Merge branch 'dygraph' into dygraph

configs/rec/ch_ppocr_v2.1/rec_chinese_lite_train_distillation_v2.1.yml

@@ -52,9 +52,10 @@ Architecture:
       Neck:
         name: SequenceEncoder
         encoder_type: rnn
-        hidden_size: 48
+        hidden_size: 64
       Head:
         name: CTCHead
+        mid_channels: 96
         fc_decay: 0.00001
     Teacher:
       pretrained:
@@ -71,9 +72,10 @@ Architecture:
       Neck:
         name: SequenceEncoder
         encoder_type: rnn
-        hidden_size: 48
+        hidden_size: 64
       Head:
         name: CTCHead
+        mid_channels: 96
         fc_decay: 0.00001
   
 

ppocr/modeling/heads/rec_ctc_head.py

@@ -33,8 +33,14 @@ def get_para_bias_attr(l2_decay, k):
 
 
 class CTCHead(nn.Layer):
-    def __init__(self, in_channels, out_channels, fc_decay=0.0004, **kwargs):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 fc_decay=0.0004,
+                 mid_channels=None,
+                 **kwargs):
         super(CTCHead, self).__init__()
+        if mid_channels is None:
             weight_attr, bias_attr = get_para_bias_attr(
                 l2_decay=fc_decay, k=in_channels)
             self.fc = nn.Linear(
@@ -42,10 +48,32 @@ class CTCHead(nn.Layer):
                 out_channels,
                 weight_attr=weight_attr,
                 bias_attr=bias_attr)
+        else:
+            weight_attr1, bias_attr1 = get_para_bias_attr(
+                l2_decay=fc_decay, k=in_channels)
+            self.fc1 = nn.Linear(
+                in_channels,
+                mid_channels,
+                weight_attr=weight_attr1,
+                bias_attr=bias_attr1)
+
+            weight_attr2, bias_attr2 = get_para_bias_attr(
+                l2_decay=fc_decay, k=mid_channels)
+            self.fc2 = nn.Linear(
+                mid_channels,
+                out_channels,
+                weight_attr=weight_attr2,
+                bias_attr=bias_attr2)
         self.out_channels = out_channels
+        self.mid_channels = mid_channels
 
     def forward(self, x, targets=None):
+        if self.mid_channels is None:
             predicts = self.fc(x)
+        else:
+            predicts = self.fc1(x)
+            predicts = self.fc2(predicts)
+            
         if not self.training:
             predicts = F.softmax(predicts, axis=2)
         return predicts

test/api.md → test1/api.md

 # PaddleStructure
 
+install layoutparser
+```sh
+wget  https://paddleocr.bj.bcebos.com/whl/layoutparser-0.0.0-py3-none-any.whl
+pip3 install layoutparser-0.0.0-py3-none-any.whl
+```
+
 ## 1. Introduction to pipeline
 
 PaddleStructure is a toolkit for complex layout text OCR, the process is as follows

test/api_ch.md → test1/api_ch.md

 # PaddleStructure
 
+安装layoutparser
+```sh
+wget  https://paddleocr.bj.bcebos.com/whl/layoutparser-0.0.0-py3-none-any.whl
+pip3 install layoutparser-0.0.0-py3-none-any.whl
+```
+
 ## 1. pipeline介绍
 
 PaddleStructure 是一个用于复杂板式文字OCR的工具包，流程如下
@@ -18,6 +24,7 @@ PaddleStructure 是一个用于复杂板式文字OCR的工具包，流程如下
 
 ## 2. LayoutParser
 
+[文档](layout/README.md)
 
