It can be seen that high Tpr can be getted when backbone is SwinTranformer_tiny, but the speed is slow. Replacing backbone with the lightweight model MobileNetV3_small_x0_35, the speed can be greatly improved, but the Tpr will be greatly reduced. Replacing backbone with faster backbone PPLCNet_x1_0, the Tpr is higher more 13 percentage points than MobileNetv3_small_x0_35. At the same time, the speed can be more than 20% faster. After additional using the SSLD pretrained model, the Tpr can be improved by about 0.7 percentage points without affecting the inference speed. Finally, after additional using the SKL-UGI knowledge distillation, the Tpr can be further improved by 0.44 percentage points. At this point, the Tpr close to that of SwinTranformer_tiny, but the speed is more than 40 times faster. The training method and deployment instructions of PULC will be introduced in detail below.
It can be seen that high Tpr can be getted when backbone is SwinTranformer_tiny, but the speed is slow. Replacing backbone with the lightweight model MobileNetV3_small_x0_35, the speed can be greatly improved, but the Tpr will be greatly reduced. Replacing backbone with faster backbone PPLCNet_x1_0, the Tpr is higher more 13 percentage points than MobileNetv3_small_x0_35. At the same time, the speed can be more than 20% faster. After additional using the SSLD pretrained model, the Tpr can be improved by about 0.7 percentage points without affecting the inference speed. Finally, after additional using the SKL-UGI knowledge distillation, the Tpr can be further improved by 0.44 percentage points. At this point, the Tpr is close to that of SwinTranformer_tiny, but the speed is more than 40 times faster. The training method and deployment instructions of PULC will be introduced in detail below.
This case provides a way for users to quickly build a lightweight, high-precision and practical classification model of language in the image using PaddleClas PULC (Practical Ultra Lightweight Classification). The model can be widely used in various scenarios involving multilingual OCR processing, such as finance and government affairs.
The following table lists the relevant indicators of the model. The first two lines means that using SwinTransformer_tiny and MobileNetV3_small_x0_35 as the backbone to training. The third to sixth lines means that the backbone is replaced by PPLCNet, additional use of EDA strategy and additional use of EDA strategy and SKL-UGI knowledge distillation strategy. When replacing the backbone with PPLCNet_x1_0, the input shape of model is changed to [192, 48], and the stride of the network is changed to [2, [2, 1], [2, 1], [2, 1]].
It can be seen that high accuracy can be getted when backbone is SwinTranformer_tiny, but the speed is slow. Replacing backbone with the lightweight model MobileNetV3_small_x0_35, the speed can be greatly improved, but the accuracy will be greatly reduced. Replacing backbone with faster backbone PPLCNet_x1_0 and changing the input shape and stride of network, the accuracy is higher more 2.43 percentage points than MobileNetv3_small_x0_35. At the same time, the speed can be more than 20% faster. After additional using the SSLD pretrained model, the accuracy can be improved by about 0.35 percentage points without affecting the inference speed. Further, additional using the EDA strategy, the accuracy can be increased by 0.42 percentage points. Finally, after additional using the SKL-UGI knowledge distillation, the accuracy can be further improved by 0.14 percentage points. At this point, the accuracy is higher than that of SwinTranformer_tiny, but the speed is more faster. The training method and deployment instructions of PULC will be introduced in detail below.
**Note**:
* The Latency is tested on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz. The MKLDNN is enabled and the number of threads is 10.
* About PP-LCNet, please refer to [PP-LCNet Introduction](../models/PP-LCNet_en.md) and [PP-LCNet Paper](https://arxiv.org/abs/2109.15099).
<aname="2"></a>
## 2. Quick Start
<aname="2.1"></a>
### 2.1 PaddlePaddle Installation
- Run the following command to install if CUDA9 or CUDA10 is available.
Please refer to [PaddlePaddle Installation](https://www.paddlepaddle.org.cn/install/quick?docurl=/documentation/docs/en/install/pip/linux-pip_en.html) for more information about installation, for examples other versions.
<aname="2.2"></a>
### 2.2 PaddleClas wheel Installation
The command of PaddleClas installation as bellow:
```bash
pip3 install paddleclas
```
<aname="2.3"></a>
### 2.3 Prediction
First, please click [here](https://paddleclas.bj.bcebos.com/data/PULC/pulc_demo_imgs.zip) to download and unzip to get the test demo images.
**Note**: The `result` returned by `model.predict()` is a generator, so you need to use the `next()` function to call it or `for` loop to loop it. And it will predict with `batch_size` size batch and return the prediction results when called. The default `batch_size` is 1, and you also specify the `batch_size` when instantiating, such as `model = paddleclas.PaddleClas(model_name="language_classification", batch_size=2)`. The result of demo above:
Please refer to [Installation](../installation/install_paddleclas_en.md) to get the description about installation.
<aname="3.2"></a>
### 3.2 Dataset
<aname="3.2.1"></a>
#### 3.2.1 Dataset Introduction
The models wo provided are trained with internal data, which is not open source yet. So it is suggested that constructing dataset based on open source dataset [Multi-lingual scene text detection and recognition](https://rrc.cvc.uab.es/?ch=15&com=downloads) to experience the this case.
Some image of the processed dataset is as follows:
The models provided support to classcify 10 languages, which as shown in the following list:
`0` : means Arabic
`1` : means chinese_cht
`2` : means cyrillic
`3` : means devanagari
`4` : means Japanese
`5` : means ka
`6` : means Korean
`7` : means ta
`8` : means te
`9` : means Latin
In the `Multi-lingual scene text detection and recognition`, only Arabic, Japanese, Korean and Latin data are included. 1600 images from each of the four languages are taken as the training data of this case, 300 images as the evaluation data, and 400 images as the supplementary data is used for the `SKL-UGI Knowledge Distillation`.
Therefore, for the demo dataset in this case, the language categories are shown in following list:
`0` : means arabic
`4` : means japan
`6` : means korean
`9` : means latin
**Note**: The images used in this task should be cropped by text from original image. Only the text line part is used as the image data.
