diff --git a/deploy/cpp_infer/src/ocr_cls.cpp b/deploy/cpp_infer/src/ocr_cls.cpp index 674630bf1e7e04841e027a7320d62af4a453ffc8..92d83600cea04419db231c0097caa53ed6fec58b 100644 --- a/deploy/cpp_infer/src/ocr_cls.cpp +++ b/deploy/cpp_infer/src/ocr_cls.cpp @@ -112,6 +112,11 @@ void Classifier::LoadModel(const std::string &model_dir) { precision = paddle_infer::Config::Precision::kInt8; } config.EnableTensorRtEngine(1 << 20, 10, 3, precision, false, false); + if (!Utility::PathExists("./trt_cls_shape.txt")){ + config.CollectShapeRangeInfo("./trt_cls_shape.txt"); + } else { + config.EnableTunedTensorRtDynamicShape("./trt_cls_shape.txt", true); + } } } else { config.DisableGpu(); diff --git a/deploy/cpp_infer/src/ocr_det.cpp b/deploy/cpp_infer/src/ocr_det.cpp index 56de195186a0d4d6c8b2482eb57c106347485928..0bfba4a2301a632696426f35b7be1dbacefe4cbf 100644 --- a/deploy/cpp_infer/src/ocr_det.cpp +++ b/deploy/cpp_infer/src/ocr_det.cpp @@ -32,49 +32,12 @@ void DBDetector::LoadModel(const std::string &model_dir) { if (this->precision_ == "int8") { precision = paddle_infer::Config::Precision::kInt8; } - config.EnableTensorRtEngine(1 << 20, 1, 20, precision, false, false); - std::map> min_input_shape = { - {"x", {1, 3, 50, 50}}, - {"conv2d_92.tmp_0", {1, 120, 20, 20}}, - {"conv2d_91.tmp_0", {1, 24, 10, 10}}, - {"conv2d_59.tmp_0", {1, 96, 20, 20}}, - {"nearest_interp_v2_1.tmp_0", {1, 256, 10, 10}}, - {"nearest_interp_v2_2.tmp_0", {1, 256, 20, 20}}, - {"conv2d_124.tmp_0", {1, 256, 20, 20}}, - {"nearest_interp_v2_3.tmp_0", {1, 64, 20, 20}}, - {"nearest_interp_v2_4.tmp_0", {1, 64, 20, 20}}, - {"nearest_interp_v2_5.tmp_0", {1, 64, 20, 20}}, - {"elementwise_add_7", {1, 56, 2, 2}}, - {"nearest_interp_v2_0.tmp_0", {1, 256, 2, 2}}}; - std::map> max_input_shape = { - {"x", {1, 3, 1536, 1536}}, - {"conv2d_92.tmp_0", {1, 120, 400, 400}}, - {"conv2d_91.tmp_0", {1, 24, 200, 200}}, - {"conv2d_59.tmp_0", {1, 96, 400, 400}}, - {"nearest_interp_v2_1.tmp_0", {1, 256, 200, 200}}, - {"nearest_interp_v2_2.tmp_0", {1, 256, 400, 400}}, - {"conv2d_124.tmp_0", {1, 256, 400, 400}}, - {"nearest_interp_v2_3.tmp_0", {1, 64, 400, 400}}, - {"nearest_interp_v2_4.tmp_0", {1, 64, 400, 400}}, - {"nearest_interp_v2_5.tmp_0", {1, 64, 400, 400}}, - {"elementwise_add_7", {1, 56, 400, 400}}, - {"nearest_interp_v2_0.tmp_0", {1, 256, 400, 400}}}; - std::map> opt_input_shape = { - {"x", {1, 3, 640, 640}}, - {"conv2d_92.tmp_0", {1, 120, 160, 160}}, - {"conv2d_91.tmp_0", {1, 24, 80, 80}}, - {"conv2d_59.tmp_0", {1, 96, 160, 160}}, - {"nearest_interp_v2_1.tmp_0", {1, 256, 80, 80}}, - {"nearest_interp_v2_2.tmp_0", {1, 256, 160, 160}}, - {"conv2d_124.tmp_0", {1, 256, 160, 160}}, - {"nearest_interp_v2_3.tmp_0", {1, 64, 160, 160}}, - {"nearest_interp_v2_4.tmp_0", {1, 64, 160, 160}}, - {"nearest_interp_v2_5.tmp_0", {1, 64, 160, 160}}, - {"elementwise_add_7", {1, 56, 40, 40}}, - {"nearest_interp_v2_0.tmp_0", {1, 256, 40, 40}}}; - - config.SetTRTDynamicShapeInfo(min_input_shape, max_input_shape, - opt_input_shape); + config.EnableTensorRtEngine(1 << 30, 1, 20, precision, false, false); + if (!Utility::PathExists("./trt_det_shape.txt")){ + config.CollectShapeRangeInfo("./trt_det_shape.txt"); + } else { + config.EnableTunedTensorRtDynamicShape("./trt_det_shape.txt", true); + } } } else { config.DisableGpu(); diff --git a/deploy/cpp_infer/src/ocr_rec.cpp b/deploy/cpp_infer/src/ocr_rec.cpp index 0f90ddfab4872f97829da081e64cb7437e72493a..90ad6598d325687bea1129d4db0a9ddb85409686 