 ## 3. Table OCR
 

test1/layout/README.md

+# 版面分析使用说明
+
+* [1. 安装whl包](#安装whl包)
+* [2. 使用](#使用)
+* [3. 后处理](#后处理)
+* [4. 指标](#指标)
+* [5. 训练版面分析模型](#训练版面分析模型)
+
+<a name="安装whl包"></a>
+
+## 1.  安装whl包
+```bash
+wget https://paddleocr.bj.bcebos.com/whl/layoutparser-0.0.0-py3-none-any.whl
+pip install -U layoutparser-0.0.0-py3-none-any.whl
+```
+
+<a name="使用"></a>
+
+## 2. 使用
+
+使用layoutparser识别给定文档的布局：
+
+```python
+import layoutparser as lp
+image = cv2.imread("imags/paper-image.jpg")
+image = image[..., ::-1]
+
+# 加载模型
+model = lp.PaddleDetectionLayoutModel(config_path="lp://PubLayNet/ppyolov2_r50vd_dcn_365e_publaynet/config", 
+                                threshold=0.5,
+                                label_map={0: "Text", 1: "Title", 2: "List", 3:"Table", 4:"Figure"},
+                                enforce_cpu=False, 
+                                enable_mkldnn=True)
+# 检测
+layout = model.detect(image)
+
+# 显示结果
+lp.draw_box(image, layout, box_width=3, show_element_type=True)
+```
+
+下图展示了结果，不同颜色的检测框表示不同的类别，并通过`show_element_type`在框的左上角显示具体类别：
+
+<div align="center">
+<img src="../../doc/table/result_all.jpg"  width = "600" />
+</div>
+
+`PaddleDetectionLayoutModel`函数参数说明如下:
+
+|      参数      |            含义             |   默认值    |                             备注                             |
+| :------------: | :-------------------------: | :---------: | :----------------------------------------------------------: |
+|  config_path   |        模型配置路径         |    None     | 指定config_path会自动下载模型(仅第一次，之后模型存在，不会再下载) |
+|   model_path   |          模型路径           |    None     | 本地模型路径，config_path和model_path必须设置一个，不能同时为None |
+|   threshold    |       预测得分的阈值        |     0.5     |                              \                               |
+|  input_shape   |     reshape之后图片尺寸     | [3,640,640] |                              \                               |
+|   batch_size   |       测试batch size        |      1      |                              \                               |
+|   label_map    |         类别映射表          |    None     | 设置config_path时，可以为None，根据数据集名称自动获取label_map |
+|  enforce_cpu   |     代码是否使用CPU运行     |    False    |         设置为False表示使用GPU，True表示强制使用CPU          |
+| enforce_mkldnn | CPU预测中是否开启MKLDNN加速 |    True     |                              \                               |
+|   thread_num   |        设置CPU线程数        |     10      |                              \                               |
+
+目前支持以下几种模型配置和label map，您可以通过修改 `--config_path`和 `--label_map`使用这些模型，从而检测不同类型的内容：
+
+| dataset                                                      | config_path                                                  | label_map                                                 |
+| ------------------------------------------------------------ | ------------------------------------------------------------ | --------------------------------------------------------- |
+| [TableBank](https://doc-analysis.github.io/tablebank-page/index.html) word | lp://TableBank/ppyolov2_r50vd_dcn_365e_tableBank_word/config | {0:"Table"}                                               |
+| TableBank latex                                              | lp://TableBank/ppyolov2_r50vd_dcn_365e_tableBank_latex/config | {0:"Table"}                                               |
+| [PubLayNet](https://github.com/ibm-aur-nlp/PubLayNet)        | lp://PubLayNet/ppyolov2_r50vd_dcn_365e_publaynet/config      | {0: "Text", 1: "Title", 2: "List", 3:"Table", 4:"Figure"} |
+
+* TableBank word和TableBank latex分别在word文档、latex文档数据集训练；
+* 下载TableBank数据集同时包含word和latex。
+
+<a name="后处理"></a>
+
+## 3. 后处理
+
+版面分析检测包含多个类别，如果只想获取指定类别(如"Text"类别)的检测框、可以使用下述代码：
+
+```python
+# 首先过滤特定文本类型的区域
+text_blocks = lp.Layout([b for b in layout if b.type=='Text'])
+figure_blocks = lp.Layout([b for b in layout if b.type=='Figure'])
+
+# 因为在图像区域内可能检测到文本区域，所以只需要删除它们
+text_blocks = lp.Layout([b for b in text_blocks \
+                   if not any(b.is_in(b_fig) for b_fig in figure_blocks)])
+
+# 对文本区域排序并分配id 
+h, w = image.shape[:2]
+
+left_interval = lp.Interval(0, w/2*1.05, axis='x').put_on_canvas(image)
+
+left_blocks = text_blocks.filter_by(left_interval, center=True)
+left_blocks.sort(key = lambda b:b.coordinates[1])
+
+right_blocks = [b for b in text_blocks if b not in left_blocks]
+right_blocks.sort(key = lambda b:b.coordinates[1])
+
+# 最终合并两个列表，并按顺序添加索引
+text_blocks = lp.Layout([b.set(id = idx) for idx, b in enumerate(left_blocks + right_blocks)])
+
+# 显示结果
+lp.draw_box(image, text_blocks,
+            box_width=3, 
+            show_element_id=True)
+```
+
+显示只有"Text"类别的结果：
+
+<div align="center">
+<img src="../../doc/table/result_text.jpg"  width = "600" />
+</div>
+
+<a name="指标"></a>
+
+## 4. 指标
+
+| Dataset   | mAP  | CPU time cost | GPU time cost |
+| --------- | ---- | ------------- | ------------- |
+| PubLayNet | 93.6 | 1713.7ms      | 66.6ms        |
+| TableBank | 96.2 | 1968.4ms      | 65.1ms        |
+
+**Envrionment：**	
+
+​	**CPU：**  Intel(R) Xeon(R) CPU E5-2650 v4 @ 2.20GHz，24core
+
+​	**GPU：**  a single NVIDIA Tesla P40
+
+<a name="训练版面分析模型"></a>
+
+## 5. 训练版面分析模型
+
+上述模型基于[PaddleDetection](https://github.com/PaddlePaddle/PaddleDetection) 训练，如果您想训练自己的版面分析模型，请参考：[train_layoutparser_model](train_layoutparser_model.md)
+

