Generally, a more complex model would achive better performance in the task, but it also leads to some redundancy in the model. Model Pruning is a technique that reduces this redundancy by removing the sub-models in the neural network model, so as to reduce model calculation complexity and improve model inference performance.
This example uses PaddleSlim provided[APIs of Pruning](https://paddlepaddle.github.io/PaddleSlim/api/prune_api/) to compress the OCR model.
This example uses PaddleSlim provided[APIs of Pruning](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/docs/zh_cn/api_cn/dygraph/pruners) to compress the OCR model.
[PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim), an open source library which integrates model pruning, quantization (including quantization training and offline quantization), distillation, neural network architecture search, and many other commonly used and leading model compression technique in the industry.
It is recommended that you could understand following pages before reading this example：
...
...
@@ -35,7 +35,7 @@ PaddleOCR also provides a series of [models](../../../doc/doc_en/models_list_en.
### 3. Pruning sensitivity analysis
After the pre-trained model is loaded, sensitivity analysis is performed on each network layer of the model to understand the redundancy of each network layer, and save a sensitivity file which named: sen.pickle. After that, user could load the sensitivity file via the [methods provided by PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/paddleslim/prune/sensitive.py#L221) and determining the pruning ratio of each network layer automatically. For specific details of sensitivity analysis, see：[Sensitivity analysis](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/docs/zh_cn/tutorials/image_classification_sensitivity_analysis_tutorial.md)
After the pre-trained model is loaded, sensitivity analysis is performed on each network layer of the model to understand the redundancy of each network layer, and save a sensitivity file which named: sen.pickle. After that, user could load the sensitivity file via the [methods provided by PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/paddleslim/prune/sensitive.py#L221) and determining the pruning ratio of each network layer automatically. For specific details of sensitivity analysis, see：[Sensitivity analysis](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/docs/en/tutorials/image_classification_sensitivity_analysis_tutorial_en.md)
The function would return a dict after loading the sensitivity file. The keys of the dict are name of parameters in each layer. And the value of key is the information about pruning sensitivity of corresponding layer. In example, pruning 10% filter of the layer corresponding to conv10_expand_weights would lead to 0.65% degradation of model performance. The details could be seen at: [Sensitivity analysis](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/docs/zh_cn/algo/algo.md#2-%E5%8D%B7%E7%A7%AF%E6%A0%B8%E5%89%AA%E8%A3%81%E5%8E%9F%E7%90%86)
The function would return a dict after loading the sensitivity file. The keys of the dict are name of parameters in each layer. And the value of key is the information about pruning sensitivity of corresponding layer. In example, pruning 10% filter of the layer corresponding to conv10_expand_weights would lead to 0.65% degradation of model performance. The details could be seen at: [Sensitivity analysis](https://github.com/PaddlePaddle/PaddleSlim/blob/release/2.0-alpha/docs/zh_cn/algo/algo.md)
Enter the PaddleOCR root directory，perform sensitivity analysis on the model with the following command：
@@ -5,11 +5,11 @@ Generally, a more complex model would achieve better performance in the task, bu
Quantization is a technique that reduces this redundancy by reducing the full precision data to a fixed number,
so as to reduce model calculation complexity and improve model inference performance.
This example uses PaddleSlim provided [APIs of Quantization](https://paddlepaddle.github.io/PaddleSlim/api/quantization_api/) to compress the OCR model.
This example uses PaddleSlim provided [APIs of Quantization](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/docs/zh_cn/api_cn/dygraph/quanter/qat.rst) to compress the OCR model.
It is recommended that you could understand following pages before reading this example：
-[The training strategy of OCR model](../../../doc/doc_en/quickstart_en.md)
| model_type | Network Type | rec | Currently support`rec`,`det`,`cls` |
| algorithm | Model name | CRNN | See [algorithm_overview](./algorithm_overview_en.md) for the support list |
| **Transform** | Set the transformation method | - | Currently only recognition algorithms are supported, see [ppocr/modeling/transform](../../ppocr/modeling/transform) for details |
| **Transform** | Set the transformation method | - | Currently only recognition algorithms are supported, see [ppocr/modeling/transforms](../../ppocr/modeling/transforms) for details |
| name | Transformation class name | TPS | Currently supports `TPS` |
| num_fiducial | Number of TPS control points | 20 | Ten on the top and bottom |
@@ -94,14 +94,14 @@ The current open source models, data sets and magnitudes are as follows:
- Chinese data set, LSVT street view data set crops the image according to the truth value, and performs position calibration, a total of 30w images. In addition, based on the LSVT corpus, 500w of synthesized data.
