        det_option.set_trt_cache_file(args.det_model)
+        cls_option.set_trt_cache_file(args.cls_model)
+        rec_option.set_trt_cache_file(args.rec_model)
+
     elif args.backend.lower() == "ort":
-        option.use_ort_backend()
+        det_option.use_ort_backend()
+        cls_option.use_ort_backend()
+        rec_option.use_ort_backend()
+
     elif args.backend.lower() == "paddle":
-        option.use_paddle_infer_backend()
+        det_option.use_paddle_infer_backend()
+        cls_option.use_paddle_infer_backend()
+        rec_option.use_paddle_infer_backend()
+
     elif args.backend.lower() == "openvino":
         assert args.device.lower(
         ) == "cpu", "OpenVINO backend require inference on device CPU."
-        option.use_openvino_backend()
+        det_option.use_openvino_backend()
-    return option
+        cls_option.use_openvino_backend()
+        rec_option.use_openvino_backend()
+
+    return det_option, cls_option, rec_option
 
 
 args = parse_arguments()
@@ -111,49 +181,18 @@ rec_params_file = os.path.join(args.rec_model, "inference.pdiparams")
 rec_label_file = args.rec_label_file
 
 # 对于三个模型，均采用同样的部署配置
-# 用户也可根据自行需求分别配置
+# 用户也可根据自己的需求,个性化配置
-runtime_option = build_option(args)
+det_option, cls_option, rec_option = build_option(args)
 
-# PPOCR的cls和rec模型现在已经支持推理一个Batch的数据
-# 定义下面两个变量后, 可用于设置trt输入shape, 并在PPOCR模型初始化后, 完成Batch推理设置
-# 当用户要把PP-OCR部署在对动态shape推理支持有限的设备上时,(例如华为昇腾)
-# 需要把cls_batch_size和rec_batch_size都设置为1.
-cls_batch_size = 1
-rec_batch_size = 6
-
-# 当使用TRT时，分别给三个模型的runtime设置动态shape,并完成模型的创建.
-# 注意: 需要在检测模型创建完成后，再设置分类模型的动态输入并创建分类模型, 识别模型同理.
-# 如果用户想要自己改动检测模型的输入shape, 我们建议用户把检测模型的长和高设置为32的倍数.
-det_option = runtime_option
-det_option.set_trt_input_shape("x", [1, 3, 64, 64], [1, 3, 640, 640],
-                               [1, 3, 960, 960])
-# 用户可以把TRT引擎文件保存至本地
-# det_option.set_trt_cache_file(args.det_model  + "/det_trt_cache.trt")
 det_model = fd.vision.ocr.DBDetector(
     det_model_file, det_params_file, runtime_option=det_option)
 
-cls_option = runtime_option
-cls_option.set_trt_input_shape("x", [1, 3, 48, 10],
-                               [cls_batch_size, 3, 48, 320],
-                               [cls_batch_size, 3, 48, 1024])
-# 用户可以把TRT引擎文件保存至本地
-# cls_option.set_trt_cache_file(args.cls_model  + "/cls_trt_cache.trt")
 cls_model = fd.vision.ocr.Classifier(
     cls_model_file, cls_params_file, runtime_option=cls_option)
 
-rec_option = runtime_option
-rec_option.set_trt_input_shape("x", [1, 3, 48, 10],
-                               [rec_batch_size, 3, 48, 320],
-                               [rec_batch_size, 3, 48, 2304])
-# 用户可以把TRT引擎文件保存至本地
-# rec_option.set_trt_cache_file(args.rec_model  + "/rec_trt_cache.trt")
 rec_model = fd.vision.ocr.Recognizer(
     rec_model_file, rec_params_file, rec_label_file, runtime_option=rec_option)
 
-# 当用户要把PP-OCR部署在对动态shape推理支持有限的设备上时,(例如华为昇腾)
-# 需要使用下行代码, 来启用rec模型的静态shape推理.
-# rec_model.preprocessor.static_shape_infer = True
-
 # 创建PP-OCR，串联3个模型，其中cls_model可选，如无需求，可设置为None
 ppocr_v3 = fd.vision.ocr.PPOCRv3(
     det_model=det_model, cls_model=cls_model, rec_model=rec_model)
@@ -161,8 +200,8 @@ ppocr_v3 = fd.vision.ocr.PPOCRv3(
 # 给cls和rec模型设置推理时的batch size
 # 此值能为-1, 和1到正无穷
 # 当此值为-1时, cls和rec模型的batch size将默认和det模型检测出的框的数量相同
-ppocr_v3.cls_batch_size = cls_batch_size
+ppocr_v3.cls_batch_size = args.cls_bs
-ppocr_v3.rec_batch_size = rec_batch_size
+ppocr_v3.rec_batch_size = args.rec_bs
 
