From c0ebad1a22fc4ea2e281c2f33a9dab409a3bca44 Mon Sep 17 00:00:00 2001 From: Guanghua Yu <742925032@qq.com> Date: Fri, 21 May 2021 16:40:22 +0800 Subject: [PATCH] fix image_shape in export_model (#3093) * fix image_shape in export_model --- configs/slim/README.md | 4 ++-- deploy/TENSOR_RT.md | 5 +++-- deploy/cpp/include/config_parser.h | 8 -------- deploy/cpp/include/object_detector.h | 4 +--- deploy/cpp/include/preprocess_op.h | 19 ++++++++----------- deploy/python/infer.py | 10 ++++------ deploy/python/keypoint_infer.py | 4 +--- deploy/python/preprocess.py | 17 ++--------------- ppdet/engine/export_utils.py | 15 +++++---------- 9 files changed, 26 insertions(+), 60 deletions(-) diff --git a/configs/slim/README.md b/configs/slim/README.md index b3b098153..b175266f8 100755 --- a/configs/slim/README.md +++ b/configs/slim/README.md @@ -20,7 +20,7 @@ **PaddleDetection、 PaddlePaddle与PaddleSlim 版本关系:** | PaddleDetection版本 | PaddlePaddle版本 | PaddleSlim版本 | 备注 | | :------------------: | :---------------: | :-------: |:---------------: | -| release/2.1 | >= 2.1.0 | 2.1 | -- | +| release/2.1 | >= 2.1.0 | 2.1 | 量化模型导出依赖最新Paddle develop分支，可在[PaddlePaddle每日版本](https://www.paddlepaddle.org.cn/documentation/docs/zh/install/Tables.html#whl-dev)中下载安装 | | release/2.0 | >= 2.0.1 | 2.0 | 量化依赖Paddle 2.1及PaddleSlim 2.1 | @@ -107,7 +107,7 @@ python tools/export_model.py -c configs/{MODEL.yml} --slim_config configs/slim/{ #### COCO上benchmark | 模型 | 压缩策略 | GFLOPs | 模型体积(MB) | 输入尺寸 | 预测时延(SD855) | Box AP | 下载 | 模型配置文件 | 压缩算法配置文件 | | :---------: | :-------: | :------------: |:-------------: | :------: | :-------------: | :------: | :-----------------------------------------------------: |:-------------: | :------: | -| PP-YOLO-MobileNetV3_large | baseline | -- | 18.5 | 608 | 25.1ms | 24.3 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | - | +| PP-YOLO-MobileNetV3_large | baseline | -- | 18.5 | 608 | 25.1ms | 23.2 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | - | | PP-YOLO-MobileNetV3_large | 剪裁-FPGM | -37% | 12.6 | 608 | - | 22.3 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolo_mbv3_large_prune_fpgm.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/prune/ppyolo_mbv3_large_prune_fpgm.yml) | | YOLOv3-DarkNet53 | baseline | -- | 238.2 | 608 | - | 39.0 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) | - | | YOLOv3-DarkNet53 | 剪裁-FPGM | -24% | - | 608 | - | 37.6 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_darknet_prune_fpgm.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/prune/yolov3_darknet_prune_fpgm.yml) | diff --git a/deploy/TENSOR_RT.md b/deploy/TENSOR_RT.md index 986933576..225f252ba 100644 --- a/deploy/TENSOR_RT.md +++ b/deploy/TENSOR_RT.md @@ -8,7 +8,9 @@ TensorRT是NVIDIA提出的用于统一模型部署的加速库，可以应用于 - 如果Python和CPP官网没有提供已编译好的安装包或预测库，请参考[源码安装](https://www.paddlepaddle.org.cn/documentation/docs/zh/install/compile/linux-compile.html) 自行编译 -注意，您的机器上TensorRT的版本需要跟您使用的预测库中TensorRT版本保持一致。 +**注意：** +- 您的机器上TensorRT的版本需要跟您使用的预测库中TensorRT版本保持一致。 +- PaddleDetection中部署预测要求TensorRT版本 > 6.0。 ## 2. 导出模型 模型导出具体请参考文档[PaddleDetection模型导出教程](../EXPORT_MODEL.md)。 @@ -31,7 +33,6 @@ config->EnableTensorRtEngine(1 << 20 /*workspace_size*/, ``` ### 3.2 TensorRT固定尺寸预测 -TensorRT版本<=5时，使用TensorRT预测时，只支持固定尺寸输入。 