100644 --- a/deploy/cpp_infer/src/ocr_rec.cpp +++ b/deploy/cpp_infer/src/ocr_rec.cpp @@ -148,19 +148,12 @@ void CRNNRecognizer::LoadModel(const std::string &model_dir) { precision = paddle_infer::Config::Precision::kInt8; } config.EnableTensorRtEngine(1 << 20, 10, 15, precision, false, false); - int imgH = this->rec_image_shape_[1]; - int imgW = this->rec_image_shape_[2]; - std::map> min_input_shape = { - {"x", {1, 3, imgH, 10}}, {"lstm_0.tmp_0", {10, 1, 96}}}; - std::map> max_input_shape = { - {"x", {this->rec_batch_num_, 3, imgH, 2500}}, - {"lstm_0.tmp_0", {1000, 1, 96}}}; - std::map> opt_input_shape = { - {"x", {this->rec_batch_num_, 3, imgH, imgW}}, - {"lstm_0.tmp_0", {25, 1, 96}}}; - - config.SetTRTDynamicShapeInfo(min_input_shape, max_input_shape, - opt_input_shape); + if (!Utility::PathExists("./trt_rec_shape.txt")){ + config.CollectShapeRangeInfo("./trt_rec_shape.txt"); + } else { + config.EnableTunedTensorRtDynamicShape("./trt_rec_shape.txt", true); + } + } } else { config.DisableGpu(); diff --git a/deploy/lite/config.txt b/deploy/lite/config.txt index dda0d2b0320544d3a82f59b0672c086c64d83d3d..404249323b6cb5de345438056a9a10abd64b38bc 100644 --- a/deploy/lite/config.txt +++ b/deploy/lite/config.txt @@ -5,4 +5,4 @@ det_db_unclip_ratio 1.6 det_db_use_dilate 0 det_use_polygon_score 1 use_direction_classify 1 -rec_image_height 32 \ No newline at end of file +rec_image_height 48 \ No newline at end of file diff --git a/deploy/lite/readme.md b/deploy/lite/readme.md index a1bef8120e52dd91db0fda4ac2a4d91cc2800818..fc91cbfa7d69f6a8c1086243e4df3f820bd78339 100644 --- a/deploy/lite/readme.md +++ b/deploy/lite/readme.md @@ -99,6 +99,8 @@ The following table also provides a series of models that can be deployed on mob |Version|Introduction|Model size|Detection model|Text Direction model|Recognition model|Paddle-Lite branch| |---|---|---|---|---|---|---| +|PP-OCRv3|extra-lightweight chinese OCR optimized model|16.2M|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.nb)|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_ppocr_mobile_v2.0_cls_infer_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_infer.nb)|v2.10| +|PP-OCRv3(slim)|extra-lightweight chinese OCR optimized model|5.9M|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_infer.nb)|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_ppocr_mobile_v2.0_cls_slim_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_slim_infer.nb)|v2.10| |PP-OCRv2|extra-lightweight chinese OCR optimized model|11M|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_det_infer_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_ppocr_mobile_v2.0_cls_infer_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_rec_infer_opt.nb)|v2.10| |PP-OCRv2(slim)|extra-lightweight chinese OCR optimized model|4.6M|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_det_slim_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_ppocr_mobile_v2.0_cls_slim_opt.nb)|[download link](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_rec_slim_opt.nb)|v2.10| @@ -134,17 +136,16 @@ Introduction to paddle_lite_opt parameters: The following takes the ultra-lightweight Chinese model of PaddleOCR as an example to introduce the use of the compiled opt file to complete the conversion of the inference model to the Paddle-Lite optimized model ``` -# 【[Recommendation] Download the Chinese and English inference model of PP-OCRv2 -wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_slim_quant_infer.tar && tar xf ch_PP-OCRv2_det_slim_quant_infer.tar -wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_infer.tar && tar xf ch_PP-OCRv2_rec_slim_quant_infer.tar +# 【[Recommendation] Download the Chinese and English inference model of PP-OCRv3 +wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_infer.tar && tar xf ch_PP-OCRv3_det_slim_infer.tar +wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_slim_infer.tar && tar xf ch_PP-OCRv2_rec_slim_quant_infer.tar wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_cls_slim_infer.tar && tar xf ch_ppocr_mobile_v2.0_cls_slim_infer.tar # Convert detection model -./opt --model_file=./ch_PP-OCRv2_det_slim_quant_infer/inference.pdmodel --param_file=./ch_PP-OCRv2_det_slim_quant_infer/inference.pdiparams --optimize_out=./ch_PP-OCRv2_det_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer +paddle_lite_opt --model_file=./ch_PP-OCRv3_det_slim_infer/inference.pdmodel --param_file=./ch_PP-OCRv3_det_slim_infer/inference.pdiparams --optimize_out=./ch_PP-OCRv3_det_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer # Convert recognition model -./opt --model_file=./ch_PP-OCRv2_rec_slim_quant_infer/inference.pdmodel --param_file=./ch_PP-OCRv2_rec_slim_quant_infer/inference.pdiparams --optimize_out=./ch_PP-OCRv2_rec_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer +paddle_lite_opt --model_file=./ch_PP-OCRv3_rec_slim_infer/inference.pdmodel --param_file=./ch_PP-OCRv3_rec_slim_infer/inference.pdiparams --optimize_out=./ch_PP-OCRv3_rec_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer # Convert angle classifier model -./opt --model_file=./ch_ppocr_mobile_v2.0_cls_slim_infer/inference.pdmodel --param_file=./ch_ppocr_mobile_v2.0_cls_slim_infer/inference.pdiparams --optimize_out=./ch_ppocr_mobile_v2.0_cls_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer - +paddle_lite_opt --model_file=./ch_ppocr_mobile_v2.0_cls_slim_infer/inference.pdmodel --param_file=./ch_ppocr_mobile_v2.0_cls_slim_infer/inference.pdiparams --optimize_out=./ch_ppocr_mobile_v2.0_cls_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer ``` After the conversion is successful, there will be more files ending with `.nb` in the inference model directory, which is the successfully converted model file. @@ -197,15 +198,15 @@ Some preparatory work is required first. cp ../../../cxx/lib/libpaddle_light_api_shared.so ./debug/ ``` -Prepare the test image, taking PaddleOCR/doc/imgs/11.jpg as an example, copy the image file to the demo/cxx/ocr/debug/ folder. Prepare the model files optimized by the lite opt tool, ch_det_mv3_db_opt.nb, ch_rec_mv3_crnn_opt.nb, and place them under the demo/cxx/ocr/debug/ folder. +Prepare the test image, taking PaddleOCR/doc/imgs/11.jpg as an example, copy the image file to the demo/cxx/ocr/debug/ folder. Prepare the model files optimized by the lite opt tool, ch_PP-OCRv3_det_slim_opt.nb , ch_PP-OCRv3_rec_slim_opt.nb , and place them under the demo/cxx/ocr/debug/ folder. The structure of the OCR demo is as follows after the above command is executed: ``` demo/cxx/ocr/ |-- debug/ -| |--ch_PP-OCRv2_det_slim_opt.nb Detection model -| |--ch_PP-OCRv2_rec_slim_opt.nb Recognition model +| |--ch_PP-OCRv3_det_slim_opt.nb Detection model +| |--ch_PP-OCRv3_rec_slim_opt.nb Recognition model | |--ch_ppocr_mobile_v2.0_cls_slim_opt.nb Text direction classification model | |--11.jpg Image for OCR | |--ppocr_keys_v1.txt Dictionary file @@ -240,7 +241,7 @@ det_db_thresh 0.3 # Used to filter the binarized image of DB prediction, det_db_box_thresh 0.5 # DDB post-processing filter box threshold, if there is a missing box detected, it can be reduced as appropriate det_db_unclip_ratio 1.6 # Indicates the compactness of the text box, the smaller the value, the closer the text box to the text use_direction_classify 0 # Whether to use the direction classifier, 0 means not to use, 1 means to use -rec_image_height 32 # The height of the input image of the recognition model, the PP-OCRv3 model needs to be set to 48, and the PP-OCRv2 model needs to be set to 32 +rec_image_height 48 # The height of the input image of the recognition model, the PP-OCRv3 model needs to be set to 48, and the PP-OCRv2 model needs to be set to 32 ``` 5. Run Model on phone @@ -260,14 +261,14 @@ After the above steps are completed, you can use adb to push the file to the pho export LD_LIBRARY_PATH=${PWD}:$LD_LIBRARY_PATH # The use of ocr_db_crnn is: # ./ocr_db_crnn Mode Detection model file Orientation classifier model file Recognition model file Hardware Precision Threads Batchsize Test image path Dictionary file path - ./ocr_db_crnn system ch_PP-OCRv2_det_slim_opt.nb ch_PP-OCRv2_rec_slim_opt.nb ch_ppocr_mobile_v2.0_cls_slim_opt.nb arm8 INT8 10 1 ./11.jpg config.txt ppocr_keys_v1.txt True + ./ocr_db_crnn system ch_PP-OCRv3_det_slim_opt.nb ch_PP-OCRv3_rec_slim_opt.nb ch_ppocr_mobile_v2.0_cls_slim_opt.nb arm8 INT8 10 1 ./11.jpg config.txt ppocr_keys_v1.txt True # precision can be INT8 for quantitative model or FP32 for normal model. # Only using detection model -./ocr_db_crnn det ch_PP-OCRv2_det_slim_opt.nb arm8 INT8 10 1 ./11.jpg config.txt +./ocr_db_crnn det ch_PP-OCRv3_det_slim_opt.nb arm8 INT8 10 1 ./11.jpg config.txt # Only using recognition model -./ocr_db_crnn rec ch_PP-OCRv2_rec_slim_opt.nb arm8 INT8 10 1 word_1.jpg ppocr_keys_v1.txt config.txt +./ocr_db_crnn rec ch_PP-OCRv3_rec_slim_opt.nb arm8 INT8 10 1 word_1.jpg ppocr_keys_v1.txt config.txt ``` If you modify the code, you need to recompile and push to the phone. diff --git a/deploy/lite/readme_ch.md b/deploy/lite/readme_ch.md index 0793827fe647c470944fc36e2b243c8f7e704e99..78e2510917e0fd85c4a724ec74eccb0b7cfc6118 100644 --- a/deploy/lite/readme_ch.md +++ b/deploy/lite/readme_ch.md @@ -97,6 +97,8 @@ Paddle-Lite 提供了多种策略来自动优化原始的模型，其中包括 |模型版本|模型简介|模型大小|检测模型|文本方向分类模型|识别模型|Paddle-Lite版本| |---|---|---|---|---|---|---| +|PP-OCRv3|蒸馏版超轻量中文OCR移动端模型|16.2M|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_infer.nb)|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_ppocr_mobile_v2.0_cls_infer_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_infer.nb)|v2.10| +|PP-OCRv3(slim)|蒸馏版超轻量中文OCR移动端模型|5.9M|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_infer.nb)|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_ppocr_mobile_v2.0_cls_slim_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_slim_infer.nb)|v2.10| |PP-OCRv2|蒸馏版超轻量中文OCR移动端模型|11M|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_det_infer_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_ppocr_mobile_v2.0_cls_infer_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_rec_infer_opt.nb)|v2.10| |PP-OCRv2(slim)|蒸馏版超轻量中文OCR移动端模型|4.6M|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_det_slim_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_ppocr_mobile_v2.0_cls_slim_opt.nb)|[下载地址](https://paddleocr.bj.bcebos.com/PP-OCRv2/lite/ch_PP-OCRv2_rec_slim_opt.nb)|v2.10| @@ -131,16 +133,16 @@ paddle_lite_opt 参数介绍： 下面以PaddleOCR的超轻量中文模型为例，介绍使用编译好的opt文件完成inference模型到Paddle-Lite优化模型的转换。 ``` -# 【推荐】 下载 PP-OCRv2版本的中英文 inference模型 -wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_slim_quant_infer.tar && tar xf ch_PP-OCRv2_det_slim_quant_infer.tar -wget https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_infer.tar && tar xf ch_PP-OCRv2_rec_slim_quant_infer.tar +# 【推荐】 下载 PP-OCRv3版本的中英文 inference模型 +wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_slim_infer.tar && tar xf ch_PP-OCRv3_det_slim_infer.tar +wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_rec_slim_infer.tar && tar xf ch_PP-OCRv2_rec_slim_quant_infer.tar wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_cls_slim_infer.tar && tar xf ch_ppocr_mobile_v2.0_cls_slim_infer.tar # 转换检测模型 -./opt --model_file=./ch_PP-OCRv2_det_slim_quant_infer/inference.pdmodel --param_file=./ch_PP-OCRv2_det_slim_quant_infer/inference.pdiparams --optimize_out=./ch_PP-OCRv2_det_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer +paddle_lite_opt --model_file=./ch_PP-OCRv3_det_slim_infer/inference.pdmodel --param_file=./ch_PP-OCRv3_det_slim_infer/inference.pdiparams --optimize_out=./ch_PP-OCRv3_det_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer # 转换识别模型 -./opt --model_file=./ch_PP-OCRv2_rec_slim_quant_infer/inference.pdmodel --param_file=./ch_PP-OCRv2_rec_slim_quant_infer/inference.pdiparams --optimize_out=./ch_PP-OCRv2_rec_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer +paddle_lite_opt --model_file=./ch_PP-OCRv3_rec_slim_infer/inference.pdmodel --param_file=./ch_PP-OCRv3_rec_slim_infer/inference.pdiparams --optimize_out=./ch_PP-OCRv3_rec_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer # 转换方向分类器模型 -./opt --model_file=./ch_ppocr_mobile_v2.0_cls_slim_infer/inference.pdmodel --param_file=./ch_ppocr_mobile_v2.0_cls_slim_infer/inference.pdiparams --optimize_out=./ch_ppocr_mobile_v2.0_cls_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer +paddle_lite_opt --model_file=./ch_ppocr_mobile_v2.0_cls_slim_infer/inference.pdmodel --param_file=./ch_ppocr_mobile_v2.0_cls_slim_infer/inference.pdiparams --optimize_out=./ch_ppocr_mobile_v2.0_cls_slim_opt --valid_targets=arm --optimize_out_type=naive_buffer ``` @@ -194,15 +196,15 @@ wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_cls ``` 准备测试图像，以`PaddleOCR/doc/imgs/11.jpg`为例，将测试的图像复制到`demo/cxx/ocr/debug/`文件夹下。 - 准备lite opt工具优化后的模型文件，比如使用`ch_PP-OCRv2_det_slim_opt.ch_PP-OCRv2_rec_slim_rec.nb, ch_ppocr_mobile_v2.0_cls_slim_opt.nb`，模型文件放置在`demo/cxx/ocr/debug/`文件夹下。 + 准备lite opt工具优化后的模型文件，比如使用`ch_PP-OCRv3_det_slim_opt.ch_PP-OCRv3_rec_slim_rec.nb, ch_ppocr_mobile_v2.0_cls_slim_opt.nb`，模型文件放置在`demo/cxx/ocr/debug/`文件夹下。 执行完成后，ocr文件夹下将有如下文件格式： ``` demo/cxx/ocr/ |-- debug/ -| |--ch_PP-OCRv2_det_slim_opt.nb 优化后的检测模型文件 -| |--ch_PP-OCRv2_rec_slim_opt.nb 优化后的识别模型文件 +| |--ch_PP-OCRv3_det_slim_opt.nb 优化后的检测模型文件 +| |--ch_PP-OCRv3_rec_slim_opt.nb 优化后的识别模型文件 | |--ch_ppocr_mobile_v2.0_cls_slim_opt.nb 优化后的文字方向分类器模型文件 | |--11.jpg 待测试图像 | |--ppocr_keys_v1.txt 中文字典文件 @@ -239,7 +241,7 @@ det_db_thresh 0.3 # 用于过滤DB预测的二值化图像，设置为0. det_db_box_thresh 0.5 # 检测器后处理过滤box的阈值，如果检测存在漏框情况，可酌情减小 det_db_unclip_ratio 1.6 # 表示文本框的紧致程度，越小则文本框更靠近文本 use_direction_classify 0 # 是否使用方向分类器，0表示不使用，1表示使用 -rec_image_height 32 # 识别模型输入图像的高度，PP-OCRv3模型设置为48，PP-OCRv2模型需要设置为32 +rec_image_height 48 # 识别模型输入图像的高度，PP-OCRv3模型设置为48，PP-OCRv2模型需要设置为32 ``` 5. 