@@ -43,13 +43,12 @@ After the quantization strategy is defined, the model can be quantified.
The code for quantization training is located in `slim/quantization/quant.py`. For example, to train a detection model, the training instructions are as follows:
Refer to [DTRB](https://arxiv.org/abs/1904.01906), the training and evaluation result of these above text recognition (using MJSynth and SynthText for training, evaluate on IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE) is as follow:
...
...
@@ -60,5 +61,6 @@ Refer to [DTRB](https://arxiv.org/abs/1904.01906), the training and evaluation r
Please refer to the document for training guide and use of PaddleOCR text recognition algorithms [Text recognition model training/evaluation/prediction](./recognition_en.md)
| label_file_list | Groundtruth file path | ["./train_data/train_list.txt"] | This parameter is not required when dataset is LMDBDateSet |
| label_file_list | Groundtruth file path | ["./train_data/train_list.txt"] | This parameter is not required when dataset is LMDBDataSet |
| ratio_list | Ratio of data set | [1.0] | If there are two train_lists in label_file_list and ratio_list is [0.4,0.6], 40% will be sampled from train_list1, and 60% will be sampled from train_list2 to combine the entire dataset |
| transforms | List of methods to transform images and labels | [DecodeImage,CTCLabelEncode,RecResizeImg,KeepKeys] | see[ppocr/data/imaug](../../ppocr/data/imaug) |
@@ -38,28 +38,17 @@ If you want to train PaddleOCR on other datasets, please build the annotation fi
## TRAINING
First download the pretrained model. The detection model of PaddleOCR currently supports 3 backbones, namely MobileNetV3, ResNet18_vd and ResNet50_vd. You can use the model in [PaddleClas](https://github.com/PaddlePaddle/PaddleClas/tree/master/ppcls/modeling/architectures) to replace backbone according to your needs.
First download the pretrained model. The detection model of PaddleOCR currently supports 3 backbones, namely MobileNetV3, ResNet18_vd and ResNet50_vd. You can use the model in [PaddleClas](https://github.com/PaddlePaddle/PaddleClas/tree/develop/ppcls/modeling/architectures) to replace backbone according to your needs.
And the responding download link of backbone pretrain weights can be found in [PaddleClas repo](https://github.com/PaddlePaddle/PaddleClas#mobile-series).
The high performance of distributed training is one of the core advantages of PaddlePaddle. In the classification task, distributed training can achieve almost linear speedup ratio. Generally, OCR training task need massive training data. Such as recognition, ppocrv2.0 model is trained based on 1800W dataset, which is very time-consuming if using single machine. Therefore, the distributed training is used in paddleocr to speedup the training task. For more information about distributed training, please refer to [distributed training quick start tutorial](https://fleet-x.readthedocs.io/en/latest/paddle_fleet_rst/parameter_server/ps_quick_start.html).
## Quick Start
### Training with single machine
Take recognition as an example. After the data is prepared locally, start the training task with the interface of `paddle.distributed.launch`. The start command as follows:
```shell
python3 -m paddle.distributed.launch \
--log_dir=./log/ \
--gpus'0,1,2,3,4,5,6,7'\
tools/train.py \
-c configs/rec/rec_mv3_none_bilstm_ctc.yml
```
### Training with multi machine
Compared with single machine, training with multi machine only needs to add the parameter `--ips` to start command, which represents the IP list of machines used for distributed training, and the IP of different machines are separated by commas. The start command as follows:
```shell
ip_list="192.168.0.1,192.168.0.2"
python3 -m paddle.distributed.launch \
--log_dir=./log/ \
--ips="${ip_list}"\
--gpus="0,1,2,3,4,5,6,7"\
tools/train.py \
-c configs/rec/rec_mv3_none_bilstm_ctc.yml
```
**Notice:**
* The IP addresses of different machines need to be separated by commas, which can be queried through `ifconfig` or `ipconfig`.
* Different machines need to be set to be secret free and can `ping` success with others directly, otherwise communication cannot establish between them.
* The code, data and start command betweent different machines must be completely consistent and then all machines need to run start command. The first machine in the `ip_list` is set to `trainer0`, and so on.
## Performance comparison
* Based on 26W public recognition dataset (LSVT, rctw, mtwi), training on single 8-card P40 and dual 8-card P40, the final time consumption is as follows.
| Model | Config file | Number of machines | Number of GPUs per machine | Training time | Recognition acc | Speedup ratio |
It can be seen that the training time is shortened from 60h to 40h, the speedup ratio can reach 150% (60h / 40h), and the efficiency is 75% (60h / (40h * 2)).
When converting to an inference model, the configuration file used is the same as the configuration file used during training. In addition, you also need to set the `Global.pretrained_model` parameter in the configuration file.
...
...
