To prepare datasets, refer to ocr_datasets .
PaddleOCR provides label files for training the icdar2015 dataset, which can be downloaded in the following ways:
# Training set label
wget -P ./train_data/ic15_data https://paddleocr.bj.bcebos.com/dataset/rec_gt_train.txt
# Test Set Label
wget -P ./train_data/ic15_data https://paddleocr.bj.bcebos.com/dataset/rec_gt_test.txt
PaddleOCR also provides a data format conversion script, which can convert ICDAR official website label to a data format
supported by PaddleOCR. The data conversion tool is in ppocr/utils/gen_label.py, here is the training set as an example:
# convert the official gt to rec_gt_label.txt
python gen_label.py --mode="rec" --input_path="{path/of/origin/label}" --output_label="rec_gt_label.txt"
The data format is as follows, (a) is the original picture, (b) is the Ground Truth text file corresponding to each picture:
- Multilingual dataset
The multi-language model training method is the same as the Chinese model. The training data set is 100w synthetic data. A small amount of fonts and test data can be downloaded using the following two methods.
- Baidu Netdisk ,Extraction code:frgi.
- Google drive
Finally, a dictionary ({word_dict_name}.txt) needs to be provided so that when the model is trained, all the characters that appear can be mapped to the dictionary index.
Therefore, the dictionary needs to contain all the characters that you want to be recognized correctly. {word_dict_name}.txt needs to be written in the following format and saved in the utf-8 encoding format:
l
d
a
d
r
n
In word_dict.txt, there is a single word in each line, which maps characters and numeric indexes together, e.g "and" will be mapped to [2 5 1]
PaddleOCR has built-in dictionaries, which can be used on demand.
ppocr/utils/ppocr_keys_v1.txt is a Chinese dictionary with 6623 characters.
ppocr/utils/ic15_dict.txt is an English dictionary with 63 characters
ppocr/utils/dict/french_dict.txt is a French dictionary with 118 characters
ppocr/utils/dict/japan_dict.txt is a Japanese dictionary with 4399 characters
ppocr/utils/dict/korean_dict.txt is a Korean dictionary with 3636 characters
ppocr/utils/dict/german_dict.txt is a German dictionary with 131 characters
ppocr/utils/en_dict.txt is a English dictionary with 96 characters
The current multi-language model is still in the demo stage and will continue to optimize the model and add languages. You are very welcome to provide us with dictionaries and fonts in other languages, If you like, you can submit the dictionary file to dict and we will thank you in the Repo.
To customize the dict file, please modify the character_dict_path field in configs/rec/rec_icdar15_train.yml .
- Custom dictionary
If you need to customize dic file, please add character_dict_path field in configs/rec/rec_icdar15_train.yml to point to your dictionary path. And set character_type to ch.
If you want to support the recognition of the space category, please set the use_space_char field in the yml file to True.
PaddleOCR provides a variety of data augmentation methods. All the augmentation methods are enabled by default.
The default perturbation methods are: cvtColor, blur, jitter, Gasuss noise, random crop, perspective, color reverse, TIA augmentation.
Each disturbance method is selected with a 40% probability during the training process. For specific code implementation, please refer to: rec_img_aug.py
PaddleOCR provides training scripts, evaluation scripts, and prediction scripts. In this section, the CRNN recognition model will be used as an example:
First download the pretrain model, you can download the trained model to finetune on the icdar2015 data:
cd PaddleOCR/
# Download the pre-trained model of en_PP-OCRv3
wget -P ./pretrain_models/ https://paddleocr.bj.bcebos.com/PP-OCRv3/english/en_PP-OCRv3_rec_train.tar
# Decompress model parameters
cd pretrain_models
tar -xf en_PP-OCRv3_rec_train.tar && rm -rf en_PP-OCRv3_rec_train.tar
Start training:
# GPU training Support single card and multi-card training
# Training icdar15 English data and The training log will be automatically saved as train.log under "{save_model_dir}"
#specify the single card training(Long training time, not recommended)
python3 tools/train.py -c configs/rec/PP-OCRv3/en_PP-OCRv3_rec.yml -o Global.pretrained_model=en_PP-OCRv3_rec_train/best_accuracy
#specify the card number through --gpus
python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py -c configs/rec/PP-OCRv3/en_PP-OCRv3_rec.yml -o Global.pretrained_model=en_PP-OCRv3_rec_train/best_accuracy
PaddleOCR supports alternating training and evaluation. You can modify eval_batch_step in configs/rec/rec_icdar15_train.yml to set the evaluation frequency. By default, it is evaluated every 500 iter and the best acc model is saved under output/rec_CRNN/best_accuracy during the evaluation process.
