[Example] Merge Download Paddle Model, Paddle->ONNX->MLIR->BModel (#1643)

* fix infer.py and README

* [Example] Merge Download Paddle Model, Paddle->Onnx->Mlir->Bmodel and
inference into infer.py. Modify README.md

* modify pp_liteseg sophgo infer.py and README.md

* fix PPOCR,PPYOLOE,PICODET,LITESEG sophgo infer.py and README.md

* fix memory overflow problem while inferring with sophgo backend

* fix memory overflow problem while inferring with sophgo backend

---------

Co-authored-by: DefTruth <31974251+DefTruth@users.noreply.github.com>
Co-authored-by: xuyizhou <yizhou.xu@sophgo.com>

examples/vision/detection/paddledetection/sophgo/cpp/infer_picodet.cc

@@ -31,7 +31,7 @@ void SophgoInfer(const std::string& model_dir, const std::string& image_file) {
   auto model = fastdeploy::vision::detection::PicoDet(
       model_file, params_file, config_file, option, format);
 
-  model.GetPostprocessor().ApplyDecodeAndNMS();
+  model.GetPostprocessor().ApplyNMS();
 
   auto im = cv::imread(image_file);
 

examples/vision/detection/paddledetection/sophgo/cpp/infer_ppyoloe.cc

@@ -31,7 +31,7 @@ void SophgoInfer(const std::string& model_dir, const std::string& image_file) {
   auto model = fastdeploy::vision::detection::PPYOLOE(
       model_file, params_file, config_file, option, format);
 
-  model.GetPostprocessor().ApplyDecodeAndNMS();
+  model.GetPostprocessor().ApplyNMS();
 
   auto im = cv::imread(image_file);
 

examples/vision/detection/paddledetection/sophgo/cpp/infer_yolov8.cc

@@ -30,7 +30,7 @@ void SophgoInfer(const std::string& model_dir, const std::string& image_file) {
   
   auto model = fastdeploy::vision::detection::PaddleYOLOv8(model_file, params_file, config_file, option, format);
 
-  model.GetPostprocessor().ApplyDecodeAndNMS();
+  model.GetPostprocessor().ApplyNMS();
 
   auto im = cv::imread(image_file);
 

examples/vision/detection/paddledetection/sophgo/python/README.md

@@ -15,13 +15,21 @@ cd FastDeploy/examples/vision/detection/paddledetection/sophgo/python
 wget https://gitee.com/paddlepaddle/PaddleDetection/raw/release/2.4/demo/000000014439.jpg
 
 # 推理
-#ppyoloe推理示例
+# ppyoloe推理示例
-python3 infer_ppyoloe.py --model_file model/ppyoloe_crn_s_300e_coco_1684x_f32.bmodel --config_file model/infer_cfg.yml --image ./000000014439.jpg
+# 指定--auto True，自动完成模型准备、转换和推理，需要指定PaddleDetection路径
+python3 infer_ppyoloe.py --auto True --pp_detect_path {Path to PaddleDetection} --model_file '' --config_file '' --image ''
 
-#picodet推理示例
+# 指定--auto False，需要用户指定模型、配置文件和图片路径，不需要指定PaddleDetection路径。
-python3 infer_picodet.py --model_file model/picodet_s_416_coco_lcnet_1684x_f32.bmodel --config_file model/infer_cfg.yml --image ./000000014439.jpg
+python3 infer_ppyoloe.py --auto False --pp_detect_path '' --model_file model/ppyoloe_crn_s_300e_coco_1684x_f32.bmodel --config_file model/infer_cfg.yml --image ./000000014439.jpg
 
-#yolov8推理示例
+# picodet推理示例
+# 指定--auto True，自动完成模型准备、转换和推理，需要指定PaddleDetection路径
+python3 infer_picodet.py --auto True --pp_detect_path {Path to PaddleDetection} --model_file '' --config_file '' --image ''
+
+# 指定--auto False，需要用户指定模型、配置文件和图片路径，不需要指定PaddleDetection路径。
+python3 infer_picodet.py --auto False --pp_detect_path '' --model_file model/ppyoloe_crn_s_300e_coco_1684x_f32.bmodel --config_file model/infer_cfg.yml --image ./000000014439.jpg
+
+# yolov8推理示例
 python3 infer_yolov8.py --model_file model/yolov8s_s_300e_coco_1684x_f32.bmodel --config_file model/infer_cfg.yml --image ./000000014439.jpg
 # 运行完成后返回结果如下所示
 可视化结果存储在sophgo_result.jpg中