test1/layout/train_layoutparser_model.md

+# 训练版面分析
+
+* [1. 安装](#安装)
+  * [1.1 环境要求](#环境要求)
+  * [1.2 安装PaddleDetection](#安装PaddleDetection)
+* [2. 准备数据](#准备数据)
+* [3. 配置文件改动和说明](#配置文件改动和说明)
+* [4. PaddleDetection训练](#训练)
+* [5. PaddleDetection预测](#预测)
+* [6. 预测部署](#预测部署)
+  * [6.1 模型导出](#模型导出)
+  * [6.2 layout parser预测](#layout_parser预测)
+
+<a name="安装"></a>
+
+## 1. 安装
+
+<a name="环境要求"></a>
+
+### 1.1 环境要求
+
+- PaddlePaddle 2.1
+- OS 64 bit
+- Python 3(3.5.1+/3.6/3.7/3.8/3.9)，64 bit
+- pip/pip3(9.0.1+), 64 bit
+- CUDA >= 10.1
+- cuDNN >= 7.6
+
+<a name="安装PaddleDetection"></a>
+
+### 1.2 安装PaddleDetection
+
+```bash
+# 克隆PaddleDetection仓库
+cd <path/to/clone/PaddleDetection>
+git clone https://github.com/PaddlePaddle/PaddleDetection.git
+
+cd PaddleDetection
+# 安装其他依赖
+pip install -r requirements.txt
+```
+
+更多安装教程，请参考: [Install doc](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.1/docs/tutorials/INSTALL_cn.md)
+
+<a name="数据准备"></a>
+
+## 2. 准备数据
+
+下载 [PubLayNet](https://github.com/ibm-aur-nlp/PubLayNet) 数据集：
+
+```bash
+cd PaddleDetection/dataset/
+mkdir publaynet
+# 执行命令，下载
+wget -O publaynet.tar.gz https://dax-cdn.cdn.appdomain.cloud/dax-publaynet/1.0.0/publaynet.tar.gz?_ga=2.104193024.1076900768.1622560733-649911202.1622560733
+# 解压
+tar -xvf publaynet.tar.gz
+```
+
+解压之后PubLayNet目录结构：
+
+| File or Folder | Description                                      | num     |
+| :------------- | :----------------------------------------------- | ------- |
+| `train/`       | Images in the training subset                    | 335,703 |
+| `val/`         | Images in the validation subset                  | 11,245  |
+| `test/`        | Images in the testing subset                     | 11,405  |
+| `train.json`   | Annotations for training images                  |         |
+| `val.json`     | Annotations for validation images                |         |
+| `LICENSE.txt`  | Plaintext version of the CDLA-Permissive license |         |
+| `README.txt`   | Text file with the file names and description    |         |
+
+如果使用其它数据集，请参考[准备训练数据](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.1/docs/tutorials/PrepareDataSet.md)
+
+<a name="配置文件改动和说明"></a>
+
+## 3. 配置文件改动和说明
+
+我们使用 `configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml`配置进行训练，配置文件摘要如下：
+
+<div align='center'>
+  <img src='../../doc/table/PaddleDetection_config.png' width='600px'/>
+</div>
+
+从上图看到 `ppyolov2_r50vd_dcn_365e_coco.yml` 配置需要依赖其他的配置文件，在该例子中需要依赖:
+
+```
+coco_detection.yml：主要说明了训练数据和验证数据的路径
+
+runtime.yml：主要说明了公共的运行参数，比如是否使用GPU、每多少个epoch存储checkpoint等
+
+optimizer_365e.yml：主要说明了学习率和优化器的配置
+
+ppyolov2_r50vd_dcn.yml：主要说明模型和主干网络的情况
+
+ppyolov2_reader.yml：主要说明数据读取器配置，如batch size，并发加载子进程数等，同时包含读取后预处理操作，如resize、数据增强等等
+```
+
+根据实际情况，修改上述文件，比如数据集路径、batch size等。
+
+<a name="训练"></a>
+
+## 4. PaddleDetection训练