If the evaluation set is large, the test will be time-consuming. It is recommended to reduce the number of evaluations, or evaluate after training.
- Tip: You can use the
-cparameter to select multiple model configurations under theconfigs/rec/path for training. The recognition algorithms supported at rec_algorithm:
For training Chinese data, it is recommended to use ch_PP-OCRv3_rec_distillation.yml. If you want to try the result of other algorithms on the Chinese data set, please refer to the following instructions to modify the configuration file:
Take ch_PP-OCRv3_rec_distillation.yml as an example:
Global:
...
# Add a custom dictionary, such as modify the dictionary, please point the path to the new dictionary
character_dict_path: ppocr/utils/ppocr_keys_v1.txt
# Modify character type
...
# Whether to recognize spaces
use_space_char: True
Optimizer:
...
# Add learning rate decay strategy
lr:
name: Cosine
learning_rate: 0.001
...
...
Train:
dataset:
# Type of dataset，we support LMDBDataSet and SimpleDataSet
name: SimpleDataSet
# Path of dataset
data_dir: ./train_data/
# Path of train list
label_file_list: ["./train_data/train_list.txt"]
transforms:
...
- RecResizeImg:
# Modify image_shape to fit long text
image_shape: [3, 48, 320]
...
loader:
...
# Train batch_size for Single card
batch_size_per_card: 256
...
Eval:
dataset:
# Type of dataset，we support LMDBDataSet and SimpleDataSet
name: SimpleDataSet
# Path of dataset
data_dir: ./train_data
# Path of eval list
label_file_list: ["./train_data/val_list.txt"]
transforms:
...
- RecResizeImg:
# Modify image_shape to fit long text
image_shape: [3, 48, 320]
...
loader:
# Eval batch_size for Single card
batch_size_per_card: 256
...
Note that the configuration file for prediction/evaluation must be consistent with the training.
If you expect to load trained model and continue the training again, you can specify the parameter Global.checkpoints as the model path to be loaded.
For example:
python3 tools/train.py -c configs/rec/rec_icdar15_train.yml -o Global.checkpoints=./your/trained/modelNote: The priority of Global.checkpoints is higher than that of Global.pretrained_model, that is, when two parameters are specified at the same time, the model specified by Global.checkpoints will be loaded first. If the model path specified by Global.checkpoints is wrong, the one specified by Global.pretrained_model will be loaded.
The network part completes the construction of the network, and PaddleOCR divides the network into four parts, which are under ppocr/modeling. The data entering the network will pass through these four parts in sequence(transforms->backbones-> necks->heads).
├── architectures # Code for building network
├── transforms # Image Transformation Module
├── backbones # Feature extraction module
├── necks # Feature enhancement module
└── heads # Output moduleIf the Backbone to be replaced has a corresponding implementation in PaddleOCR, you can directly modify the parameters in the Backbone part of the configuration yml file.
However, if you want to use a new Backbone, an example of replacing the backbones is as follows:
- Create a new file under the ppocr/modeling/backbones folder, such as my_backbone.py.
- Add code in the my_backbone.py file, the sample code is as follows:
import paddle
import paddle.nn as nn
import paddle.nn.functional as F
class MyBackbone(nn.Layer):
def __init__(self, *args, **kwargs):
super(MyBackbone, self).__init__()
# your init code
self.conv = nn.xxxx
def forward(self, inputs):
# your network forward
y = self.conv(inputs)
return y- Import the added module in the ppocr/modeling/backbones/_init_.py file.
After adding the four-part modules of the network, you only need to configure them in the configuration file to use, such as:
Backbone:
name: MyBackbone
args1: args1NOTE: More details about replace Backbone and other mudule can be found in doc.