examples/vision/detection/paddledetection/sophgo/python/infer_picodet.py

@@ -14,12 +14,101 @@
 import fastdeploy as fd
 import cv2
 import os
+from subprocess import run
+from prepare_npz import prepare
+
+def export_model(args):
+    PPDetection_path = args.pp_detect_path
+
+    export_str = 'python3 tools/export_model.py \
+                -c configs/picodet/picodet_s_320_coco_lcnet.yml \
+                --output_dir=output_inference \
+                -o weights=https://paddledet.bj.bcebos.com/models/picodet_s_320_coco_lcnet.pdparams'
+    cur_path = os.getcwd()
+    os.chdir(PPDetection_path)
+    print(export_str)
+    run(export_str, shell=True)
+    cp_str = 'cp -r ./output_inference/picodet_s_320_coco_lcnet ' + cur_path
+    print(cp_str)
+    run(cp_str, shell=True)
+    os.chdir(cur_path)
+
+def paddle2onnx():
+    convert_str = 'paddle2onnx --model_dir picodet_s_320_coco_lcnet/ \
+                    --model_filename model.pdmodel \
+                    --params_filename model.pdiparams \
+                    --save_file picodet_s_320_coco_lcnet.onnx \
+                    --enable_dev_version True'
+    print(convert_str)
+    run(convert_str, shell=True)
+    fix_shape_str = 'python3 -m paddle2onnx.optimize \
+                    --input_model picodet_s_320_coco_lcnet.onnx \
+                    --output_model picodet_s_320_coco_lcnet.onnx \
+                    --input_shape_dict "{\'image\':[1,3,640,640]}"'
+    print(fix_shape_str)
+    run(fix_shape_str, shell=True)
+
+def mlir_prepare():
+    mlir_path = os.getenv("MODEL_ZOO_PATH")
+    mlir_path = mlir_path[:-13]
+    regression_path = os.path.join(mlir_path, 'regression')
+    mv_str_list = ['mkdir picodet', 
+                'cp -rf ' + os.path.join(regression_path, 'dataset/COCO2017/') + ' ./picodet', 
+                'cp -rf ' + os.path.join(regression_path, 'image/') + ' ./picodet',
+                'cp picodet_s_320_coco_lcnet.onnx ./picodet',
+                'mkdir ./picodet/workspace']
+    for str in mv_str_list:
+        print(str)
+        run(str, shell=True)
+
+def image_prepare():
+    img_str = 'wget https://gitee.com/paddlepaddle/PaddleDetection/raw/release/2.4/demo/000000014439.jpg'
+    if not os.path.exists('000000014439.jpg'):
+        print(img_str)
+        run(img_str, shell=True)
+    prepare('000000014439.jpg', [320, 320])
+    cp_npz_str = 'cp ./inputs.npz ./picodet'
+    print(cp_npz_str)
+    run(cp_npz_str, shell=True)
+
+def onnx2mlir():
+    transform_str = 'model_transform.py \
+            --model_name picodet_s_320_coco_lcnet \
+            --model_def ../picodet_s_320_coco_lcnet.onnx \
+            --input_shapes [[1,3,320,320],[1,2]] \
+            --keep_aspect_ratio \
+            --pixel_format rgb \
+            --output_names p2o.Div.79,p2o.Concat.9 \
+            --test_input ../inputs.npz \
+            --test_result picodet_s_320_coco_lcnet_top_outputs.npz \
+            --mlir picodet_s_320_coco_lcnet.mlir'
+    os.chdir('./picodet/workspace')
+    print(transform_str)
+    run(transform_str, shell=True)
+    os.chdir('../../')
+
+def mlir2bmodel():
+    deploy_str = 'model_deploy.py \
+            --mlir picodet_s_320_coco_lcnet.mlir \
+            --quantize F32 \
+            --chip bm1684x \
+            --test_input picodet_s_320_coco_lcnet_in_f32.npz \
+            --test_reference picodet_s_320_coco_lcnet_top_outputs.npz \
+            --model picodet_s_320_coco_lcnet_1684x_f32.bmodel'
+    os.chdir('./picodet/workspace')
+    print(deploy_str)
+    run(deploy_str, shell=True)
+    os.chdir('../../')
 
 
 def parse_arguments():
     import argparse
     import ast
     parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--auto", required=True, help="Auto download, convert, compile and infer if True")
+    parser.add_argument(
+        "--pp_detect_path", default='/workspace/PaddleDetection', help="Path of PaddleDetection folder")
     parser.add_argument(
         "--model_file", required=True, help="Path of sophgo model.")
     parser.add_argument("--config_file", required=True, help="Path of config.")
@@ -31,10 +120,18 @@ def parse_arguments():
 if __name__ == "__main__":
     args = parse_arguments()
 
-    model_file = args.model_file
+    if args.auto:
-    params_file = ""
+        export_model()
-    config_file = args.config_file
+        paddle2onnx()
+        mlir_prepare()
+        image_prepare()
+        onnx2mlir()
+        mlir2bmodel()
 
+    model_file = './picodet/workspace/picodet_s_320_coco_lcnet_1684x_f32.bmodel' if args.auto else args.model_file
+    params_file = ""
+    config_file = './picodet_s_320_coco_lcnet/infer_cfg.yml' if args.auto else args.config_file
+    img_file = './000000014439.jpg' if args.auto else args.image
     # 配置runtime，加载模型
     runtime_option = fd.RuntimeOption()
     runtime_option.use_sophgo()
@@ -46,14 +143,14 @@ if __name__ == "__main__":
         runtime_option=runtime_option,
         model_format=fd.ModelFormat.SOPHGO)
 
-    model.postprocessor.apply_decode_and_nms()
+    model.postprocessor.apply_nms()
 
     # 预测图片分割结果
-    im = cv2.imread(args.image)
+    im = cv2.imread(img_file)
     result = model.predict(im)
     print(result)
 
     # 可视化结果
     vis_im = fd.vision.vis_detection(im, result, score_threshold=0.5)
     cv2.imwrite("sophgo_result.jpg", vis_im)
-    print("Visualized result save in ./sophgo_result.jpg")
+    print("Visualized result save in ./sophgo_result_picodet.jpg")

examples/vision/detection/paddledetection/sophgo/python/infer_ppyoloe.py

@@ -14,12 +14,17 @@
 import fastdeploy as fd
 import cv2
 import os
-
+from subprocess import run
+from prepare_npz import prepare
 
 def parse_arguments():
     import argparse
     import ast
     parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--auto", required=True, help="Auto download, convert, compile and infer if True")
+    parser.add_argument(
+        "--pp_detect_path", default='/workspace/PaddleDetection', help="Path of PaddleDetection folder")
     parser.add_argument(
         "--model_file", required=True, help="Path of sophgo model.")
     parser.add_argument("--config_file", required=True, help="Path of config.")
@@ -27,13 +32,102 @@ def parse_arguments():
         "--image", type=str, required=True, help="Path of test image file.")
     return parser.parse_args()
 