- Small language data set, using different corpora and fonts, respectively generated 100w synthetic data set, and using ICDAR-MLT as the verification set.
Among them, the public data sets are all open source, users can search and download by themselves, or refer to [Chinese data set](../doc_ch/datasets.md), synthetic data is not open source, users can use open source synthesis tools to synthesize by themselves. Synthesis tools include [text_renderer](https://github.com/Sanster/text_renderer), [SynthText](https://github.com/ankush-me/SynthText), [TextRecognitionDataGenerator](https://github.com/Belval/TextRecognitionDataGenerator) etc.
Among them, the public data sets are all open source, users can search and download by themselves, or refer to [Chinese data set](./datasets_en.md), synthetic data is not open source, users can use open source synthesis tools to synthesize by themselves. Synthesis tools include [text_renderer](https://github.com/Sanster/text_renderer), [SynthText](https://github.com/ankush-me/SynthText), [TextRecognitionDataGenerator](https://github.com/Belval/TextRecognitionDataGenerator) etc.
<aname="22-vertical-scene"></a>
### 3.2 Vertical Scene
PaddleOCR mainly focuses on general OCR. If you have vertical requirements, you can use PaddleOCR + vertical data to train yourself;
If there is a lack of labeled data, or if you do not want to invest in research and development costs, it is recommended to directly call the open API, which covers some of the more common vertical categories.
If there is a lack of labeled data, or if you do not want to invest in research and development costs, it is recommended to directly call the open API, which covers some of the more common vertical categories.
<aname="23-build-your-own-data-set"></a>
...
...
@@ -147,8 +147,8 @@ There are several experiences for reference when constructing the data set:
***
Click the following links for detailed training tutorial:
Click the following links for detailed training tutorial:
-[text detection model training](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/doc/doc_ch/detection.md)
-[text recognition model training](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/doc/doc_ch/recognition.md)
-[text direction classification model training](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/doc/doc_ch/angle_class.md)
-[text detection model training](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/doc/doc_ch/detection.md)
-[text recognition model training](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/doc/doc_ch/recognition.md)
-[text direction classification model training](https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.3/doc/doc_ch/angle_class.md)
@@ -12,25 +12,25 @@ Here we have sorted out some Chinese OCR training and prediction tricks, which a
At present, ResNet_vd series and MobileNetV3 series are the backbone networks used in PaddleOCR, whether replacing the other backbone networks will help to improve the accuracy? What should be paid attention to when replacing?
-**Tips**
- Whether text detection or text recognition, the choice of backbone network is a trade-off between prediction effect and prediction efficiency. Generally, a larger backbone network is selected, e.g. ResNet101_vd, then the performance of the detection or recognition is more accurate, but the time cost will increase accordingly. And a smaller backbone network is selected, e.g. MobileNetV3_small_x0_35, the prediction speed is faster, but the accuracy of detection or recognition will be reduced. Fortunately, the detection or recognition effect of different backbone networks is positively correlated with the performance of ImageNet 1000 classification task. [**PaddleClas**](https://github.com/PaddlePaddle/PaddleClas/blob/master/README_en.md) have sorted out the 23 series of classification network structures, such as ResNet_vd、Res2Net、HRNet、MobileNetV3、GhostNet. It provides the top1 accuracy of classification, the time cost of GPU(V100 and T4) and CPU(SD 855), and the 117 pretrained models [**download addresses**](https://paddleclas-en.readthedocs.io/en/latest/models/models_intro_en.html).