 # 预测图片准备
 im = cv2.imread(args.image)

examples/vision/ocr/PP-OCRv3/python/infer_static_shape.py

@@ -0,0 +1,114 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import fastdeploy as fd
+import cv2
+import os
+
+
+def parse_arguments():
+    import argparse
+    import ast
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--det_model", required=True, help="Path of Detection model of PPOCR.")
+    parser.add_argument(
+        "--cls_model",
+        required=True,
+        help="Path of Classification model of PPOCR.")
+    parser.add_argument(
+        "--rec_model",
+        required=True,
+        help="Path of Recognization model of PPOCR.")
+    parser.add_argument(
+        "--rec_label_file",
+        required=True,
+        help="Path of Recognization model of PPOCR.")
+    parser.add_argument(
+        "--image", type=str, required=True, help="Path of test image file.")
+    parser.add_argument(
+        "--device",
+        type=str,
+        default='cpu',
+        help="Type of inference device, support 'cpu', 'kunlunxin' or 'gpu'.")
+    parser.add_argument(
+        "--cpu_thread_num",
+        type=int,
+        default=9,
+        help="Number of threads while inference on CPU.")
+    return parser.parse_args()
+
+
+def build_option(args):
+
+    det_option = fd.RuntimeOption()
+    cls_option = fd.RuntimeOption()
+    rec_option = fd.RuntimeOption()
+
+    # 当前需要对PP-OCR启用静态shape推理的硬件只有昇腾.
+    if args.device.lower() == "ascend":
+        det_option.use_ascend()
+        cls_option.use_ascend()
+        rec_option.use_ascend()
+
+    return det_option, cls_option, rec_option
+
+
+args = parse_arguments()
+
+# Detection模型, 检测文字框
+det_model_file = os.path.join(args.det_model, "inference.pdmodel")
+det_params_file = os.path.join(args.det_model, "inference.pdiparams")
+# Classification模型，方向分类，可选
+cls_model_file = os.path.join(args.cls_model, "inference.pdmodel")
+cls_params_file = os.path.join(args.cls_model, "inference.pdiparams")
+# Recognition模型，文字识别模型
+rec_model_file = os.path.join(args.rec_model, "inference.pdmodel")
+rec_params_file = os.path.join(args.rec_model, "inference.pdiparams")
+rec_label_file = args.rec_label_file
+
+det_option, cls_option, rec_option = build_option(args)
+
+det_model = fd.vision.ocr.DBDetector(
+    det_model_file, det_params_file, runtime_option=det_option)
+
+cls_model = fd.vision.ocr.Classifier(
+    cls_model_file, cls_params_file, runtime_option=cls_option)
+
+rec_model = fd.vision.ocr.Recognizer(
+    rec_model_file, rec_params_file, rec_label_file, runtime_option=rec_option)
+
+# Rec模型启用静态shape推理
+rec_model.preprocessor.static_shape_infer = True
+
+# 创建PP-OCR，串联3个模型，其中cls_model可选，如无需求，可设置为None
+ppocr_v3 = fd.vision.ocr.PPOCRv3(
+    det_model=det_model, cls_model=cls_model, rec_model=rec_model)
+
+# Cls模型和Rec模型的batch size 必须设置为1, 开启静态shape推理
+ppocr_v3.cls_batch_size = 1
+ppocr_v3.rec_batch_size = 1
+
+# 预测图片准备
+im = cv2.imread(args.image)
+
+#预测并打印结果
+result = ppocr_v3.predict(im)
+
+print(result)
+
+# 可视化结果
+vis_im = fd.vision.vis_ppocr(im, result)
+cv2.imwrite("visualized_result.jpg", vis_im)
+print("Visualized result save in ./visualized_result.jpg")

examples/vision/segmentation/paddleseg/cpp/README.md

@@ -46,6 +46,9 @@ wget https://paddleseg.bj.bcebos.com/dygraph/demo/cityscapes_demo.png
 以上命令只适用于Linux或MacOS, Windows下SDK的使用方式请参考:  
 - [如何在Windows中使用FastDeploy C++ SDK](../../../../../docs/cn/faq/use_sdk_on_windows.md)
 
+如果用户使用华为昇腾NPU部署, 请参考以下方式在部署前初始化部署环境:
+- [如何使用华为昇腾NPU部署](../../../../../docs/cn/faq/use_sdk_on_ascend.md)
+
 ## PaddleSeg C++接口
 
 ### PaddleSeg类