+def export_model(args):
+    PPDetection_path = args.pp_detect_path
+    export_str = 'python3 tools/export_model.py \
+                -c configs/ppyoloe/ppyoloe_crn_s_300e_coco.yml \
+                -output_dir=output_inference \
+                -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_crn_s_300e_coco.pdparams'
+    cur_path = os.getcwd()
+    os.chdir(PPDetection_path)
+    print(export_str)
+    run(export_str, shell=True)
+    cp_str = 'cp -r ./output_inference/ppyoloe_crn_s_300e_coco ' + cur_path
+    print(cp_str)
+    run(cp_str, shell=True)
+    os.chdir(cur_path)
+
+def paddle2onnx():
+    convert_str = 'paddle2onnx --model_dir ppyoloe_crn_s_300e_coco \
+                --model_filename model.pdmodel \
+                --params_filename model.pdiparams \
+                --save_file ppyoloe_crn_s_300e_coco.onnx \
+                --enable_dev_version True'
+    print(convert_str)
+    run(convert_str, shell=True)
+    fix_shape_str = 'python3 -m paddle2onnx.optimize --input_model ppyoloe_crn_s_300e_coco.onnx \
+                                --output_model ppyoloe_crn_s_300e_coco.onnx \
+                                --input_shape_dict "{\'image\':[1,3,640,640]}"'
+    print(fix_shape_str)
+    run(fix_shape_str, shell=True)
+
+def mlir_prepare():
+    mlir_path = os.getenv("MODEL_ZOO_PATH")
+    mlir_path = mlir_path[:-13]
+    regression_path = os.path.join(mlir_path, 'regression')
+    mv_str_list = ['mkdir ppyoloe', 
+                'cp -rf ' + os.path.join(regression_path, 'dataset/COCO2017/') + ' ./ppyoloe', 
+                'cp -rf ' + os.path.join(regression_path, 'image/') + ' ./ppyoloe',
+                'cp ppyoloe_crn_s_300e_coco.onnx ./ppyoloe',
+                'mkdir ./ppyoloe/workspace']
+    for str in mv_str_list:
+        print(str)
+        run(str, shell=True)
+
+def image_prepare():
+    img_str = 'wget https://gitee.com/paddlepaddle/PaddleDetection/raw/release/2.4/demo/000000014439.jpg'
+    if not os.path.exists('000000014439.jpg'):
+        print(img_str)
+        run(img_str, shell=True)
+    prepare('000000014439.jpg', [640, 640])
+    cp_npz_str = 'cp ./inputs.npz ./ppyoloe'
+    print(cp_npz_str)
+    run(cp_npz_str, shell=True)
+
+def onnx2mlir():
+    transform_str = 'model_transform.py \
+            --model_name ppyoloe_crn_s_300e_coco \
+            --model_def ../ppyoloe_crn_s_300e_coco.onnx \
+            --input_shapes [[1,3,640,640],[1,2]] \
+            --keep_aspect_ratio \
+            --pixel_format rgb \
+            --output_names p2o.Div.1,p2o.Concat.29 \
+            --test_input ../inputs.npz \
+            --test_result ppyoloe_crn_s_300e_coco_top_outputs.npz \
+            --mlir ppyoloe_crn_s_300e_coco.mlir'
+    os.chdir('./ppyoloe/workspace')
+    print(transform_str)
+    run(transform_str, shell=True)
+    os.chdir('../../')
+
+def mlir2bmodel():
+    deploy_str = 'model_deploy.py \
+            --mlir ppyoloe_crn_s_300e_coco.mlir \
+            --quantize F32 \
+            --chip bm1684x \
+            --test_input ppyoloe_crn_s_300e_coco_in_f32.npz \
+            --test_reference ppyoloe_crn_s_300e_coco_top_outputs.npz \
+            --model ppyoloe_crn_s_300e_coco_1684x_f32.bmodel'
+    os.chdir('./ppyoloe/workspace')
+    print(deploy_str)
+    run(deploy_str, shell=True)
+    os.chdir('../../')
 
 if __name__ == "__main__":
     args = parse_arguments()
 
-    model_file = args.model_file
+    if args.auto:
+        export_model(args)
+        paddle2onnx()
+        mlir_prepare()
+        image_prepare()
+        onnx2mlir()
+        mlir2bmodel()
+
+    model_file = './ppyoloe/workspace/ppyoloe_crn_s_300e_coco_1684x_f32.bmodel' if args.auto else args.model_file
     params_file = ""
-    config_file = args.config_file
+    config_file = './ppyoloe_crn_s_300e_coco/infer_cfg.yml' if args.auto else args.config_file
+    image_file = './000000014439.jpg' if args.auto else args.image
 
     # 配置runtime，加载模型
     runtime_option = fd.RuntimeOption()
@@ -46,14 +140,14 @@ if __name__ == "__main__":
         runtime_option=runtime_option,
         model_format=fd.ModelFormat.SOPHGO)
 
-    model.postprocessor.apply_decode_and_nms()
+    model.postprocessor.apply_nms()
 
     # 预测图片分割结果
-    im = cv2.imread(args.image)
+    im = cv2.imread(image_file)
     result = model.predict(im)
     print(result)
 