- Whether text detection or text recognition, the choice of backbone network is a trade-off between prediction effect and prediction efficiency. Generally, a larger backbone network is selected, e.g. ResNet101_vd, then the performance of the detection or recognition is more accurate, but the time cost will increase accordingly. And a smaller backbone network is selected, e.g. MobileNetV3_small_x0_35, the prediction speed is faster, but the accuracy of detection or recognition will be reduced. Fortunately, the detection or recognition effect of different backbone networks is positively correlated with the performance of ImageNet 1000 classification task. [**PaddleClas**](https://github.com/PaddlePaddle/PaddleClas/blob/release/2.3/docs/en/models/models_intro_en.md) have sorted out the 23 series of classification network structures, such as ResNet_vd、Res2Net、HRNet、MobileNetV3、GhostNet. It provides the top1 accuracy of classification, the time cost of GPU(V100 and T4) and CPU(SD 855), and the 117 pretrained models [**download addresses**](https://paddleclas-en.readthedocs.io/en/latest/models/models_intro_en.html).
- Similar as the 4 stages of ResNet, the replacement of text detection backbone network is to determine those four stages to facilitate the integration of FPN like the object detection heads. In addition, for the text detection problem, the pre trained model in ImageNet1000 can accelerate the convergence and improve the accuracy.
- In order to replace the backbone network of text recognition, we need to pay attention to the descending position of network width and height stride. Since the ratio between width and height is large in chinese text recognition, the frequency of height decrease is less and the frequency of width decrease is more. You can refer the [modifies of MobileNetV3](https://github.com/PaddlePaddle/PaddleOCR/blob/develop/ppocr/modeling/backbones/rec_mobilenet_v3.py) in PaddleOCR.
<aname="LongChineseTextRecognition"></a>
#### 2、Long Chinese Text Recognition
-**Problem Description**
-**Problem Description**
The maximum resolution of Chinese recognition model during training is [3,32,320], if the text image to be recognized is too long, as shown in the figure below, how to adapt?
During the training, the training samples are not directly resized to [3,32,320]. At first, the height of samples are resized to 32 and keep the ratio between the width and the height. When the width is less than 320, the excess parts are padding 0. Besides, when the ratio between the width and the height of the samples is larger than 10, these samples will be ignored. When the prediction for one image, do as above, but do not limit the max ratio between the width and the height. When the prediction for an images batch, do as training, but the resized target width is the longest width of the images in the batch. [Code as following](https://github.com/PaddlePaddle/PaddleOCR/blob/develop/tools/infer/predict_rec.py)：
```
def resize_norm_img(self, img, max_wh_ratio):
imgC, imgH, imgW = self.rec_image_shape
...
...
@@ -58,11 +58,11 @@ Here we have sorted out some Chinese OCR training and prediction tricks, which a
-**Problem Description**
As shown in the figure below, for Chinese and English mixed scenes, in order to facilitate reading and using the recognition results, it is often necessary to recognize the spaces between words. How can this situation be adapted?
There are two possible methods for space recognition. (1) Optimize the text detection. For spliting the text at the space in detection results, it needs to divide the text line with space into many segments when label the data for detection. (2) Optimize the text recognition. The space character is introduced into the recognition dictionary. Label the blank line in the training data for text recognition. In addition, we can also concat multiple word lines to synthesize the training data with spaces. PaddleOCR currently uses the second method.
-[7.2 OCR and table recognition model](#72-ocr-and-table-recognition-model)
-[7.3 DOC-VQA model](#73-doc-vqa-model)
## 1. Introduction
PP-Structure is an OCR toolkit that can be used for document analysis and processing with complex structures, designed to help developers better complete document understanding tasks
<aname="2"></a>
## 2. Update log
* 2022.02.12 DOC-VQA add LayoutLMv2 model。
* 2021.12.07 add [DOC-VQA SER and RE tasks](vqa/README.md)。
<aname="3"></a>
## 3. Features
The main features of PP-Structure are as follows:
...
...
@@ -38,21 +36,14 @@ The main features of PP-Structure are as follows:
- Support custom training for layout analysis and table structure tasks
- Support Document Visual Question Answering (DOC-VQA) tasks: Semantic Entity Recognition (SER) and Relation Extraction (RE)
The figure shows the pipeline of layout analysis + table recognition. The image is first divided into four areas of image, text, title and table by layout analysis, and then OCR detection and recognition is performed on the three areas of image, text and title, and the table is performed table recognition, where the image will also be stored for use.
<aname="42"></a>
### 4.2 DOC-VQA
* SER
...
...