+
+PaddleDetection提供了单卡/多卡训练模式，满足用户多种训练需求
+
+* GPU 单卡训练
+
+```bash
+export CUDA_VISIBLE_DEVICES=0 #windows和Mac下不需要执行该命令
+python tools/train.py -c configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml
+```
+
+* GPU多卡训练
+
+```bash
+export CUDA_VISIBLE_DEVICES=0,1,2,3
+python -m paddle.distributed.launch --gpus 0,1,2,3 tools/train.py -c configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml --eval
+```
+
+--eval：表示边训练边验证
+
+* 模型恢复训练
+
+在日常训练过程中，有的用户由于一些原因导致训练中断，用户可以使用-r的命令恢复训练:
+
+```bash
+export CUDA_VISIBLE_DEVICES=0,1,2,3
+python -m paddle.distributed.launch --gpus 0,1,2,3 tools/train.py -c configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml --eval -r output/ppyolov2_r50vd_dcn_365e_coco/10000
+```
+
+注意：如果遇到 "`Out of memory error`" 问题, 尝试在 `ppyolov2_reader.yml` 文件中调小`batch_size`
+
+<a name="预测"></a>
+
+## 5. PaddleDetection预测
+
+设置参数，使用PaddleDetection预测：
+
+```bash
+export CUDA_VISIBLE_DEVICES=0
+python tools/infer.py -c configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml --infer_img=images/paper-image.jpg --output_dir=infer_output/ --draw_threshold=0.5 -o weights=output/ppyolov2_r50vd_dcn_365e_coco/model_final --use_vdl=Ture
+```
+
+`--draw_threshold` 是个可选参数. 根据 [NMS](https://ieeexplore.ieee.org/document/1699659) 的计算，不同阈值会产生不同的结果 `keep_top_k`表示设置输出目标的最大数量，默认值为100，用户可以根据自己的实际情况进行设定。
+
+<a name="预测部署"></a>
+
+## 6. 预测部署
+
+在layout parser中使用自己训练好的模型，
+
+<a name="模型导出"></a>
+
+### 6.1 模型导出
+
+在模型训练过程中保存的模型文件是包含前向预测和反向传播的过程，在实际的工业部署则不需要反向传播，因此需要将模型进行导成部署需要的模型格式。 在PaddleDetection中提供了 `tools/export_model.py`脚本来导出模型。
+
+导出模型名称默认是`model.*`，layout parser代码模型名称是`inference.*`,  所以修改[PaddleDetection/ppdet/engine/trainer.py ](https://github.com/PaddlePaddle/PaddleDetection/blob/b87a1ea86fa18ce69e44a17ad1b49c1326f19ff9/ppdet/engine/trainer.py#L512) (点开链接查看详细代码行)，将`model`改为`inference`即可。
+
+执行导出模型脚本：
+
+```bash
+python tools/export_model.py -c configs/ppyolo/ppyolov2_r50vd_dcn_365e_coco.yml --output_dir=./inference -o weights=output/ppyolov2_r50vd_dcn_365e_coco/model_final.pdparams
+```
+
+预测模型会导出到`inference/ppyolov2_r50vd_dcn_365e_coco`目录下，分别为`infer_cfg.yml`(预测不需要), `inference.pdiparams`, `inference.pdiparams.info`,`inference.pdmodel` 。
+
+更多模型导出教程，请参考：[EXPORT_MODEL](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.1/deploy/EXPORT_MODEL.md)
+
+<a name="layout parser预测"></a>
+
+### 6.2 layout_parser预测
+
+`model_path`指定训练好的模型路径，使用layout parser进行预测：
+
+```bash
+import layoutparser as lp
+model = lp.PaddleDetectionLayoutModel(model_path="inference/ppyolov2_r50vd_dcn_365e_coco", threshold=0.5,label_map={0: "Text", 1: "Title", 2: "List", 3:"Table", 4:"Figure"},enforce_cpu=True,enable_mkldnn=True)
+```
+
+
+
+***
+
+更多PaddleDetection训练教程，请参考：[PaddleDetection训练](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.1/docs/tutorials/GETTING_STARTED_cn.md)
+
+***
+