If you want to create your own dataset, you can collect and sort out the data of the required languages in your task as required. And you can also download the data processed directly.
```
cd path_to_PaddleClas
```
Enter the `dataset/` directory, download and unzip the dataset.
The datas under `language_classification` directory:
```
├── img
│ ├── word_1.png
│ ├── word_2.png
...
├── train_list.txt
├── train_list_for_distill.txt
├── test_list.txt
└── label_list.txt
```
Where `img/` is the directory including 9200 images in 4 languages. The `train_list.txt` and `test_list.txt` are label files of training data and validation data respectively. `label_list.txt` is the mapping file corresponding to the four languages. `train_list_for_distill.txt` is the label list of images used for `SKL-UGI Knowledge Distillation`.
**Note**:
* About the contents format of `train_list.txt` and `val_list.txt`, please refer to [Description about Classification Dataset in PaddleClas](../data_preparation/classification_dataset_en.md).
* About the `train_list_for_distill.txt`, please refer to [Knowledge Distillation Label](../advanced_tutorials/distillation/distillation_en.md).
<aname="3.3"></a>
### 3.3 Training
The details of training config in `ppcls/configs/PULC/person_exists/PPLCNet_x1_0.yaml`. The command about training as follows:
**Note**: Because the class num of demo dataset is 4, the argument `-o Arch.class_num=4` should be specifed to change the prediction class num of model to 4.
<aname="3.4"></a>
### 3.4 Evaluation
After training, you can use the following commands to evaluate the model.
Among the above command, the argument `-o Global.pretrained_model="output/PPLCNet_x1_0/best_model"` specify the path of the best model weight file. You can specify other path if needed.
<aname="3.5"></a>
### 3.5 Inference
After training, you can use the model that trained to infer. Command is as follow:
* Among the above command, argument `-o Global.pretrained_model="output/PPLCNet_x1_0/best_model"` specify the path of the best model weight file. You can specify other path if needed.
* The default test image is `deploy/images/PULC/person_exists/objects365_02035329.jpg`. And you can test other image, only need to specify the argument `-o Infer.infer_imgs=path_to_test_image`.
* Among the prediction results, `japan` means japanese and `korean` means korean.
<aname="4"></a>
## 4. Model Compression
<aname="4.1"></a>
### 4.1 SKL-UGI Knowledge Distillation
SKL-UGI is a simple but effective knowledge distillation algrithem proposed by PaddleClas.
<!-- todo -->
<!-- Please refer to [SKL-UGI](../advanced_tutorials/distillation/distillation_en.md) for more details. -->
<aname="4.1.1"></a>
#### 4.1.1 Teacher Model Training
Training the teacher model with hyperparameters specified in `ppcls/configs/PULC/language_classification/PPLCNet/PPLCNet_x1_0.yaml`. The command is as follow:
The best teacher model weight would be saved in file `output/ResNet101_vd/best_model.pdparams`.
**Note**: Training the ResNet101_vd model requires more GPU memory. If the memory is not enough, you can reduce the learning rate and batch size in the same proportion.
<aname="4.1.2"></a>
#### 4.1.2 Knowledge Distillation Training
The training strategy, specified in training config file `ppcls/configs/PULC/language_classification/PPLCNet_x1_0_distillation.yaml`, the teacher model is `ResNet101_vd`, the student model is `PPLCNet_x1_0`. The command is as follow:
The best student model weight would be saved in file `output/DistillationModel/best_model_student.pdparams`.
<aname="5"></a>
## 5. Hyperparameters Searching
The hyperparameters used by [3.2 section](#3.2) and [4.1 section](#4.1) are according by `Hyperparameters Searching` in PaddleClas. If you want to get better results on your own dataset, you can refer to [Hyperparameters Searching](PULC_train_en.md#4) to get better hyperparameters.
**Note**: This section is optional. Because the search process will take a long time, you can selectively run according to your specific. If not replace the dataset, you can ignore this section.
<aname="6"></a>
## 6. Inference Deployment
<aname="6.1"></a>
### 6.1 Getting Paddle Inference Model
Paddle Inference is the original Inference Library of the PaddlePaddle, provides high-performance inference for server deployment. And compared with directly based on the pretrained model, Paddle Inference can use tools to accelerate prediction, so as to achieve better inference performance. Please refer to [Paddle Inference](https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/infer/inference/inference_cn.html) for more information.
Paddle Inference need Paddle Inference Model to predict. Two process provided to get Paddle Inference Model. If want to use the provided by PaddleClas, you can download directly, click [Downloading Inference Model](#6.1.2).
<aname="6.1.1"></a>
### 6.1.1 Exporting Paddle Inference Model
The command about exporting Paddle Inference Model is as follow:
After running above command, the inference model files would be saved in `deploy/models/PPLCNet_x1_0_language_classification_infer`, as shown below:
```
├── PPLCNet_x1_0_language_classification_infer
│ ├── inference.pdiparams
│ ├── inference.pdiparams.info
│ └── inference.pdmodel
```
**Note**: The best model is from knowledge distillation training. If knowledge distillation training is not used, the best model would be saved in `output/PPLCNet_x1_0/best_model.pdparams`.
<aname="6.1.2"></a>
### 6.1.2 Downloading Inference Model
You can also download directly.
```
cd deploy/models
# download the inference model and decompression
wget https://paddleclas.bj.bcebos.com/models/PULC/language_classification_infer.tar && tar -xf language_classification_infer.tar
```
After decompression, the directory `models` should be shown below.
```
├── language_classification_infer
│ ├── inference.pdiparams
│ ├── inference.pdiparams.info
│ └── inference.pdmodel
```
<aname="6.2"></a>
### 6.2 Prediction with Python
<aname="6.2.1"></a>
#### 6.2.1 Image Prediction
Return the directory `deploy`:
```
cd ../
```
Run the following command to classify language about the image `./images/PULC/language_classification/word_35404.png`.