     # 可视化结果
     vis_im = fd.vision.vis_detection(im, result, score_threshold=0.5)
     cv2.imwrite("sophgo_result.jpg", vis_im)
-    print("Visualized result save in ./sophgo_result.jpg")
+    print("Visualized result save in ./sophgo_result_ppyoloe.jpg")

examples/vision/detection/paddledetection/sophgo/python/prepare_npz.py

@@ -0,0 +1,17 @@
+import cv2
+import numpy as np
+
+def prepare(img_path, sz):
+    im = cv2.imread(img_path)
+    im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
+    im = cv2.resize(im, sz)
+    im = im.transpose((2,0,1))
+    im = im[np.newaxis,...]
+    im_scale_y = sz[0] / float(im.shape[2])
+    im_scale_x = sz[1] / float(im.shape[3])
+    inputs = {}
+    inputs['image'] = np.array(im).astype('float32')
+    # scale = np.array([im_scale_y, im_scale_x])
+    # scale = scale[np.newaxis,...]
+    inputs['scale_factor'] = np.array(([im_scale_y, im_scale_x], )).astype('float32')
+    np.savez('inputs.npz', image=inputs['image'], scale_factor=inputs['scale_factor'])

examples/vision/detection/yolov5/sophgo/python/README.md

@@ -15,7 +15,10 @@ cd FastDeploy/examples/vision/detection/yolov5/sophgo/python
 # Download images.
 wget https://gitee.com/paddlepaddle/PaddleDetection/raw/release/2.4/demo/000000014439.jpg
 
-# Inference.
+# Set --auto True, automatic inference.
+python3 infer.py --auto True
+
+# Set --auto False, need to set the model path and image path manually.
 python3 infer.py --model_file ./bmodel/yolov5s_1684x_f32.bmodel --image 000000014439.jpg
 
 # The returned result.

examples/vision/detection/yolov5/sophgo/python/README_CN.md

@@ -15,8 +15,11 @@ cd FastDeploy/examples/vision/detection/yolov5/sophgo/python
 # 下载图片
 wget https://gitee.com/paddlepaddle/PaddleDetection/raw/release/2.4/demo/000000014439.jpg
 
-# 推理
+# 指定--auto True，自动完成下载数据、模型转换、推理
-python3 infer.py --model_file ./bmodel/yolov5s_1684x_f32.bmodel --image 000000014439.jpg
+python3 infer.py --auto True
+
+# 指定--auto False，则需要设置模型目录、图片目录
+python3 infer.py --auto False --model_file ./bmodel/yolov5s_1684x_f32.bmodel --image 000000014439.jpg
 
 # 运行完成后返回结果如下所示
 DetectionResult: [xmin, ymin, xmax, ymax, score, label_id]

examples/vision/detection/yolov5/sophgo/python/infer.py

@@ -1,27 +1,94 @@
 import fastdeploy as fd
 import cv2
 import os
-
+from subprocess import run
 
 def parse_arguments():
     import argparse
     import ast
     parser = argparse.ArgumentParser()
-    parser.add_argument("--model", required=True, help="Path of model.")
     parser.add_argument(
-        "--image", type=str, required=True, help="Path of test image file.")
+        "--auto", required=True, help="Auto download, convert, compile and infer if True")
+    parser.add_argument("--model", help="Path of model.")
+    parser.add_argument(
+        "--image", type=str, help="Path of test image file.")
 
     return parser.parse_args()
 
 
+def download():
+    download_model_str = 'wget https://bj.bcebos.com/paddlehub/fastdeploy/yolov5s.onnx'
+    if not os.path.exists('yolov5s.onnx'):
+        print(download_model_str)
+        run(download_model_str, shell=True)
+    download_img_str = 'wget https://gitee.com/paddlepaddle/PaddleDetection/raw/release/2.4/demo/000000014439.jpg'
+    if not os.path.exists('000000014439.jpg'):
+        print(download_img_str)
+        run(download_img_str, shell=True)
+    
+    
+def mlir_prepare():
+    mlir_path = os.getenv("MODEL_ZOO_PATH")
+    mlir_path = mlir_path[:-13]
+    regression_path = os.path.join(mlir_path, 'regression')
+    mv_str_list = ['mkdir YOLOv5s', 
+                'cp -rf ' + os.path.join(regression_path, 'dataset/COCO2017/') + ' ./YOLOv5s', 
+                'cp -rf ' + os.path.join(regression_path, 'image/') + ' ./YOLOv5s',
+                'cp yolov5s.onnx ./YOLOv5s',
+                'mkdir ./YOLOv5s/workspace']
+    for str in mv_str_list:
+        print(str)
+        run(str, shell=True)
+
+
+def onnx2mlir():
+    transform_str = 'model_transform.py \
+            --model_name yolov5s \
+            --model_def ../yolov5s.onnx \
+            --input_shapes [[1,3,640,640]] \
+            --mean 0.0,0.0,0.0 \
+            --scale 0.0039216,0.0039216,0.0039216 \
+            --keep_aspect_ratio \
+            --pixel_format rgb \
+            --output_names output0 \
+            --test_input ../image/dog.jpg \
+            --test_result yolov5s_top_outputs.npz \
+            --mlir yolov5s.mlir'
+    os.chdir('./YOLOv5s/workspace')
+    print(transform_str)
+    run(transform_str, shell=True)
+    os.chdir('../../')
+
+def mlir2bmodel():
+    deploy_str = 'model_deploy.py \
+                --mlir yolov5s.mlir \
+                --quantize F32 \
+                --chip bm1684x \
+                --test_input yolov5s_in_f32.npz \
+                --test_reference yolov5s_top_outputs.npz \
+                --model yolov5s_1684x_f32.bmodel'
+
+    os.chdir('./YOLOv5s/workspace')
+    print(deploy_str)
+    run(deploy_str, shell=True)
+    os.chdir('../../')
+
+
 args = parse_arguments()
 