test/paddlestructure.py → test1/paddlestructure.py

@@ -24,9 +24,9 @@ import numpy as np
 from pathlib import Path
 
 from ppocr.utils.logging import get_logger
-from test.predict_system import OCRSystem, save_res
-from test.table.predict_table import to_excel
-from test.utility import init_args, draw_result
+from test1.predict_system import OCRSystem, save_res
+from test1.table.predict_table import to_excel
+from test1.utility import init_args, draw_result
 
 logger = get_logger()
 from ppocr.utils.utility import check_and_read_gif, get_image_file_list

test/predict_system.py → test1/predict_system.py

@@ -31,8 +31,8 @@ import layoutparser as lp
 from ppocr.utils.utility import get_image_file_list, check_and_read_gif
 from ppocr.utils.logging import get_logger
 from tools.infer.predict_system import TextSystem
-from test.table.predict_table import TableSystem, to_excel
-from test.utility import parse_args, draw_result
+from test1.table.predict_table import TableSystem, to_excel
+from test1.utility import parse_args, draw_result
 
 logger = get_logger()
 

test/setup.py → test1/setup.py

@@ -30,9 +30,9 @@ def readme():
     return README
 
 
-shutil.copytree('/table', './test/table')
-shutil.copyfile('/predict_system.py', './test/predict_system.py')
-shutil.copyfile('/utility.py', './test/utility.py')
+shutil.copytree('./table', './test1/table')
+shutil.copyfile('./predict_system.py', './test1/predict_system.py')
+shutil.copyfile('./utility.py', './test1/utility.py')
 shutil.copytree('../ppocr', './ppocr')
 shutil.copytree('../tools', './tools')
 shutil.copyfile('../LICENSE', './LICENSE')
@@ -68,5 +68,5 @@ setup(
 