**Note**: Among the prediction results, `japan` means japanese and `korean` means korean.
<aname="6.2.2"></a>
#### 6.2.2 Images Prediction
If you want to predict images in directory, please specify the argument `Global.infer_imgs` as directory path by `-o Global.infer_imgs`. The command is as follow.
```shell
# Use the following command to predict with GPU. If want to replace with CPU, you can add argument -o Global.use_gpu=False
Among the prediction results above, `japan` means japanese, `latin` means latin, `arabic` means arabic and `korean` means korean.
<aname="6.3"></a>
### 6.3 Deployment with C++
PaddleClas provides an example about how to deploy with C++. Please refer to [Deployment with C++](../inference_deployment/cpp_deploy_en.md).
<aname="6.4"></a>
### 6.4 Deployment as Service
Paddle Serving is a flexible, high-performance carrier for machine learning models, and supports different protocol, such as RESTful, gRPC, bRPC and so on, which provides different deployment solutions for a variety of heterogeneous hardware and operating system environments. Please refer [Paddle Serving](https://github.com/PaddlePaddle/Serving) for more information.
PaddleClas provides an example about how to deploy as service by Paddle Serving. Please refer to [Paddle Serving Deployment](../inference_deployment/paddle_serving_deploy_en.md).
<aname="6.5"></a>
### 6.5 Deployment on Mobile
Paddle-Lite is an open source deep learning framework that designed to make easy to perform inference on mobile, embeded, and IoT devices. Please refer to [Paddle-Lite](https://github.com/PaddlePaddle/Paddle-Lite) for more information.
PaddleClas provides an example of how to deploy on mobile by Paddle-Lite. Please refer to [Paddle-Lite deployment](../inference_deployment/paddle_lite_deploy_en.md).
<aname="6.6"></a>
### 6.6 Converting To ONNX and Deployment
Paddle2ONNX support convert Paddle Inference model to ONNX model. And you can deploy with ONNX model on different inference engine, such as TensorRT, OpenVINO, MNN/TNN, NCNN and so on. About Paddle2ONNX details, please refer to [Paddle2ONNX](https://github.com/PaddlePaddle/Paddle2ONNX).
PaddleClas provides an example of how to convert Paddle Inference model to ONNX model by paddle2onnx toolkit and predict by ONNX model. You can refer to [paddle2onnx](../../../deploy/paddle2onnx/readme_en.md) for deployment details.
It can be seen that high Tpr can be getted when backbone is SwinTranformer_tiny, but the speed is slow. Replacing backbone with the lightweight model MobileNetV3_small_x0_35, the speed can be greatly improved, but the Tpr will be greatly reduced. Replacing backbone with faster backbone PPLCNet_x1_0, the Tpr is higher more 20 percentage points than MobileNetv3_small_x0_35. At the same time, the speed can be more than 20% faster. After additional using the SSLD pretrained model, the Tpr can be improved by about 2.6 percentage points without affecting the inference speed. Further, additional using the EDA strategy, the Tpr can be increased by 1.3 percentage points. Finally, after additional using the SKL-UGI knowledge distillation, the Tpr can be further improved by 2.2 percentage points. At this point, the Tpr close to that of SwinTranformer_tiny, but the speed is more than 40 times faster. The training method and deployment instructions of PULC will be introduced in detail below.
It can be seen that high Tpr can be getted when backbone is SwinTranformer_tiny, but the speed is slow. Replacing backbone with the lightweight model MobileNetV3_small_x0_35, the speed can be greatly improved, but the Tpr will be greatly reduced. Replacing backbone with faster backbone PPLCNet_x1_0, the Tpr is higher more 20 percentage points than MobileNetv3_small_x0_35. At the same time, the speed can be more than 20% faster. After additional using the SSLD pretrained model, the Tpr can be improved by about 2.6 percentage points without affecting the inference speed. Further, additional using the EDA strategy, the Tpr can be increased by 1.3 percentage points. Finally, after additional using the SKL-UGI knowledge distillation, the Tpr can be further improved by 2.2 percentage points. At this point, the Tpr is close to that of SwinTranformer_tiny, but the speed is more than 40 times faster. The training method and deployment instructions of PULC will be introduced in detail below.
# PULC Classification Model of Wheather Wearing Safety Helmet or Not
-----
## Catalogue
-[1. Introduction](#1)
-[2. Quick Start](#2)
-[2.1 PaddlePaddle Installation](#2.1)
-[2.2 PaddleClas Installation](#2.2)
-[2.3 Prediction](#2.3)
-[3. Training, Evaluation and Inference](#3)
-[3.1 Installation](#3.1)
-[3.2 Dataset](#3.2)
-[3.2.1 Dataset Introduction](#3.2.1)
-[3.2.2 Getting Dataset](#3.2.2)
-[3.3 Training](#3.3)
-[3.4 Evaluation](#3.4)
-[3.5 Inference](#3.5)
-[4. Model Compression](#4)
-[4.1 SKL-UGI Knowledge Distillation](#4.1)
-[4.1.1 Teacher Model Training](#4.1.1)
-[4.1.2 Knowledge Distillation Training](#4.1.2)
-[5. SHAS](#5)
-[6. Inference Deployment](#6)
-[6.1 Getting Paddle Inference Model](#6.1)
-[6.1.1 Exporting Paddle Inference Model](#6.1.1)
-[6.1.2 Downloading Inference Model](#6.1.2)
-[6.2 Prediction with Python](#6.2)
-[6.2.1 Image Prediction](#6.2.1)
-[6.2.2 Images Prediction](#6.2.2)
-[6.3 Deployment with C++](#6.3)
-[6.4 Deployment as Service](#6.4)
-[6.5 Deployment on Mobile](#6.5)
-[6.6 Converting To ONNX and Deployment](#6.6)
<aname="1"></a>
## 1. Introduction
This case provides a way for users to quickly build a lightweight, high-precision and practical classification model of wheather wearing safety helmet using PaddleClas PULC (Practical Ultra Lightweight Classification). The model can be widely used in construction scenes, factory workshop scenes, traffic scenes and so on.