+if args.auto:
+    download()
+    mlir_prepare()
+    onnx2mlir()
+    mlir2bmodel()
+
 # 配置runtime，加载模型
 runtime_option = fd.RuntimeOption()
 runtime_option.use_sophgo()
 
-model_file = args.model
+model_file = './YOLOv5s/workspace/yolov5s_1684x_f32.bmodel' if args.auto else args.model
 params_file = ""
+img_file = './000000014439.jpg' if args.auto else args.image
 
 model = fd.vision.detection.YOLOv5(
     model_file,
@@ -29,8 +96,9 @@ model = fd.vision.detection.YOLOv5(
     runtime_option=runtime_option,
     model_format=fd.ModelFormat.SOPHGO)
 
+
 # 预测图片分类结果
-im = cv2.imread(args.image)
+im = cv2.imread(img_file)
 result = model.predict(im)
 print(result)
 

examples/vision/ocr/PP-OCR/sophgo/python/README.md

@@ -27,11 +27,16 @@ cd PaddleOCR/deploy/fastdeploy/sophgo/python
 # 下载图片
 wget https://gitee.com/paddlepaddle/PaddleOCR/raw/release/2.6/doc/imgs/12.jpg
 
-#下载字典文件
+# 下载字典文件
 wget https://gitee.com/paddlepaddle/PaddleOCR/raw/release/2.6/ppocr/utils/ppocr_keys_v1.txt
 
 # 推理
-python3 infer.py --det_model ocr_bmodel/ch_PP-OCRv3_det_1684x_f32.bmodel \
+# --auto True，自动完成下载数据、模型转换、推理
+python3 infer.py --auto True --det_model '' --cls_model '' --rec_model '' --rec_label_file '' --image ''
+
+# --auto False，需要用户设置模型、字典、图片路径，进行推理
+python3 infer.py --auto False \
+                 --det_model ocr_bmodel/ch_PP-OCRv3_det_1684x_f32.bmodel \
                  --cls_model ocr_bmodel/ch_ppocr_mobile_v2.0_cls_1684x_f32.bmodel \
                  --rec_model ocr_bmodel/ch_PP-OCRv3_rec_1684x_f32.bmodel \
                  --rec_label_file ../ppocr_keys_v1.txt \  

examples/vision/ocr/PP-OCR/sophgo/python/infer.py

@@ -1,12 +1,15 @@
 import fastdeploy as fd
 import cv2
 import os
+from subprocess import run
 
 
 def parse_arguments():
     import argparse
     import ast
     parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--auto", required=True, help="Auto download, convert, compile and infer if True")
     parser.add_argument(
         "--det_model", required=True, help="Path of Detection model of PPOCR.")
     parser.add_argument(
@@ -27,22 +30,210 @@ def parse_arguments():
     return parser.parse_args()
 