启动调试 @@ -259,13 +261,13 @@ rec_image_height 32 # 识别模型输入图像的高度，PP-OCRv3模型 export LD_LIBRARY_PATH=${PWD}:$LD_LIBRARY_PATH # 开始使用，ocr_db_crnn可执行文件的使用方式为: # ./ocr_db_crnn 预测模式 检测模型文件 方向分类器模型文件 识别模型文件 运行硬件 运行精度 线程数 batchsize 测试图像路径 参数配置路径 字典文件路径 是否使用benchmark参数 - ./ocr_db_crnn system ch_PP-OCRv2_det_slim_opt.nb ch_PP-OCRv2_rec_slim_opt.nb ch_ppocr_mobile_v2.0_cls_slim_opt.nb arm8 INT8 10 1 ./11.jpg config.txt ppocr_keys_v1.txt True + ./ocr_db_crnn system ch_PP-OCRv3_det_slim_opt.nb ch_PP-OCRv3_rec_slim_opt.nb ch_ppocr_mobile_v2.0_cls_slim_opt.nb arm8 INT8 10 1 ./11.jpg config.txt ppocr_keys_v1.txt True # 仅使用文本检测模型，使用方式如下： -./ocr_db_crnn det ch_PP-OCRv2_det_slim_opt.nb arm8 INT8 10 1 ./11.jpg config.txt +./ocr_db_crnn det ch_PP-OCRv3_det_slim_opt.nb arm8 INT8 10 1 ./11.jpg config.txt # 仅使用文本识别模型，使用方式如下： -./ocr_db_crnn rec ch_PP-OCRv2_rec_slim_opt.nb arm8 INT8 10 1 word_1.jpg ppocr_keys_v1.txt config.txt +./ocr_db_crnn rec ch_PP-OCRv3_rec_slim_opt.nb arm8 INT8 10 1 word_1.jpg ppocr_keys_v1.txt config.txt ``` 如果对代码做了修改，则需要重新编译并push到手机上。 diff --git a/deploy/slim/quantization/README.md b/deploy/slim/quantization/README.md index 4c1d784b99aade614d78b4bd6fb20afef15f0f6f..8b29693c9803f004f123b5497c9224ae5c31041d 100644 --- a/deploy/slim/quantization/README.md +++ b/deploy/slim/quantization/README.md @@ -22,7 +22,7 @@ ### 1. 安装PaddleSlim ```bash -pip3 install paddleslim==2.2.2 +pip3 install paddleslim==2.3.2 ``` ### 2. 准备训练好的模型 @@ -32,18 +32,7 @@ PaddleOCR提供了一系列训练好的[模型](../../../doc/doc_ch/models_list. ### 3. 量化训练 量化训练包括离线量化训练和在线量化训练，在线量化训练效果更好，需加载预训练模型，在定义好量化策略后即可对模型进行量化。 - -量化训练的代码位于slim/quantization/quant.py 中，比如训练检测模型，训练指令如下： -```bash -python deploy/slim/quantization/quant.py -c configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml -o Global.pretrained_model='your trained model' Global.save_model_dir=./output/quant_model - -# 比如下载提供的训练模型 -wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_det_train.tar -tar -xf ch_ppocr_mobile_v2.0_det_train.tar -python deploy/slim/quantization/quant.py -c configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml -o Global.pretrained_model=./ch_ppocr_mobile_v2.0_det_train/best_accuracy Global.save_model_dir=./output/quant_model -``` - -模型蒸馏和模型量化可以同时使用，以PPOCRv3检测模型为例： +量化训练的代码位于slim/quantization/quant.py 中，比如训练检测模型，以PPOCRv3检测模型为例，训练指令如下： ``` # 下载检测预训练模型： wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_distill_train.tar @@ -58,7 +47,7 @@ python deploy/slim/quantization/quant.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_ 在得到量化训练保存的模型后，我们可以将其导出为inference_model，用于预测部署： ```bash -python deploy/slim/quantization/export_model.py -c configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml -o Global.checkpoints=output/quant_model/best_accuracy Global.save_inference_dir=./output/quant_inference_model +python deploy/slim/quantization/export_model.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml -o Global.checkpoints=output/quant_model/best_accuracy Global.save_inference_dir=./output/quant_inference_model ``` ### 5. 