The following table lists the relevant indicators of the model. The first two lines means that using SwinTransformer_tiny and MobileNetV3_small_x0_35 as the backbone to training. The third to seventh lines means that the backbone is replaced by PPLCNet, additional use of EDA strategy and additional use of EDA strategy and SKL-UGI knowledge distillation strategy.
It can be seen that high Tpr can be getted when backbone is Res2Net200_vd_26w_4s, but the speed is slow. Replacing backbone with the lightweight model MobileNetV3_small_x0_35, the speed can be greatly improved, but the Tpr will be greatly reduced. Replacing backbone with faster backbone PPLCNet_x1_0, the Tpr is higher more 8.5 percentage points than MobileNetv3_small_x0_35. At the same time, the speed can be more than 20% faster. After additional using the SSLD pretrained model, the Tpr can be improved by about 4.9 percentage points without affecting the inference speed. Further, additional using the EDA strategy, the Tpr can be increased by 1.1 percentage points. Finally, after additional using the UDML knowledge distillation, the Tpr can be further improved by 2.2 percentage points. At this point, the Tpr is higher than that of Res2Net200_vd_26w_4s, but the speed is more than 70 times faster. The training method and deployment instructions of PULC will be introduced in detail below.
**Note**:
* About `Tpr` metric, please refer to [3.2 section](#3.2) for more information .
* The Latency is tested on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz. The MKLDNN is enabled and the number of threads is 10.
* About PP-LCNet, please refer to [PP-LCNet Introduction](../models/PP-LCNet_en.md) and [PP-LCNet Paper](https://arxiv.org/abs/2109.15099).
<aname="2"></a>
## 2. Quick Start
<aname="2.1"></a>
### 2.1 PaddlePaddle Installation
- Run the following command to install if CUDA9 or CUDA10 is available.
Please refer to [PaddlePaddle Installation](https://www.paddlepaddle.org.cn/install/quick?docurl=/documentation/docs/en/install/pip/linux-pip_en.html) for more information about installation, for examples other versions.
<aname="2.2"></a>
### 2.2 PaddleClas wheel Installation
The command of PaddleClas installation as bellow:
```bash
pip3 install paddleclas
```
<aname="2.3"></a>
### 2.3 Prediction
First, please click [here](https://paddleclas.bj.bcebos.com/data/PULC/pulc_demo_imgs.zip) to download and unzip to get the test demo images.
**Note**: The `result` returned by `model.predict()` is a generator, so you need to use the `next()` function to call it or `for` loop to loop it. And it will predict with `batch_size` size batch and return the prediction results when called. The default `batch_size` is 1, and you also specify the `batch_size` when instantiating, such as `model = paddleclas.PaddleClas(model_name="safety_helmet", batch_size=2)`. The result of demo above:
Please refer to [Installation](../installation/install_paddleclas_en.md) to get the description about installation.
<aname="3.2"></a>
### 3.2 Dataset
<aname="3.2.1"></a>
#### 3.2.1 Dataset Introduction
All datasets used in this case are open source data. Train data is the subset of [Safety-Helmet-Wearing-Dataset](https://github.com/njvisionpower/Safety-Helmet-Wearing-Dataset), [hard-hat-detection](https://www.kaggle.com/datasets/andrewmvd/hard-hat-detection) and [Large-scale CelebFaces Attributes (CelebA) Dataset](https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html).
<aname="3.2.2"></a>
#### 3.2.2 Getting Dataset
The data used in this case can be getted by processing the open source data. The detailed processes are as follows:
*`Safety-Helmet-Wearing-Dataset`: according to the bbox label data, the image is cropped by enlarging width and height of bbox by 3 times. The label is 0 if wearing safety helmet in the image, and the label is 1 if not;
*`hard-hat-detection`: Only use the image that labeled "hat" and crop it with bbox. The label is 0;
*`CelebA`: Only use the image labeled "wearing_hat" and crop it with bbox. The label is 0;
After processing, the dataset totals about 150000 images, of which the number of images with and without wearing safety helmet is about 28000 and 121000 respectively. Then 5600 images are randomly selected in the two labels as the valuation data, a total of about 11200 images, and about 138000 other images as the training data.
Some image of the processed dataset is as follows:
The `train_list.txt` and `val_list.txt` are label files of training data and validation data respectively. All images in `images/` directory.
**Note**:
* About the contents format of `train_list.txt` and `val_list.txt`, please refer to [Description about Classification Dataset in PaddleClas](../data_preparation/classification_dataset_en.md).
<aname="3.3"></a>
### 3.3 Training
The details of training config in `ppcls/configs/PULC/person_exists/PPLCNet_x1_0.yaml`. The command about training as follows:
The best metric of validation data is between `0.985` and `0.993`. There would be fluctuations because the data size is small.
**Note**:
* The metric Tpr, that describe the True Positive Rate when False Positive Rate is less than a certain threshold(1/10000 used in this case), is one of the commonly used metric for binary classification. About the details of Fpr and Tpr, please refer [here](https://en.wikipedia.org/wiki/Receiver_operating_characteristic).
* When evaluation, the best metric TprAtFpr will be printed that include `Fpr`, `Tpr` and the current `threshold`. The `Tpr` means the Recall rate under the current `Fpr`. The `Tpr` higher, the model better. The `threshold` would be used in deployment, which means the classification threshold under best `Fpr` metric.
<aname="3.4"></a>
### 3.4 Evaluation
After training, you can use the following commands to evaluate the model.
Among the above command, the argument `-o Global.pretrained_model="output/PPLCNet_x1_0/best_model"` specify the path of the best model weight file. You can specify other path if needed.
<aname="3.5"></a>
### 3.5 Inference
After training, you can use the model that trained to infer. Command is as follow:
* Among the above command, argument `-o Global.pretrained_model="output/PPLCNet_x1_0/best_model"` specify the path of the best model weight file. You can specify other path if needed.