 
+def getPPOCRv3():
+    cmd_str_det = 'wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.tar'
+    tar_str_det = 'tar xvf ch_PP-OCRv3_det_infer.tar'
+    cmd_str_cls = 'wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar'
+    tar_str_cls = 'tar xvf ch_ppocr_mobile_v2.0_cls_infer.tar'
+    cmd_str_rec = 'wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_infer.tar'
+    tar_str_rec = 'tar xvf ch_PP-OCRv3_rec_infer.tar'
+    cmd_str_img = 'wget https://gitee.com/paddlepaddle/PaddleOCR/raw/release/2.6/doc/imgs/12.jpg'
+    cmd_str_label = 'wget https://gitee.com/paddlepaddle/PaddleOCR/raw/release/2.6/ppocr/utils/ppocr_keys_v1.txt'
+    script_str = 'wget https://raw.githubusercontent.com/PaddlePaddle/Paddle2ONNX/develop/tools/paddle/paddle_infer_shape.py'
+    if not os.path.exists('ch_PP-OCRv3_det_infer.tar'):
+        print(cmd_str_det, tar_str_det)
+        run(cmd_str_det, shell=True)
+        run(tar_str_det, shell=True)
+    if not os.path.exists('ch_ppocr_mobile_v2.0_cls_infer.tar'):
+        print(cmd_str_cls, tar_str_cls)
+        run(cmd_str_cls, shell=True)
+        run(tar_str_cls, shell=True)
+    if not os.path.exists('ch_PP-OCRv3_rec_infer.tar'):
+        print(cmd_str_rec, tar_str_rec)
+        run(cmd_str_rec, shell=True)
+        run(tar_str_rec, shell=True)
+    if not os.path.exists('12.jpg'):
+        print(cmd_str_img)
+        run(cmd_str_img, shell=True)
+    if not os.path.exists('ppocr_keys_v1.txt'):
+        print(cmd_str_label)
+        run(cmd_str_label, shell=True)
+    if not os.path.exists('paddle_infer_shape.py'):
+        print(script_str)
+        run(script_str, shell=True)
+
+def fix_input_shape():
+    fix_det_str = 'python paddle_infer_shape.py --model_dir ch_PP-OCRv3_det_infer \
+                    --model_filename inference.pdmodel \
+                    --params_filename inference.pdiparams \
+                    --save_dir ch_PP-OCRv3_det_infer_fix \
+                    --input_shape_dict="{\'x\':[1,3,960,608]}"'
+    fix_rec_str = 'python paddle_infer_shape.py --model_dir ch_PP-OCRv3_rec_infer \
+                    --model_filename inference.pdmodel \
+                    --params_filename inference.pdiparams \
+                    --save_dir ch_PP-OCRv3_rec_infer_fix \
+                    --input_shape_dict="{\'x\':[1,3,48,320]}"'
+    fix_cls_str = 'python paddle_infer_shape.py --model_dir ch_ppocr_mobile_v2.0_cls_infer \
+                    --model_filename inference.pdmodel \
+                    --params_filename inference.pdiparams \
+                    --save_dir ch_PP-OCRv3_cls_infer_fix \
+                    --input_shape_dict="{\'x\':[1,3,48,192]}"'
+    print(fix_det_str)
+    run(fix_det_str, shell=True)
+    print(fix_rec_str)
+    run(fix_rec_str, shell=True)
+    print(fix_cls_str)
+    run(fix_cls_str, shell=True)
+
+
+def paddle2onnx():
+    cmd_str_det = 'paddle2onnx --model_dir ch_PP-OCRv3_det_infer_fix \
+            --model_filename inference.pdmodel \
+            --params_filename inference.pdiparams \
+            --save_file ch_PP-OCRv3_det_infer.onnx \
+            --enable_dev_version True'
+    cmd_str_cls = 'paddle2onnx --model_dir ch_PP-OCRv3_cls_infer_fix \
+            --model_filename inference.pdmodel \
+            --params_filename inference.pdiparams \
+            --save_file ch_PP-OCRv3_cls_infer.onnx \
+            --enable_dev_version True'
+    cmd_str_rec = 'paddle2onnx --model_dir ch_PP-OCRv3_rec_infer_fix \
+            --model_filename inference.pdmodel \
+            --params_filename inference.pdiparams \
+            --save_file ch_PP-OCRv3_rec_infer.onnx \
+            --enable_dev_version True'
+    print(cmd_str_det)
+    run(cmd_str_det, shell=True)
+    print(cmd_str_cls)
+    run(cmd_str_cls, shell=True)
+    print(cmd_str_rec)
+    run(cmd_str_rec, shell=True)
+
+def mlir_prepare():
+    mlir_path = os.getenv("MODEL_ZOO_PATH")
+    mlir_path = mlir_path[:-13]
+    regression_path = os.path.join(mlir_path, 'regression')
+    mv_str_list = ['mkdir ch_PP-OCRv3', 
+                'cp -rf ' + os.path.join(regression_path, 'dataset/COCO2017/') + ' ./ch_PP-OCRv3', 
+                'cp -rf ' + os.path.join(regression_path, 'image/') + ' ./ch_PP-OCRv3',
+                'mv ch_PP-OCRv3_det_infer.onnx ./ch_PP-OCRv3',
+                'mv ch_PP-OCRv3_rec_infer.onnx ./ch_PP-OCRv3',
+                'mv ch_PP-OCRv3_cls_infer.onnx ./ch_PP-OCRv3',
+                'mkdir ./ch_PP-OCRv3/workspace']
+    for str in mv_str_list:
+        print(str)
+        run(str, shell=True)
+
+
+def onnx2mlir():
+    transform_str_det = 'model_transform.py \
+        --model_name ch_PP-OCRv3_det \
+        --model_def ../ch_PP-OCRv3_det_infer.onnx \
+        --input_shapes [[1,3,960,608]] \
+        --mean 0.0,0.0,0.0 \
+        --scale 0.0039216,0.0039216,0.0039216 \
+        --keep_aspect_ratio \
+        --pixel_format rgb \
+        --output_names sigmoid_0.tmp_0 \
+        --test_input ../image/dog.jpg \
+        --test_result ch_PP-OCRv3_det_top_outputs.npz \
+        --mlir ./ch_PP-OCRv3_det.mlir'
+    transform_str_rec = 'model_transform.py \
+        --model_name ch_PP-OCRv3_rec \
+        --model_def ../ch_PP-OCRv3_rec_infer.onnx \
+        --input_shapes [[1,3,48,320]] \
+        --mean 0.0,0.0,0.0 \
+        --scale 0.0039216,0.0039216,0.0039216 \
+        --keep_aspect_ratio \
+        --pixel_format rgb \
+        --output_names softmax_5.tmp_0 \
+        --test_input ../image/dog.jpg \
+        --test_result ch_PP-OCRv3_rec_top_outputs.npz \
+        --mlir ./ch_PP-OCRv3_rec.mlir'
+    transform_str_cls = 'model_transform.py \
+        --model_name ch_PP-OCRv3_cls \
+        --model_def ../ch_PP-OCRv3_cls_infer.onnx \
+        --input_shapes [[1,3,48,192]] \
+        --mean 0.0,0.0,0.0 \
+        --scale 0.0039216,0.0039216,0.0039216 \
+        --keep_aspect_ratio \
+        --pixel_format rgb \
+        --output_names softmax_0.tmp_0 \
+        --test_input ../image/dog.jpg \
+        --test_result ch_PP-OCRv3_cls_top_outputs.npz \
+        --mlir ./ch_PP-OCRv3_cls.mlir'
+
+    os.chdir('./ch_PP-OCRv3/workspace/')
+
+    print(transform_str_det)
+    run(transform_str_det, shell=True)
+
+    print(transform_str_rec)
+    run(transform_str_rec, shell=True)
+
+    print(transform_str_cls)
+    run(transform_str_cls, shell=True)
+
+    os.chdir('../../')
+
+def mlir2bmodel():
+    det_str = 'model_deploy.py \
+        --mlir ./ch_PP-OCRv3_det.mlir \
+        --quantize F32 \
+        --chip bm1684x \
+        --test_input ./ch_PP-OCRv3_det_in_f32.npz \
+        --test_reference ./ch_PP-OCRv3_det_top_outputs.npz \
+        --model ./ch_PP-OCRv3_det_1684x_f32.bmodel'
+    rec_str = 'model_deploy.py \
+        --mlir ./ch_PP-OCRv3_rec.mlir \
+        --quantize F32 \
+        --chip bm1684x \
+        --test_input ./ch_PP-OCRv3_rec_in_f32.npz \
+        --test_reference ./ch_PP-OCRv3_rec_top_outputs.npz \
+        --model ./ch_PP-OCRv3_rec_1684x_f32.bmodel'
+    cls_str = 'model_deploy.py \
+        --mlir ./ch_PP-OCRv3_cls.mlir \
+        --quantize F32 \
+        --chip bm1684x \
+        --test_input ./ch_PP-OCRv3_cls_in_f32.npz \
+        --test_reference ./ch_PP-OCRv3_cls_top_outputs.npz \
+        --model ./ch_PP-OCRv3_cls_1684x_f32.bmodel'
+    os.chdir('./ch_PP-OCRv3/workspace/')
+    print(det_str)
+    run(det_str, shell=True)
+
+    print(rec_str)
+    run(rec_str, shell=True)
+
+    print(cls_str)
+    run(cls_str, shell=True)
+    os.chdir('../../')
+
 args = parse_arguments()
 