 shutil.rmtree('ppocr')
 shutil.rmtree('tools')
-shutil.rmtree('test')
+shutil.rmtree('test1')
 os.remove('LICENSE')

test/table/eval_table.py → test1/table/eval_table.py

@@ -20,9 +20,9 @@ sys.path.append(os.path.abspath(os.path.join(__dir__, '../..')))
 import cv2
 import json
 from tqdm import tqdm
-from test.table.table_metric import TEDS
-from test.table.predict_table import TableSystem
-from test.utility import init_args
+from test1.table.table_metric import TEDS
+from test1.table.predict_table import TableSystem
+from test1.utility import init_args
 from ppocr.utils.logging import get_logger
 
 logger = get_logger()

test/table/predict_structure.py → test1/table/predict_structure.py

@@ -32,7 +32,7 @@ from ppocr.data import create_operators, transform
 from ppocr.postprocess import build_post_process
 from ppocr.utils.logging import get_logger
 from ppocr.utils.utility import get_image_file_list, check_and_read_gif
-from test.utility import parse_args
+from test1.utility import parse_args
 
 logger = get_logger()
 

test/table/predict_table.py → test1/table/predict_table.py

@@ -30,9 +30,9 @@ import tools.infer.predict_rec as predict_rec
 import tools.infer.predict_det as predict_det
 from ppocr.utils.utility import get_image_file_list, check_and_read_gif
 from ppocr.utils.logging import get_logger
-from test.table.matcher import distance, compute_iou
-from test.utility import parse_args
-import test.table.predict_structure as predict_strture
+from test1.table.matcher import distance, compute_iou
+from test1.utility import parse_args
+import test1.table.predict_structure as predict_strture
 
 logger = get_logger()
 

tools/eval.py

@@ -44,8 +44,15 @@ def main():
     # build model
     # for rec algorithm
     if hasattr(post_process_class, 'character'):
-        config['Architecture']["Head"]['out_channels'] = len(
-            getattr(post_process_class, 'character'))
+        char_num = len(getattr(post_process_class, 'character'))
+        if config['Architecture']["algorithm"] in ["Distillation",
+                                                   ]:  # distillation model
+            for key in config['Architecture']["Models"]:
+                config['Architecture']["Models"][key]["Head"][
+                    'out_channels'] = char_num
+        else:  # base rec model
+            config['Architecture']["Head"]['out_channels'] = char_num
+
     model = build_model(config['Architecture'])
     use_srn = config['Architecture']['algorithm'] == "SRN"
     model_type = config['Architecture']['model_type']

tools/infer/predict_det.py

@@ -31,7 +31,7 @@ from ppocr.utils.utility import get_image_file_list, check_and_read_gif
 from ppocr.data import create_operators, transform
 from ppocr.postprocess import build_post_process
 
-import tools.infer.benchmark_utils as benchmark_utils
+# import tools.infer.benchmark_utils as benchmark_utils
 
 logger = get_logger()
 
@@ -100,8 +100,6 @@ class TextDetector(object):
         self.predictor, self.input_tensor, self.output_tensors, self.config = utility.create_predictor(
             args, 'det', logger)
 
-        self.det_times = utility.Timer()
-
     def order_points_clockwise(self, pts):
         """
         reference from: https://github.com/jrosebr1/imutils/blob/master/imutils/perspective.py
@@ -158,8 +156,8 @@ class TextDetector(object):
     def __call__(self, img):
         ori_im = img.copy()
         data = {'image': img}
-        self.det_times.total_time.start()
-        self.det_times.preprocess_time.start()
+
+        st = time.time()
         data = transform(data, self.preprocess_op)
         img, shape_list = data
         if img is None:
@@ -168,16 +166,12 @@ class TextDetector(object):
         shape_list = np.expand_dims(shape_list, axis=0)
         img = img.copy()
 