* The default test image is `deploy/images/PULC/safety_helmet/safety_helmet_test_1.png`. And you can test other image, only need to specify the argument `-o Infer.infer_imgs=path_to_test_image`.
* The default threshold is `0.5`. If needed, you can specify the argument `Infer.PostProcess.threshold`, such as: `-o Infer.PostProcess.threshold=0.9167`. And the argument `threshold` is needed to be specified according by specific case. The `0.9167` is the best threshold when `Fpr` is less than `1/10000` in this valuation dataset.
<aname="4"></a>
## 4. Model Compression
<aname="4.1"></a>
### 4.1 UDML Knowledge Distillation
UDML is a simple but effective knowledge distillation algrithem proposed by PaddleClas. Please refer to [UDML 知识蒸馏](../advanced_tutorials/knowledge_distillation_en.md#1.2.3) for more details.
<aname="4.1.1"></a>
#### 4.1.1 Knowledge Distillation Training
Training with hyperparameters specified in `ppcls/configs/PULC/safety_helmet/PPLCNet_x1_0_distillation.yaml`. The command is as follow:
The best metric is between `0.990` and `0.993`. The best student model weight would be saved in file `output/DistillationModel/best_model_student.pdparams`.
<aname="5"></a>
## 5. Hyperparameters Searching
The hyperparameters used by [3.2 section](#3.2) and [4.1 section](#4.1) are according by `Hyperparameters Searching` in PaddleClas. If you want to get better results on your own dataset, you can refer to [Hyperparameters Searching](PULC_train_en.md#4) to get better hyperparameters.
**Note**: This section is optional. Because the search process will take a long time, you can selectively run according to your specific. If not replace the dataset, you can ignore this section.
<aname="6"></a>
## 6. Inference Deployment
<aname="6.1"></a>
### 6.1 Getting Paddle Inference Model
Paddle Inference is the original Inference Library of the PaddlePaddle, provides high-performance inference for server deployment. And compared with directly based on the pretrained model, Paddle Inference can use tools to accelerate prediction, so as to achieve better inference performance. Please refer to [Paddle Inference](https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/infer/inference/inference_cn.html) for more information.
Paddle Inference need Paddle Inference Model to predict. Two process provided to get Paddle Inference Model. If want to use the provided by PaddleClas, you can download directly, click [Downloading Inference Model](#6.1.2).
<aname="6.1.1"></a>
### 6.1.1 Exporting Paddle Inference Model
The command about exporting Paddle Inference Model is as follow:
After running above command, the inference model files would be saved in `deploy/models/PPLCNet_x1_0_safety_helmet_infer`, as shown below:
```
├── PPLCNet_x1_0_safety_helmet_infer
│ ├── inference.pdiparams
│ ├── inference.pdiparams.info
│ └── inference.pdmodel
```
**Note**: The best model is from knowledge distillation training. If knowledge distillation training is not used, the best model would be saved in `output/PPLCNet_x1_0/best_model.pdparams`.
<aname="6.1.2"></a>
### 6.1.2 Downloading Inference Model
You can also download directly.
```
cd deploy/models
# download the inference model and decompression
wget https://paddleclas.bj.bcebos.com/models/PULC/safety_helmet_infer.tar && tar -xf safety_helmet_infer.tar
```
After decompression, the directory `models` should be shown below.
```
├── safety_helmet_infer
│ ├── inference.pdiparams
│ ├── inference.pdiparams.info
│ └── inference.pdmodel
```
<aname="6.2"></a>
### 6.2 Prediction with Python
<aname="6.2.1"></a>
#### 6.2.1 Image Prediction
Return the directory `deploy`:
```
cd ../
```
Run the following command to classify whether wearing safety helmet about the image `./images/PULC/safety_helmet/safety_helmet_test_1.png`.
safety_helmet_test_1.png: class id(s): [1], score(s): [1.00], label_name(s): ['unwearing_helmet']
```
**Note**: The default threshold is `0.5`. If needed, you can specify the argument `Infer.PostProcess.threshold`, such as: `-o Infer.PostProcess.threshold=0.9167`. And the argument `threshold` is needed to be specified according by specific case. The `0.9167` is the best threshold when `Fpr` is less than `1/10000` in this valuation dataset. Please refer to [3.3 section](#3.3) for details.
<aname="6.2.2"></a>
#### 6.2.2 Images Prediction
If you want to predict images in directory, please specify the argument `Global.infer_imgs` as directory path by `-o Global.infer_imgs`. The command is as follow.
```shell
# Use the following command to predict with GPU. If want to replace with CPU, you can add argument -o Global.use_gpu=False
All prediction results will be printed, as shown below.
```
safety_helmet_test_1.png: class id(s): [1], score(s): [1.00], label_name(s): ['unwearing_helmet']
safety_helmet_test_2.png: class id(s): [0], score(s): [1.00], label_name(s): ['wearing_helmet']
```
Among the prediction results above, `wearing_helmet` means that wearing safety helmet about the image, `unwearing_helmet` means not.
<aname="6.3"></a>
### 6.3 Deployment with C++
PaddleClas provides an example about how to deploy with C++. Please refer to [Deployment with C++](../inference_deployment/cpp_deploy_en.md).
<aname="6.4"></a>
### 6.4 Deployment as Service
Paddle Serving is a flexible, high-performance carrier for machine learning models, and supports different protocol, such as RESTful, gRPC, bRPC and so on, which provides different deployment solutions for a variety of heterogeneous hardware and operating system environments. Please refer [Paddle Serving](https://github.com/PaddlePaddle/Serving) for more information.
PaddleClas provides an example about how to deploy as service by Paddle Serving. Please refer to [Paddle Serving Deployment](../inference_deployment/paddle_serving_deploy_en.md).
<aname="6.5"></a>
### 6.5 Deployment on Mobile
Paddle-Lite is an open source deep learning framework that designed to make easy to perform inference on mobile, embeded, and IoT devices. Please refer to [Paddle-Lite](https://github.com/PaddlePaddle/Paddle-Lite) for more information.