@@ -80,10 +79,9 @@ The recognition model is converted to the inference model in the same way as the
# -c Set the training algorithm yml configuration file
# -o Set optional parameters
# Global.pretrained_model parameter Set the training model address to be converted without adding the file suffix .pdmodel, .pdopt or .pdparams.
# Global.load_static_weights needs to be set to False
# Global.save_inference_dir Set the address where the converted model will be saved.
If you have a model trained on your own dataset with a different dictionary file, please make sure that you modify the `character_dict_path` in the configuration file to your dictionary file path.
...
...
@@ -109,10 +107,9 @@ The angle classification model is converted to the inference model in the same w
# -c Set the training algorithm yml configuration file
# -o Set optional parameters
# Global.pretrained_model parameter Set the training model address to be converted without adding the file suffix .pdmodel, .pdopt or .pdparams.
# Global.load_static_weights needs to be set to False
# Global.save_inference_dir Set the address where the converted model will be saved.
After the conversion is successful, there are two files in the directory:
...
...
@@ -157,12 +154,12 @@ Set as `limit_type='min', det_limit_side_len=960`, it means that the shortest si
If the resolution of the input picture is relatively large and you want to use a larger resolution prediction, you can set det_limit_side_len to the desired value, such as 1216:
First, convert the model saved in the DB text detection training process into an inference model. Taking the model based on the Resnet50_vd backbone network and trained on the ICDAR2015 English dataset as an example ([model download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_r50_vd_db_v2.0_train.tar)), you can use the following command to convert:
DB text detection model inference, you can execute the following command:
...
...
@@ -192,7 +189,7 @@ The visualized text detection results are saved to the `./inference_results` fol
First, convert the model saved in the EAST text detection training process into an inference model. Taking the model based on the Resnet50_vd backbone network and trained on the ICDAR2015 English dataset as an example ([model download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_r50_vd_east_v2.0_train.tar)), you can use the following command to convert:
**For EAST text detection model inference, you need to set the parameter ``--det_algorithm="EAST"``**, run the following command:
...
...
@@ -213,7 +210,7 @@ The visualized text detection results are saved to the `./inference_results` fol
First, convert the model saved in the SAST text detection training process into an inference model. Taking the model based on the Resnet50_vd backbone network and trained on the ICDAR2015 English dataset as an example ([model download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_r50_vd_sast_icdar15_v2.0_train.tar)), you can use the following command to convert:
**For SAST quadrangle text detection model inference, you need to set the parameter `--det_algorithm="SAST"`**, run the following command:
...
...
@@ -230,10 +227,10 @@ The visualized text detection results are saved to the `./inference_results` fol
First, convert the model saved in the SAST text detection training process into an inference model. Taking the model based on the Resnet50_vd backbone network and trained on the Total-Text English dataset as an example ([model download link](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/det_r50_vd_sast_totaltext_v2.0_train.tar)), you can use the following command to convert:
**For SAST curved text detection model inference, you need to set the parameter `--det_algorithm="SAST"` and `--det_sast_polygon=True`**, run the following command:
For SAST curved text detection model inference, you need to set the parameter `--det_algorithm="SAST"` and `--det_sast_polygon=True`, run the following command:
@@ -279,7 +276,7 @@ Taking CRNN as an example, we introduce the recognition model inference based on
First, convert the model saved in the CRNN text recognition training process into an inference model. Taking the model based on Resnet34_vd backbone network, using MJSynth and SynthText (two English text recognition synthetic datasets) for training, as an example ([model download address](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_r34_vd_none_bilstm_ctc_v2.0_train.tar)). It can be converted as follow:
For CRNN text recognition model inference, execute the following commands:
...
...
@@ -379,14 +376,18 @@ After executing the command, the prediction results (classification angle and sc
<aname="LIGHTWEIGHT_CHINESE_MODEL"></a>
### 1. LIGHTWEIGHT CHINESE MODEL
When performing prediction, you need to specify the path of a single image or a folder of images through the parameter `image_dir`, the parameter `det_model_dir` specifies the path to detect the inference model, the parameter `cls_model_dir` specifies the path to angle classification inference model and the parameter `rec_model_dir` specifies the path to identify the inference model. The parameter `use_angle_cls` is used to control whether to enable the angle classification model.The visualized recognition results are saved to the `./inference_results` folder by default.
When performing prediction, you need to specify the path of a single image or a folder of images through the parameter `image_dir`, the parameter `det_model_dir` specifies the path to detect the inference model, the parameter `cls_model_dir` specifies the path to angle classification inference model and the parameter `rec_model_dir` specifies the path to identify the inference model. The parameter `use_angle_cls` is used to control whether to enable the angle classification model. The parameter `use_mp` specifies whether to use multi-process to infer `total_process_num` specifies process number when using multi-process. The parameter . The visualized recognition results are saved to the `./inference_results` folder by default.