-        self.det_times.preprocess_time.end()
-        self.det_times.inference_time.start()
-
         self.input_tensor.copy_from_cpu(img)
         self.predictor.run()
         outputs = []
         for output_tensor in self.output_tensors:
             output = output_tensor.copy_to_cpu()
             outputs.append(output)
-        self.det_times.inference_time.end()
 
         preds = {}
         if self.det_algorithm == "EAST":
@@ -193,8 +187,6 @@ class TextDetector(object):
         else:
             raise NotImplementedError
 
-        self.det_times.postprocess_time.start()
-
         self.predictor.try_shrink_memory()
         post_result = self.postprocess_op(preds, shape_list)
         dt_boxes = post_result[0]['points']
@@ -203,10 +195,8 @@ class TextDetector(object):
         else:
             dt_boxes = self.filter_tag_det_res(dt_boxes, ori_im.shape)
 
-        self.det_times.postprocess_time.end()
-        self.det_times.total_time.end()
-        self.det_times.img_num += 1
-        return dt_boxes, self.det_times.total_time.value()
+        et = time.time()
+        return dt_boxes, et - st
 
 
 if __name__ == "__main__":
@@ -216,12 +206,13 @@ if __name__ == "__main__":
     count = 0
     total_time = 0
     draw_img_save = "./inference_results"
-    cpu_mem, gpu_mem, gpu_util = 0, 0, 0
 
-    # warmup 10 times
-    fake_img = np.random.uniform(-1, 1, [640, 640, 3]).astype(np.float32)
+    if args.warmup:
+        img = np.random.uniform(0, 255, [640, 640, 3]).astype(np.uint8)
         for i in range(10):
-        dt_boxes, _ = text_detector(fake_img)
+            res = text_detector(img)
+
+    cpu_mem, gpu_mem, gpu_util = 0, 0, 0
 
     if not os.path.exists(draw_img_save):
         os.makedirs(draw_img_save)
@@ -239,12 +230,6 @@ if __name__ == "__main__":
             total_time += elapse
         count += 1
 
-        if args.benchmark:
-            cm, gm, gu = utility.get_current_memory_mb(0)
-            cpu_mem += cm
-            gpu_mem += gm
-            gpu_util += gu
-
         logger.info("Predict time of {}: {}".format(image_file, elapse))
         src_im = utility.draw_text_det_res(dt_boxes, image_file)
         img_name_pure = os.path.split(image_file)[-1]
@@ -252,36 +237,3 @@ if __name__ == "__main__":
                                 "det_res_{}".format(img_name_pure))
 
         logger.info("The visualized image saved in {}".format(img_path))
-    # print the information about memory and time-spent
-    if args.benchmark:
-        mems = {
-            'cpu_rss_mb': cpu_mem / count,
-            'gpu_rss_mb': gpu_mem / count,
-            'gpu_util': gpu_util * 100 / count
-        }
-    else:
-        mems = None
-    logger.info("The predict time about detection module is as follows: ")
-    det_time_dict = text_detector.det_times.report(average=True)
-    det_model_name = args.det_model_dir
-
-    if args.benchmark:
-        # construct log information
-        model_info = {
-            'model_name': args.det_model_dir.split('/')[-1],
-            'precision': args.precision
-        }
-        data_info = {
-            'batch_size': 1,
-            'shape': 'dynamic_shape',
-            'data_num': det_time_dict['img_num']
-        }
-        perf_info = {
-            'preprocess_time_s': det_time_dict['preprocess_time'],
-            'inference_time_s': det_time_dict['inference_time'],
-            'postprocess_time_s': det_time_dict['postprocess_time'],
-            'total_time_s': det_time_dict['total_time']
-        }
-        benchmark_log = benchmark_utils.PaddleInferBenchmark(
-            text_detector.config, model_info, data_info, perf_info, mems)
-        benchmark_log("Det")

tools/infer/predict_rec.py

@@ -257,13 +257,15 @@ def main(args):
     text_recognizer = TextRecognizer(args)
     valid_image_file_list = []
     img_list = []
-    cpu_mem, gpu_mem, gpu_util = 0, 0, 0
-    count = 0
 