PaddleClas provides an example of how to deploy on mobile by Paddle-Lite. Please refer to [Paddle-Lite deployment](../inference_deployment/paddle_lite_deploy_en.md).
<aname="6.6"></a>
### 6.6 Converting To ONNX and Deployment
Paddle2ONNX support convert Paddle Inference model to ONNX model. And you can deploy with ONNX model on different inference engine, such as TensorRT, OpenVINO, MNN/TNN, NCNN and so on. About Paddle2ONNX details, please refer to [Paddle2ONNX](https://github.com/PaddlePaddle/Paddle2ONNX).
PaddleClas provides an example of how to convert Paddle Inference model to ONNX model by paddle2onnx toolkit and predict by ONNX model. You can refer to [paddle2onnx](../../../deploy/paddle2onnx/readme_en.md) for deployment details.
@@ -48,7 +48,7 @@ The following table lists the relevant indicators of the model. The first two li
| PPLCNet_x1_0 | 98.02 | 2.16 | 6.5 | using SSLD pretrained model |
| **PPLCNet_x1_0** | **99.06** | **2.16** | **6.5** | using SSLD pretrained model + hyperparameters searching strategy |
It can be seen that high Tpr can be getted when backbone is SwinTranformer_tiny, but the speed is slow. Replacing backbone with the lightweight model MobileNetV3_small_x0_35, the speed can be greatly improved, but the Tpr will be greatly reduced. Replacing backbone with faster backbone PPLCNet_x1_0, the Tpr is higher more 14 percentage points than MobileNetv3_small_x0_35. At the same time, the speed can be more faster. After additional using the SSLD pretrained model, the accuracy can be improved by about 0.17 percentage points without affecting the inference speed. Finally, after additional using the hyperparameters searching strategy, the accuracy can be further improved by 1.04 percentage points. At this point, the Tpr close to that of SwinTranformer_tiny, but the speed is more faster. The training method and deployment instructions of PULC will be introduced in detail below.
It can be seen that high accuracy can be getted when backbone is SwinTranformer_tiny, but the speed is slow. Replacing backbone with the lightweight model MobileNetV3_small_x0_35, the speed can be greatly improved, but the accuracy will be greatly reduced. Replacing backbone with faster backbone PPLCNet_x1_0, the accuracy is higher more 14 percentage points than MobileNetv3_small_x0_35. At the same time, the speed can be more faster. After additional using the SSLD pretrained model, the accuracy can be improved by about 0.17 percentage points without affecting the inference speed. Finally, after additional using the hyperparameters searching strategy, the accuracy can be further improved by 1.04 percentage points. At this point, the accuracy is close to that of SwinTranformer_tiny, but the speed is more faster. The training method and deployment instructions of PULC will be introduced in detail below.
**Note**:
...
...
@@ -161,7 +161,6 @@ Some image of the processed dataset is as follows:
And you can also download the data processed directly.
```
cd path_to_PaddleClas
```
...
...
@@ -307,7 +306,7 @@ The best metric of validation data is about `0.996`. The best teacher model weig
#### 4.1.2 Knowledge Distillation Training
The training strategy, specified in training config file `ppcls/configs/PULC/text_image_orientation/PPLCNet_x1_0_distillation.yaml`, the teacher model is `ResNet101_vd`, the student model is `PPLCNet_x1_0` and the additional unlabeled training data is validation data of ImageNet1k.
The training strategy, specified in training config file `ppcls/configs/PULC/text_image_orientation/PPLCNet_x1_0_distillation.yaml`, the teacher model is `ResNet101_vd` and the student model is `PPLCNet_x1_0`.
# PULC Classification Model of Textline Orientation
------
## Catalogue
-[1. Introduction](#1)
-[2. Quick Start](#2)
-[2.1 PaddlePaddle Installation](#2.1)
-[2.2 PaddleClas Installation](#2.2)
-[2.3 Prediction](#2.3)
-[3. Training, Evaluation and Inference](#3)
-[3.1 Installation](#3.1)
-[3.2 Dataset](#3.2)
-[3.2.1 Dataset Introduction](#3.2.1)
-[3.2.2 Getting Dataset](#3.2.2)
-[3.3 Training](#3.3)
-[3.4 Evaluation](#3.4)
-[3.5 Inference](#3.5)
-[4. Model Compression](#4)
-[4.1 SKL-UGI Knowledge Distillation](#4.1)
-[4.1.1 Teacher Model Training](#4.1.1)
-[4.1.2 Knowledge Distillation Training](#4.1.2)
-[5. SHAS](#5)
-[6. Inference Deployment](#6)
-[6.1 Getting Paddle Inference Model](#6.1)
-[6.1.1 Exporting Paddle Inference Model](#6.1.1)
-[6.1.2 Downloading Inference Model](#6.1.2)
-[6.2 Prediction with Python](#6.2)
-[6.2.1 Image Prediction](#6.2.1)
-[6.2.2 Images Prediction](#6.2.2)
-[6.3 Deployment with C++](#6.3)
-[6.4 Deployment as Service](#6.4)
-[6.5 Deployment on Mobile](#6.5)
-[6.6 Converting To ONNX and Deployment](#6.6)
<aname="1"></a>
## 1. Introduction
This case provides a way for users to quickly build a lightweight, high-precision and practical classification model of textline orientation using PaddleClas PULC (Practical Ultra Lightweight Classification). The model can be widely used in character correction, character recognition, etc.
The following table lists the relevant indicators of the model. The first two lines means that using SwinTransformer_tiny and MobileNetV3_small_x0_35 as the backbone to training. The third to seventh lines means that the backbone is replaced by PPLCNet, additional use of EDA strategy and additional use of EDA strategy and SKL-UGI knowledge distillation strategy.