在导出模型时指定模型输入尺寸，设置`TestReader.inputs_def.image_shape=[3,640,640]`，具体请参考[PaddleDetection模型导出教程](../EXPORT_MODEL.md) 。 diff --git a/deploy/cpp/include/config_parser.h b/deploy/cpp/include/config_parser.h index c38049d31..3392b3ed7 100644 --- a/deploy/cpp/include/config_parser.h +++ b/deploy/cpp/include/config_parser.h @@ -91,13 +91,6 @@ class ConfigPaser { return false; } - if (config["image_shape"].IsDefined()) { - image_shape_ = config["image_shape"].as>(); - } else { - std::cerr << "Please set image_shape." << std::endl; - return false; - } - return true; } std::string mode_; @@ -106,7 +99,6 @@ class ConfigPaser { int min_subgraph_size_; YAML::Node preprocess_info_; std::vector label_list_; - std::vector image_shape_; }; } // namespace PaddleDetection diff --git a/deploy/cpp/include/object_detector.h b/deploy/cpp/include/object_detector.h index aeffca0b3..572224a3b 100644 --- a/deploy/cpp/include/object_detector.h +++ b/deploy/cpp/include/object_detector.h @@ -82,8 +82,7 @@ class ObjectDetector { config_.load_config(model_dir); this->min_subgraph_size_ = config_.min_subgraph_size_; threshold_ = config_.draw_threshold_; - image_shape_ = config_.image_shape_; - preprocessor_.Init(config_.preprocess_info_, image_shape_); + preprocessor_.Init(config_.preprocess_info_); LoadModel(model_dir, batch_size, run_mode); } @@ -134,7 +133,6 @@ class ObjectDetector { std::vector out_bbox_num_data_; float threshold_; ConfigPaser config_; - std::vector image_shape_; }; } // namespace PaddleDetection diff --git a/deploy/cpp/include/preprocess_op.h b/deploy/cpp/include/preprocess_op.h index 26a91cc9e..aad528d67 100644 --- a/deploy/cpp/include/preprocess_op.h +++ b/deploy/cpp/include/preprocess_op.h @@ -48,19 +48,19 @@ class ImageBlob { // Abstraction of preprocessing opration class class PreprocessOp { public: - virtual void Init(const YAML::Node& item, const std::vector image_shape) = 0; + virtual void Init(const YAML::Node& item) = 0; virtual void Run(cv::Mat* im, ImageBlob* data) = 0; }; class InitInfo : public PreprocessOp{ public: - virtual void Init(const YAML::Node& item, const std::vector image_shape) {} + virtual void Init(const YAML::Node& item) {} virtual void Run(cv::Mat* im, ImageBlob* data); }; class NormalizeImage : public PreprocessOp { public: - virtual void Init(const YAML::Node& item, const std::vector image_shape) { + virtual void Init(const YAML::Node& item) { mean_ = item["mean"].as>(); scale_ = item["std"].as>(); is_scale_ = item["is_scale"].as(); @@ -77,21 +77,18 @@ class NormalizeImage : public PreprocessOp { class Permute : public PreprocessOp { public: - virtual void Init(const YAML::Node& item, const std::vector image_shape) {} + virtual void Init(const YAML::Node& item) {} virtual void Run(cv::Mat* im, ImageBlob* data); }; class Resize : public PreprocessOp { public: - virtual void Init(const YAML::Node& item, const std::vector image_shape) { + virtual void Init(const YAML::Node& item) { interp_ = item["interp"].as(); //max_size_ = item["target_size"].as(); keep_ratio_ = item["keep_ratio"].as(); target_size_ = item["target_size"].as>(); - if (item["keep_ratio"]) { - in_net_shape_ = image_shape; - } } // Compute best resize scale for x-dimension, y-dimension @@ -109,7 +106,7 @@ class Resize : public PreprocessOp { // Models with FPN need input shape % stride == 0 class PadStride : public PreprocessOp { public: - virtual void Init(const YAML::Node& item, const std::vector image_shape) { + virtual void Init(const YAML::Node& item) { stride_ = item["stride"].as(); } @@ -121,14 +118,14 @@ class PadStride : public PreprocessOp { class Preprocessor { public: - void Init(const YAML::Node& config_node, const std::vector image_shape) { + void Init(const YAML::Node& config_node) { // initialize image info at first ops_["InitInfo"] = std::make_shared(); for (const auto& item : config_node) { auto op_name = item["type"].as(); ops_[op_name] = CreateOp(op_name); - ops_[op_name]->Init(item, image_shape); + ops_[op_name]->Init(item); } } diff --git a/deploy/python/infer.py b/deploy/python/infer.py index c92ab89c6..6ea92ed3d 100644 --- a/deploy/python/infer.py +++ b/deploy/python/infer.py @@ -99,8 +99,7 @@ class Detector(object): input_im_lst = [] input_im_info_lst = [] for im_path in image_list: - im, im_info = preprocess(im_path, preprocess_ops, - self.pred_config.input_shape) + im, im_info = preprocess(im_path, preprocess_ops) input_im_lst.append(im) input_im_info_lst.append(im_info) inputs = create_inputs(input_im_lst, input_im_info_lst) @@ -141,12 +140,12 @@ class Detector(object): ''' self.det_times.preprocess_time_s.start() inputs = self.preprocess(image_list) + self.det_times.preprocess_time_s.end() np_boxes, np_masks = None, None input_names = self.predictor.get_input_names() for i in range(len(input_names)): input_tensor = self.predictor.get_input_handle(input_names[i]) input_tensor.copy_from_cpu(inputs[input_names[i]]) - self.det_times.preprocess_time_s.end() for i in range(warmup): self.predictor.run() output_names = self.predictor.get_output_names() @@ -236,14 +235,14 @@ class DetectorSOLOv2(Detector): 'cate_label': label of segm, shape:[N] 'cate_score': confidence score of segm, shape:[N] ''' - self.det_times.postprocess_time_s.start() + self.det_times.preprocess_time_s.start() inputs = self.preprocess(image) + self.det_times.preprocess_time_s.end() np_label, np_score, np_segms = None, None, None input_names = self.predictor.get_input_names() for i in range(len(input_names)): input_tensor = self.predictor.get_input_handle(input_names[i]) input_tensor.copy_from_cpu(inputs[input_names[i]]) - self.det_times.postprocess_time_s.end() for i in range(warmup): self.predictor.run() output_names = self.predictor.get_output_names() @@ -331,7 +330,6 @@ class PredictConfig(): self.mask = False if 'mask' in yml_conf: self.mask = yml_conf['mask'] - self.input_shape = yml_conf['image_shape'] self.print_config() def check_model(self, yml_conf): diff --git a/deploy/python/keypoint_infer.py b/deploy/python/keypoint_infer.py index 49d8b3441..0b7ece269 100644 --- a/deploy/python/keypoint_infer.py +++ b/deploy/python/keypoint_infer.py @@ -88,8 +88,7 @@ class KeyPoint_Detector(object): new_op_info = op_info.copy() op_type = new_op_info.pop('type') preprocess_ops.append(eval(op_type)(**new_op_info)) - im, im_info = preprocess(im, preprocess_ops, - self.pred_config.input_shape) + im, im_info = preprocess(im, preprocess_ops) inputs = create_inputs(im, im_info) return inputs @@ -213,7 +212,6 @@ class PredictConfig_KeyPoint(): self.tagmap = False if 'keypoint_bottomup' == self.archcls: self.tagmap = True - self.input_shape = yml_conf['image_shape'] self.print_config() def