     # warmup 10 times
-    fake_img = np.random.uniform(-1, 1, [1, 32, 320, 3]).astype(np.float32)
+    if args.warmup:
+        img = np.random.uniform(0, 255, [32, 320, 3]).astype(np.uint8)
         for i in range(10):
-        dt_boxes, _ = text_recognizer(fake_img)
+            res = text_recognizer([img])
+
+    cpu_mem, gpu_mem, gpu_util = 0, 0, 0
+    count = 0
 
     for image_file in image_file_list:
         img, flag = check_and_read_gif(image_file)

tools/infer/predict_system.py

@@ -13,6 +13,7 @@
 # limitations under the License.
 import os
 import sys
+import subprocess
 
 __dir__ = os.path.dirname(os.path.abspath(__file__))
 sys.path.append(__dir__)
@@ -93,7 +94,6 @@ class TextSystem(object):
         dt_boxes, elapse = self.text_detector(img)
 
         logger.debug("dt_boxes num : {}, elapse : {}".format(
-
             len(dt_boxes), elapse))
         if dt_boxes is None:
             return None, None
@@ -147,15 +147,24 @@ def sorted_boxes(dt_boxes):
 
 def main(args):
     image_file_list = get_image_file_list(args.image_dir)
+    image_file_list = image_file_list[args.process_id::args.total_process_num]
     text_sys = TextSystem(args)
     is_visualize = True
     font_path = args.vis_font_path
     drop_score = args.drop_score
+
+    # warm up 10 times
+    if args.warmup:
+        img = np.random.uniform(0, 255, [640, 640, 3]).astype(np.uint8)
+        for i in range(10):
+            res = text_sys(img)
+            
     total_time = 0
     cpu_mem, gpu_mem, gpu_util = 0, 0, 0
     _st = time.time()
     count = 0
     for idx, image_file in enumerate(image_file_list):
+
         img, flag = check_and_read_gif(image_file)
         if not flag:
             img = cv2.imread(image_file)
@@ -264,4 +273,18 @@ def main(args):
 
 
 if __name__ == "__main__":
-    main(utility.parse_args())
+    args = utility.parse_args()
+    if args.use_mp:
+        p_list = []
+        total_process_num = args.total_process_num
+        for process_id in range(total_process_num):
+            cmd = [sys.executable, "-u"] + sys.argv + [
+                "--process_id={}".format(process_id),
+                "--use_mp={}".format(False)
+            ]
+            p = subprocess.Popen(cmd, stdout=sys.stdout, stderr=sys.stdout)
+            p_list.append(p)
+        for p in p_list:
+            p.wait()
+    else:
+        main(args)

tools/infer/utility.py

@@ -105,7 +105,9 @@ def init_args():
     parser.add_argument("--enable_mkldnn", type=str2bool, default=False)
     parser.add_argument("--cpu_threads", type=int, default=10)
     parser.add_argument("--use_pdserving", type=str2bool, default=False)
+    parser.add_argument("--warmup", type=str2bool, default=True)
 
+    # multi-process
     parser.add_argument("--use_mp", type=str2bool, default=False)
     parser.add_argument("--total_process_num", type=int, default=1)
     parser.add_argument("--process_id", type=int, default=0)
@@ -113,7 +115,6 @@ def init_args():
     parser.add_argument("--benchmark", type=bool, default=False)
     parser.add_argument("--save_log_path", type=str, default="./log_output/")
 
-
     parser.add_argument("--show_log", type=str2bool, default=True)
     return parser