It can be seen that high accuracy can be getted when backbone is SwinTranformer_tiny, but the speed is slow. Replacing backbone with the lightweight model MobileNetV3_small_x0_35, the speed can be greatly improved, but the accuracy will be greatly reduced. Replacing backbone with faster backbone PPLCNet_x1_0, the accuracy is higher more 8.6 percentage points than MobileNetv3_small_x0_35. At the same time, the speed can be more than 10% faster. On this basis, by changing the resolution and stripe (refer to [PaddleOCR](https://github.com/PaddlePaddle/PaddleOCR)), the speed becomes 27% slower, but the accuracy can be improved by 4.5 percentage points. After additional using the SSLD pretrained model, the accuracy can be improved by about 0.05 percentage points without affecting the inference speed. Finally, additional using the EDA strategy, the accuracy can be increased by 1.9 percentage points. The training method and deployment instructions of PULC will be introduced in detail below.
**Note**:
* Backbone name without \* means the resolution is 224x224, and with \* means the resolution is 48x192 (h\*w). The stride of the network is changed to `[2, [2, 1], [2, 1], [2, 1]`. Please refer to [PaddleOCR]（ https://github.com/PaddlePaddle/PaddleOCR）for more details.
* Backbone name with \*\* means that the resolution is 80x160 (h\*w), and the stride of the network is changed to `[2, [2, 1], [2, 1], [2, 1]]`. This resolution is searched by [Hyperparameter Searching](pulc_train_en.md#4).
* The Latency is tested on Intel(R) Xeon(R) Gold 6148 CPU @ 2.40GHz. The MKLDNN is enabled and the number of threads is 10.
* About PP-LCNet, please refer to [PP-LCNet Introduction](../models/PP-LCNet_en.md) and [PP-LCNet Paper](https://arxiv.org/abs/2109.15099).
<aname="2"></a>
## 2. Quick Start
<aname="2.1"></a>
### 2.1 PaddlePaddle Installation
- Run the following command to install if CUDA9 or CUDA10 is available.
Please refer to [PaddlePaddle Installation](https://www.paddlepaddle.org.cn/install/quick?docurl=/documentation/docs/en/install/pip/linux-pip_en.html) for more information about installation, for examples other versions.
<aname="2.2"></a>
### 2.2 PaddleClas wheel Installation
The command of PaddleClas installation as bellow:
```bash
pip3 install paddleclas
```
<aname="2.3"></a>
### 2.3 Prediction
First, please click [here](https://paddleclas.bj.bcebos.com/data/PULC/pulc_demo_imgs.zip) to download and unzip to get the test demo images.
**Note**: The `result` returned by `model.predict()` is a generator, so you need to use the `next()` function to call it or `for` loop to loop it. And it will predict with `batch_size` size batch and return the prediction results when called. The default `batch_size` is 1, and you also specify the `batch_size` when instantiating, such as `model = paddleclas.PaddleClas(model_name="textline_orientation", batch_size=2)`. The result of demo above:
Please refer to [Installation](../installation/install_paddleclas_en.md) to get the description about installation.
<aname="3.2"></a>
### 3.2 Dataset
<aname="3.2.1"></a>
#### 3.2.1 Dataset Introduction
The data used in this case come from internal data. If you want to experience the training process, you can use open source data, such as [ICDAR2019-LSVT](https://aistudio.baidu.com/aistudio/datasetdetail/8429).
<aname="3.2.2"></a>
#### 3.2.2 Getting Dataset
Take ICDAR2019-LSVT for example, images with ID numbers from 0 to 1999 would be processed and used. After rotation, it is divided into class 0 or class 1. Class 0 means that the textline rotation angle is 0 degrees, and class 1 means 180 degrees.
- Training data: The images with ID number from 0 to 1799 are used as the training set. 3600 images in total.
- Evaluation data: The images with ID number from 1800 to 1999 are used as the evaluation set. 400 images in total.
Some image of the processed dataset is as follows:
Where `0/` and `1/` are class 0 and class 1 data respectively. The `train_list.txt` and `val_list.txt` are label files of training data and validation data respectively.
**Note**:
* About the contents format of `train_list.txt` and `val_list.txt`, please refer to [Description about Classification Dataset in PaddleClas](../data_preparation/classification_dataset_en.md).
<aname="3.3"></a>
### 3.3 Training
The details of training config in `ppcls/configs/PULC/textline_orientation/PPLCNet_x1_0.yaml`. The command about training as follows:
* Because the ICDAR2019-LSVT data set is different from the dataset used in the provided pretrained model. If you want to get higher accuracy, you can process [ICDAR2019-LSVT](https://aistudio.baidu.com/aistudio/datasetdetail/8429).
<aname="3.4"></a>
### 3.4 Evaluation
After training, you can use the following commands to evaluate the model.
Among the above command, the argument `-o Global.pretrained_model="output/PPLCNet_x1_0/best_model"` specify the path of the best model weight file. You can specify other path if needed.
<aname="3.5"></a>
### 3.5 Inference
After training, you can use the model that trained to infer. Command is as follow:
* Among the above command, argument `-o Global.pretrained_model="output/PPLCNet_x1_0/best_model"` specify the path of the best model weight file. You can specify other path if needed.
* The default test image is `deploy/images/PULC/textline_orientation/textline_orientation_test_0_0.png`. And you can test other image, only need to specify the argument `-o Infer.infer_imgs=path_to_test_image`.
<aname="4"></a>
## 4. Model Compression
<aname="4.1"></a>
### 4.1 SKL-UGI Knowledge Distillation
SKL-UGI is a simple but effective knowledge distillation algrithem proposed by PaddleClas.
<!-- todo -->
<!-- Please refer to [SKL-UGI](../advanced_tutorials/distillation/distillation_en.md) for more details. -->
<aname="4.1.1"></a>
#### 4.1.1 Teacher Model Training
Training the teacher model with hyperparameters specified in `ppcls/configs/PULC/textline_orientation/PPLCNet/PPLCNet_x1_0.yaml`. The command is as follow:
The best metric of validation data is between `0.96` and `0.98`. The best teacher model weight would be saved in file `output/ResNet101_vd/best_model.pdparams`.