+if (args.auto):
+    getPPOCRv3()
+    fix_input_shape()
+    paddle2onnx()
+    mlir_prepare()
+    onnx2mlir()
+    mlir2bmodel()
+
 # 配置runtime，加载模型
 runtime_option = fd.RuntimeOption()
 runtime_option.use_sophgo()
 
 # Detection模型, 检测文字框
-det_model_file = args.det_model
+det_model_file = './ch_PP-OCRv3/workspace/ch_PP-OCRv3_det_1684x_f32.bmodel' if args.auto else args.det_model
 det_params_file = ""
 # Classification模型，方向分类，可选
-cls_model_file = args.cls_model
+cls_model_file = './ch_PP-OCRv3/workspace/ch_PP-OCRv3_cls_1684x_f32.bmodel' if args.auto else args.cls_model
 cls_params_file = ""
 # Recognition模型，文字识别模型
-rec_model_file = args.rec_model
+rec_model_file = './ch_PP-OCRv3/workspace/ch_PP-OCRv3_rec_1684x_f32.bmodel' if args.auto else args.rec_model
 rec_params_file = ""
-rec_label_file = args.rec_label_file
+rec_label_file = './ppocr_keys_v1.txt' if args.auto else args.rec_label_file
+image_file = './12.jpg' if args.auto else args.image
 
 # 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
@@ -92,9 +283,9 @@ rec_model = fd.vision.ocr.Recognizer(
 ppocr_v3 = fd.vision.ocr.PPOCRv3(
     det_model=det_model, cls_model=cls_model, rec_model=rec_model)
 
-# 需要使用下行代码, 来启用rec模型的静态shape推理，这里rec模型的静态输入为[3, 48, 584]
+# 需要使用下行代码, 来启用rec模型的静态shape推理，这里rec模型的静态输入为[3, 48, 320]
 rec_model.preprocessor.static_shape_infer = True
-rec_model.preprocessor.rec_image_shape = [3, 48, 584]
+rec_model.preprocessor.rec_image_shape = [3, 48, 320]
 
 # 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
@@ -103,7 +294,7 @@ ppocr_v3.cls_batch_size = cls_batch_size
 ppocr_v3.rec_batch_size = rec_batch_size
 
 # 预测图片准备
-im = cv2.imread(args.image)
+im = cv2.imread(image_file)
 
 #预测并打印结果
 result = ppocr_v3.predict(im)

examples/vision/segmentation/paddleseg/semantic_segmentation/sophgo/python/README.md

@@ -28,7 +28,11 @@ wget https://paddleseg.bj.bcebos.com/dygraph/demo/cityscapes_demo.png
 将Paddle模型转换为SOPHGO bmodel模型，转换步骤参考[文档](../README_CN.md)
 
 # 推理
-python3 infer.py --model_file ./bmodel/pp_liteseg_1684x_f32.bmodel --config_file ./bmodel/deploy.yaml --image cityscapes_demo.png
+# --auto True: 自动完成下载数据、模型转换、推理 
+python3 infer.py --auto True --model '' --config_file '' --image ''
+
+# --auto False，需要用户设置模型、配置文件、图片路径，进行推理
+python3 infer.py --auto False --model_file ./bmodel/pp_liteseg_1684x_f32.bmodel --config_file ./bmodel/deploy.yaml --image cityscapes_demo.png
 
 # 运行完成后返回结果如下所示
 运行结果保存在sophgo_img.png中

examples/vision/segmentation/paddleseg/semantic_segmentation/sophgo/python/infer.py

@@ -1,12 +1,14 @@
 import fastdeploy as fd
 import cv2
 import os
-
+from subprocess import run
 
 def parse_arguments():
     import argparse
     import ast
     parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--auto", required=True, help="Auto download, convert, compile and infer if True")
     parser.add_argument("--model", required=True, help="Path of model.")
     parser.add_argument(
         "--config_file", required=True, help="Path of config file.")
@@ -16,15 +18,102 @@ def parse_arguments():
     return parser.parse_args()
 