量化模型部署 diff --git a/deploy/slim/quantization/README_en.md b/deploy/slim/quantization/README_en.md index c6796ae9dc256496308e432023c45ef1026c3d92..f82c3d844e292ee76b95624f7632ed40301e5a4c 100644 --- a/deploy/slim/quantization/README_en.md +++ b/deploy/slim/quantization/README_en.md @@ -25,7 +25,7 @@ After training, if you want to further compress the model size and accelerate th ### 1. Install PaddleSlim ```bash -pip3 install paddleslim==2.2.2 +pip3 install paddleslim==2.3.2 ``` @@ -39,18 +39,7 @@ Quantization training includes offline quantization training and online quantiza Online quantization training is more effective. It is necessary to load the pre-trained model. After the quantization strategy is defined, the model can be quantified. -The code for quantization training is located in `slim/quantization/quant.py`. For example, to train a detection model, the training instructions are as follows: -```bash -python deploy/slim/quantization/quant.py -c configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml -o Global.pretrained_model='your trained model' Global.save_model_dir=./output/quant_model - -# download provided model -wget https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_det_train.tar -tar -xf ch_ppocr_mobile_v2.0_det_train.tar -python deploy/slim/quantization/quant.py -c configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml -o Global.pretrained_model=./ch_ppocr_mobile_v2.0_det_train/best_accuracy Global.save_model_dir=./output/quant_model -``` - - -Model distillation and model quantization can be used at the same time, taking the PPOCRv3 detection model as an example: +The code for quantization training is located in `slim/quantization/quant.py`. For example, the training instructions of slim PPOCRv3 detection model are as follows: ``` # download provided model wget https://paddleocr.bj.bcebos.com/PP-OCRv3/chinese/ch_PP-OCRv3_det_distill_train.tar @@ -66,7 +55,7 @@ If you want to quantify the text recognition model, you can modify the configura Once we got the model after pruning and fine-tuning, we can export it as an inference model for the deployment of predictive tasks: ```bash -python deploy/slim/quantization/export_model.py -c configs/det/ch_ppocr_v2.0/ch_det_mv3_db_v2.0.yml -o Global.checkpoints=output/quant_model/best_accuracy Global.save_inference_dir=./output/quant_inference_model +python deploy/slim/quantization/export_model.py -c configs/det/ch_PP-OCRv3/ch_PP-OCRv3_det_cml.yml -o Global.checkpoints=output/quant_model/best_accuracy Global.save_inference_dir=./output/quant_inference_model ``` ### 5. Deploy diff --git a/tools/infer/utility.py b/tools/infer/utility.py index 9baf66d7f469a3bf6c9a140e034aee3a635a5c8e..b9793123163185d823cd0bd3fe53b99719f66669 100644 --- a/tools/infer/utility.py +++ b/tools/infer/utility.py @@ -226,19 +226,18 @@ def create_predictor(args, mode, logger): use_calib_mode=False) # collect shape - if args.shape_info_filename is not None: - if not os.path.exists(args.shape_info_filename): - config.collect_shape_range_info( - args.shape_info_filename) + trt_shape_f = f"{os.path.dirname(args.shape_info_filename)}/{mode}_{os.path.basename(args.shape_info_filename)}" + if trt_shape_f is not None: + if not os.path.exists(trt_shape_f): + config.collect_shape_range_info(trt_shape_f) logger.info( - f"collect dynamic shape info into : {args.shape_info_filename}" + f"collect dynamic shape info into : {trt_shape_f}" ) else: logger.info( - f"dynamic shape info file( {args.shape_info_filename} ) already exists, not need to generate again." + f"dynamic shape info file( {trt_shape_f} ) already exists, not need to generate again." ) - config.enable_tuned_tensorrt_dynamic_shape( - args.shape_info_filename, True) + config.enable_tuned_tensorrt_dynamic_shape(trt_shape_f, True) else: logger.info( f"when using tensorrt, dynamic shape is a suggested option, you can use '--shape_info_filename=shape.txt' for offline dygnamic shape tuning"