<aname="4.1.2"></a>
#### 4.1.2 Knowledge Distillation Training
The training strategy, specified in training config file `ppcls/configs/PULC/textline_orientation/PPLCNet_x1_0_distillation.yaml`, the teacher model is `ResNet101_vd` and the student model is `PPLCNet_x1_0`. The command is as follow:
The best metric is between `0.95` and `0.97`. The best student model weight would be saved in file `output/DistillationModel/best_model_student.pdparams`.
<aname="5"></a>
## 5. Hyperparameters Searching
The hyperparameters used by [3.2 section](#3.2) and [4.1 section](#4.1) are according by `Hyperparameters Searching` in PaddleClas. If you want to get better results on your own dataset, you can refer to [Hyperparameters Searching](PULC_train_en.md#4) to get better hyperparameters.
**Note**: This section is optional. Because the search process will take a long time, you can selectively run according to your specific. If not replace the dataset, you can ignore this section.
<aname="6"></a>
## 6. Inference Deployment
<aname="6.1"></a>
### 6.1 Getting Paddle Inference Model
Paddle Inference is the original Inference Library of the PaddlePaddle, provides high-performance inference for server deployment. And compared with directly based on the pretrained model, Paddle Inference can use tools to accelerate prediction, so as to achieve better inference performance. Please refer to [Paddle Inference](https://www.paddlepaddle.org.cn/documentation/docs/zh/guides/infer/inference/inference_cn.html) for more information.
Paddle Inference need Paddle Inference Model to predict. Two process provided to get Paddle Inference Model. If want to use the provided by PaddleClas, you can download directly, click [Downloading Inference Model](#6.1.2).
<aname="6.1.1"></a>
### 6.1.1 Exporting Paddle Inference Model
The command about exporting Paddle Inference Model is as follow:
After running above command, the inference model files would be saved in `deploy/models/PPLCNet_x1_0_textline_orientation_infer`, as shown below:
```
├── PPLCNet_x1_0_textline_orientation_infer
│ ├── inference.pdiparams
│ ├── inference.pdiparams.info
│ └── inference.pdmodel
```
**Note**: The best model is from knowledge distillation training. If knowledge distillation training is not used, the best model would be saved in `output/PPLCNet_x1_0/best_model.pdparams`.
<aname="6.1.2"></a>
### 6.1.2 Downloading Inference Model
You can also download directly.
```
cd deploy/models
# 下载 inference 模型并解压
wget https://paddleclas.bj.bcebos.com/models/PULC/textline_orientation_infer.tar && tar -xf textline_orientation_infer.tar
```
After decompression, the directory `models` should be shown below.
```
├── textline_orientation_infer
│ ├── inference.pdiparams
│ ├── inference.pdiparams.info
│ └── inference.pdmodel
```
<aname="6.2"></a>
### 6.2 Prediction with Python
<aname="6.2.1"></a>
#### 6.2.1 Image Prediction
Return the directory `deploy`:
```
cd ../
```
Run the following command to classify the rotation of image `./images/PULC/textline_orientation/objects365_02035329.jpg`.
textline_orientation_test_0_0.png: class id(s): [0], score(s): [1.00], label_name(s): ['0_degree']
```
<aname="6.2.2"></a>
#### 6.2.2 Images Prediction
If you want to predict images in directory, please specify the argument `Global.infer_imgs` as directory path by `-o Global.infer_imgs`. The command is as follow.
```shell
# Use the following command to predict with GPU. If want to replace with CPU, you can add argument -o Global.use_gpu=False
All prediction results will be printed, as shown below.
```
textline_orientation_test_0_0.png: class id(s): [0], score(s): [1.00], label_name(s): ['0_degree']
textline_orientation_test_0_1.png: class id(s): [0], score(s): [1.00], label_name(s): ['0_degree']
textline_orientation_test_1_0.png: class id(s): [1], score(s): [1.00], label_name(s): ['180_degree']
textline_orientation_test_1_1.png: class id(s): [1], score(s): [1.00], label_name(s): ['180_degree']
```
Among the prediction results above, `0_degree` means that the rotation angle of the textline image is 0, and `180_degree` means that 180.
<aname="6.3"></a>
### 6.3 Deployment with C++
PaddleClas provides an example about how to deploy with C++. Please refer to [Deployment with C++](../inference_deployment/cpp_deploy_en.md).
<aname="6.4"></a>
### 6.4 Deployment as Service
Paddle Serving is a flexible, high-performance carrier for machine learning models, and supports different protocol, such as RESTful, gRPC, bRPC and so on, which provides different deployment solutions for a variety of heterogeneous hardware and operating system environments. Please refer [Paddle Serving](https://github.com/PaddlePaddle/Serving) for more information.
PaddleClas provides an example about how to deploy as service by Paddle Serving. Please refer to [Paddle Serving Deployment](../inference_deployment/paddle_serving_deploy_en.md).
<aname="6.5"></a>
### 6.5 Deployment on Mobile
Paddle-Lite is an open source deep learning framework that designed to make easy to perform inference on mobile, embeded, and IoT devices. Please refer to [Paddle-Lite](https://github.com/PaddlePaddle/Paddle-Lite) for more information.
PaddleClas provides an example of how to deploy on mobile by Paddle-Lite. Please refer to [Paddle-Lite deployment](../inference_deployment/paddle_lite_deploy_en.md).
<aname="6.6"></a>
### 6.6 Converting To ONNX and Deployment
Paddle2ONNX support convert Paddle Inference model to ONNX model. And you can deploy with ONNX model on different inference engine, such as TensorRT, OpenVINO, MNN/TNN, NCNN and so on. About Paddle2ONNX details, please refer to [Paddle2ONNX](https://github.com/PaddlePaddle/Paddle2ONNX).
PaddleClas provides an example of how to convert Paddle Inference model to ONNX model by paddle2onnx toolkit and predict by ONNX model. You can refer to [paddle2onnx](../../../deploy/paddle2onnx/readme_en.md) for deployment details.