 
+def download():
+    download_model_str = 'wget https://bj.bcebos.com/paddlehub/fastdeploy/PP_LiteSeg_B_STDC2_cityscapes_without_argmax_infer.tgz'
+    if not os.path.exists('PP_LiteSeg_B_STDC2_cityscapes_without_argmax_infer.tgz'):
+        print(download_model_str)
+        run(download_model_str, shell=True)
+    tar_str = 'tar xvf PP_LiteSeg_B_STDC2_cityscapes_without_argmax_infer.tgz'
+    if not os.path.exists('./PP_LiteSeg_B_STDC2_cityscapes_without_argmax_infer'):
+        print(tar_str)
+        run(tar_str, shell=True)
+    download_script_str = 'wget https://raw.githubusercontent.com/PaddlePaddle/Paddle2ONNX/develop/tools/paddle/paddle_infer_shape.py'
+    if not os.path.exists('paddle_infer_shape.py'):
+        print(download_script_str)
+        run(download_script_str, shell=True)
+    download_img_str = 'wget https://paddleseg.bj.bcebos.com/dygraph/demo/cityscapes_demo.png'
+    if not os.path.exists('cityscapes_demo.png'):
+        print(download_img_str)
+        run(download_img_str, shell=True)
+
+def paddle2onnx():
+    paddle_infer_shape_str = 'python3 paddle_infer_shape.py --model_dir PP_LiteSeg_B_STDC2_cityscapes_without_argmax_infer \
+                             --model_filename model.pdmodel \
+                             --params_filename model.pdiparams \
+                             --save_dir pp_liteseg_fix \
+                             --input_shape_dict="{\'x\':[1,3,512,512]}"'
+    print(paddle_infer_shape_str)
+    run(paddle_infer_shape_str, shell=True)
+    pp2onnx_str = 'paddle2onnx --model_dir pp_liteseg_fix \
+                --model_filename model.pdmodel \
+                --params_filename model.pdiparams \
+                --save_file pp_liteseg.onnx \
+                --enable_dev_version True'
+    print(pp2onnx_str)
+    run(pp2onnx_str, shell=True)
+
+def mlir_prepare():
+    mlir_path = os.getenv("MODEL_ZOO_PATH")
+    mlir_path = mlir_path[:-13]
+    regression_path = os.path.join(mlir_path, 'regression')
+    mv_str_list = ['mkdir pp_liteseg', 
+                'cp -rf ' + os.path.join(regression_path, 'dataset/COCO2017/') + ' ./pp_liteseg', 
+                'cp -rf ' + os.path.join(regression_path, 'image/') + ' ./pp_liteseg',
+                'mv pp_liteseg.onnx ./pp_liteseg',
+                'mkdir ./pp_liteseg/workspace']
+    for str in mv_str_list:
+        print(str)
+        run(str, shell=True)
+
+def onnx2mlir():
+    transform_str = 'model_transform.py \
+        --model_name pp_liteseg \
+        --model_def ../pp_liteseg.onnx \
+        --input_shapes [[1,3,512,512]] \
+        --mean 0.0,0.0,0.0 \
+        --scale 0.0039216,0.0039216,0.0039216 \
+        --keep_aspect_ratio \
+        --pixel_format rgb \
+        --output_names bilinear_interp_v2_6.tmp_0 \
+        --test_input ../image/dog.jpg \
+        --test_result pp_liteseg_top_outputs.npz \
+        --mlir pp_liteseg.mlir'
+    print(transform_str)
+    os.chdir('./pp_liteseg/workspace')
+    run(transform_str, shell=True)
+    os.chdir('../../')
+
+def mlir2bmodel():
+    deploy_str = 'model_deploy.py \
+                --mlir pp_liteseg.mlir \
+                --quantize F32 \
+                --chip bm1684x \
+                --test_input pp_liteseg_in_f32.npz \
+                --test_reference pp_liteseg_top_outputs.npz \
+                --model pp_liteseg_1684x_f32.bmodel'
+    print(deploy_str)
+    os.chdir('./pp_liteseg/workspace')
+    run(deploy_str, shell=True)
+    os.chdir('../../')
+
 args = parse_arguments()
 
+
+if args.auto:
+    download()
+    paddle2onnx()
+    mlir_prepare()
+    onnx2mlir()
+    mlir2bmodel()
+
 # 配置runtime，加载模型
 runtime_option = fd.RuntimeOption()
 runtime_option.use_sophgo()
 
-model_file = args.model
+model_file = './pp_liteseg/workspace/pp_liteseg_1684x_f32.bmodel' if args.auto else args.model
 params_file = ""
-config_file = args.config_file
+config_file = './PP_LiteSeg_B_STDC2_cityscapes_without_argmax_infer/deploy.yaml' if args.auto else args.config_file
+img_file = './cityscapes_demo.png' if args.auto else args.image
 
 model = fd.vision.segmentation.PaddleSegModel(
     model_file,
@@ -34,7 +123,7 @@ model = fd.vision.segmentation.PaddleSegModel(
     model_format=fd.ModelFormat.SOPHGO)
 
 # 预测图片分类结果
-im_org = cv2.imread(args.image)
+im_org = cv2.imread(img_file)
 #bmodel 是静态模型，模型输入固定，这里设置为[512, 512]
 im = cv2.resize(im_org, [512, 512], interpolation=cv2.INTER_LINEAR)
 result = model.predict(im)
@@ -42,4 +131,5 @@ print(result)
 
 # 预测结果可视化
 vis_im = fd.vision.vis_segmentation(im, result, weight=0.5)
+vis_im = cv2.resize(vis_im, [im_org.shape[1], im_org.shape[0]], interpolation=cv2.INTER_LINEAR)
 cv2.imwrite("sophgo_img.png", vis_im)

fastdeploy/runtime/backends/sophgo/sophgo_backend.cc

@@ -225,6 +225,13 @@ bool SophgoBackend::Infer(std::vector<FDTensor>& inputs,
     free(temp_out);
   }
 
+  for (int i = 0; i < input_size; i++) {
+    bm_free_device(handle_, input_tensors[i].device_mem);
+  }
+  for (int i = 0; i < output_size; i++) {
+    bm_free_device(handle_, output_tensors[i].device_mem);
+  }
+
   return true;
 }