check_model(self, yml_conf): diff --git a/deploy/python/preprocess.py b/deploy/python/preprocess.py index 700926ea8..5e44596f1 100644 --- a/deploy/python/preprocess.py +++ b/deploy/python/preprocess.py @@ -47,11 +47,7 @@ class Resize(object): interp (int): method of resize """ - def __init__( - self, - target_size, - keep_ratio=True, - interp=cv2.INTER_LINEAR, ): + def __init__(self, target_size, keep_ratio=True, interp=cv2.INTER_LINEAR): if isinstance(target_size, int): target_size = [target_size, target_size] self.target_size = target_size @@ -81,14 +77,6 @@ class Resize(object): im_info['im_shape'] = np.array(im.shape[:2]).astype('float32') im_info['scale_factor'] = np.array( [im_scale_y, im_scale_x]).astype('float32') - # padding im when image_shape fixed by infer_cfg.yml - if self.keep_ratio and im_info['input_shape'][1] != -1: - max_size = im_info['input_shape'][1] - padding_im = np.zeros( - (max_size, max_size, im_channel), dtype=np.float32) - im_h, im_w = im.shape[:2] - padding_im[:im_h, :im_w, :] = im - im = padding_im return im, im_info def generate_scale(self, im): @@ -205,13 +193,12 @@ class PadStride(object): return padding_im, im_info -def preprocess(im, preprocess_ops, input_shape): +def preprocess(im, preprocess_ops): # process image by preprocess_ops im_info = { 'scale_factor': np.array( [1., 1.], dtype=np.float32), 'im_shape': None, - 'input_shape': input_shape, } im, im_info = decode_image(im, im_info) for operator in preprocess_ops: diff --git a/ppdet/engine/export_utils.py b/ppdet/engine/export_utils.py index 69ac1bfc1..0bff32cdb 100644 --- a/ppdet/engine/export_utils.py +++ b/ppdet/engine/export_utils.py @@ -58,9 +58,7 @@ def _parse_reader(reader_cfg, dataset_cfg, metric, arch, image_shape): for key, value in st.items(): p = {'type': key} if key == 'Resize': - if value.get('keep_ratio', False) and int(image_shape[1]) != -1: - max_size = max(image_shape[1:]) - image_shape = [3, max_size, max_size] + if int(image_shape[1]) != -1: value['target_size'] = image_shape[1:] p.update(value) preprocess_list.append(p) @@ -76,7 +74,7 @@ def _parse_reader(reader_cfg, dataset_cfg, metric, arch, image_shape): }) break - return preprocess_list, label_list, image_shape + return preprocess_list, label_list def _dump_infer_config(config, path, image_shape, model): @@ -87,7 +85,6 @@ def _dump_infer_config(config, path, image_shape, model): 'mode': 'fluid', 'draw_threshold': 0.5, 'metric': config['metric'], - 'image_shape': image_shape }) infer_arch = config['architecture'] @@ -107,10 +104,9 @@ def _dump_infer_config(config, path, image_shape, model): label_arch = 'detection_arch' if infer_arch in KEYPOINT_ARCH: label_arch = 'keypoint_arch' - infer_cfg['Preprocess'], infer_cfg[ - 'label_list'], image_shape = _parse_reader( - config['TestReader'], config['TestDataset'], config['metric'], - label_arch, image_shape) + infer_cfg['Preprocess'], infer_cfg['label_list'] = _parse_reader( + config['TestReader'], config['TestDataset'], config['metric'], + label_arch, image_shape) if infer_arch == 'S2ANet': # TODO: move background to num_classes @@ -119,4 +115,3 @@ def _dump_infer_config(config, path, image_shape, model): yaml.dump(infer_cfg, open(path, 'w')) logger.info("Export inference config file to {}".format(os.path.join(path))) - return image_shape -- GitLab