@@ -77,19 +68,12 @@ The corresponding category and OCR recognition results are also marked at the to
In the figure, the red box represents the question, the blue box represents the answer, and the question and answer are connected by green lines. The corresponding category and OCR recognition results are also marked at the top left of the OCR detection box.
<aname="5"></a>
## 5. Quick start
Start from [Quick Installation](./docs/quickstart.md)
<aname="6"></a>
## 6. PP-Structure System
<aname="61"></a>
### 6.1 Layout analysis and table recognition
...
...
@@ -104,39 +88,33 @@ Layout analysis classifies image by region, including the use of Python scripts
Table recognition converts table images into excel documents, which include the detection and recognition of table text and the prediction of table structure and cell coordinates. For detailed instructions, please refer to [document](table/README.md)
<aname="62"></a>
### 6.2 DOC-VQA
Document Visual Question Answering (DOC-VQA) if a type of Visual Question Answering (VQA), which includes Semantic Entity Recognition (SER) and Relation Extraction (RE) tasks. Based on SER task, text recognition and classification in images can be completed. Based on THE RE task, we can extract the relation of the text content in the image, such as judge the problem pair. For details, please refer to [document](vqa/README.md)
<aname="7"></a>
## 7. Model List
PP-Structure系列模型列表（更新中）
PP-Structure Series Model List (Updating)
* Layout analysis model
### 7.1 Layout analysis model
|model name|description|download|
| --- | --- | --- |
| ppyolov2_r50vd_dcn_365e_publaynet | The layout analysis model trained on the PubLayNet dataset can divide image into 5 types of areas **text, title, table, picture, and list** | [PubLayNet](https://paddle-model-ecology.bj.bcebos.com/model/layout-parser/ppyolov2_r50vd_dcn_365e_publaynet.tar) |
|ch_ppocr_mobile_slim_v2.0_rec|Slim pruned and quantized lightweight model, supporting Chinese, English and number recognition|6M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_train.tar) |
|en_ppocr_mobile_v2.0_table_structure|Table structure prediction of English table scene trained on PubLayNet dataset|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_structure_train.tar) |
|ch_PP-OCRv2_det_slim|[New] Slim quantization with distillation lightweight model, supporting Chinese, English, multilingual text detection| 3M |[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_slim_quant_infer.tar)|
|ch_PP-OCRv2_rec_slim|[New] Slim qunatization with distillation lightweight model, supporting Chinese, English, multilingual text recognition| 9M |[inference model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_slim_quant_train.tar) |
|en_ppocr_mobile_v2.0_table_structure|Table structure prediction of English table scene trained on PubLayNet dataset| 18.6M |[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_structure_train.tar) |
* DOC-VQA model
### 7.3 DOC-VQA model
|model name|description|model size|download|
| --- | --- | --- | --- |
|PP-Layout_v1.0_ser_pretrained|SER model trained on xfun Chinese dataset based on LayoutXLM|1.4G|[inference model coming soon]() / [trained model](https://paddleocr.bj.bcebos.com/pplayout/PP-Layout_v1.0_ser_pretrained.tar) |
|PP-Layout_v1.0_re_pretrained|RE model trained on xfun Chinese dataset based on LayoutXLM|1.4G|[inference model coming soon]() / [trained model](https://paddleocr.bj.bcebos.com/pplayout/PP-Layout_v1.0_re_pretrained.tar) |
|ser_LayoutXLM_xfun_zhd|SER model trained on xfun Chinese dataset based on LayoutXLM|1.4G|[inference model coming soon]() / [trained model](https://paddleocr.bj.bcebos.com/pplayout/ser_LayoutXLM_xfun_zh.tar) |
|re_LayoutXLM_xfun_zh|RE model trained on xfun Chinese dataset based on LayoutXLM|1.4G|[inference model coming soon]() / [trained model](https://paddleocr.bj.bcebos.com/pplayout/re_LayoutXLM_xfun_zh.tar) |
If you need to use other models, you can download the model in [PPOCR model_list](../doc/doc_en/models_list_en.md) and [PPStructure model_list](./docs/model_list.md)
If you need to use other models, you can download the model in [PPOCR model_list](../doc/doc_en/models_list_en.md) and [PPStructure model_list](./docs/models_list.md)