If you want to speed up your training further, you can use Auto Mixed Precision Training, taking a single machine and a single gpu as an example, the commands are as follows:
python3 tools/train.py -c configs/rec/rec_icdar15_train.yml \
-o Global.pretrained_model=./pretrain_models/rec_mv3_none_bilstm_ctc_v2.0_train \
Global.use_amp=True Global.scale_loss=1024.0 Global.use_dynamic_loss_scaling=TrueDuring multi-machine multi-gpu training, use the --ips parameter to set the used machine IP address, and the --gpus parameter to set the used GPU ID:
python3 -m paddle.distributed.launch --ips="xx.xx.xx.xx,xx.xx.xx.xx" --gpus '0,1,2,3' tools/train.py -c configs/rec/rec_icdar15_train.yml \
-o Global.pretrained_model=./pretrain_models/rec_mv3_none_bilstm_ctc_v2.0_trainNote: (1) When using multi-machine and multi-gpu training, you need to replace the ips value in the above command with the address of your machine, and the machines need to be able to ping each other. (2) Training needs to be launched separately on multiple machines. The command to view the ip address of the machine is ifconfig. (3) For more details about the distributed training speedup ratio, please refer to Distributed Training Tutorial.
Knowledge distillation is supported in PaddleOCR for text recognition training process. For more details, please refer to doc.
Currently, the multi-language algorithms supported by PaddleOCR are:
| Configuration file | Algorithm name | backbone | trans | seq | pred | language |
|---|---|---|---|---|---|---|
| rec_chinese_cht_lite_train.yml | CRNN | Mobilenet_v3 small 0.5 | None | BiLSTM | ctc | chinese traditional |
| rec_en_lite_train.yml | CRNN | Mobilenet_v3 small 0.5 | None | BiLSTM | ctc | English(Case sensitive) |
| rec_french_lite_train.yml | CRNN | Mobilenet_v3 small 0.5 | None | BiLSTM | ctc | French |
| rec_ger_lite_train.yml | CRNN | Mobilenet_v3 small 0.5 | None | BiLSTM | ctc | German |
| rec_japan_lite_train.yml | CRNN | Mobilenet_v3 small 0.5 | None | BiLSTM | ctc | Japanese |
| rec_korean_lite_train.yml | CRNN | Mobilenet_v3 small 0.5 | None | BiLSTM | ctc | Korean |
| rec_latin_lite_train.yml | CRNN | Mobilenet_v3 small 0.5 | None | BiLSTM | ctc | Latin |
| rec_arabic_lite_train.yml | CRNN | Mobilenet_v3 small 0.5 | None | BiLSTM | ctc | arabic |
| rec_cyrillic_lite_train.yml | CRNN | Mobilenet_v3 small 0.5 | None | BiLSTM | ctc | cyrillic |
| rec_devanagari_lite_train.yml | CRNN | Mobilenet_v3 small 0.5 | None | BiLSTM | ctc | devanagari |
For more supported languages, please refer to : Multi-language model
If you want to finetune on the basis of the existing model effect, please refer to the following instructions to modify the configuration file:
Take rec_french_lite_train as an example:
Global:
...
# Add a custom dictionary, such as modify the dictionary, please point the path to the new dictionary
character_dict_path: ./ppocr/utils/dict/french_dict.txt
...
# Whether to recognize spaces
use_space_char: True
...
Train:
dataset:
# Type of dataset，we support LMDBDataSet and SimpleDataSet
name: SimpleDataSet
# Path of dataset
data_dir: ./train_data/
# Path of train list
label_file_list: ["./train_data/french_train.txt"]
...
Eval:
dataset:
# Type of dataset，we support LMDBDataSet and SimpleDataSet
name: SimpleDataSet
# Path of dataset
data_dir: ./train_data
# Path of eval list
label_file_list: ["./train_data/french_val.txt"]
...
-
Windows GPU/CPU The Windows platform is slightly different from the Linux platform: Windows platform only supports
single gputraining and inference, specify GPU for trainingset CUDA_VISIBLE_DEVICES=0On the Windows platform, DataLoader only supports single-process mode, so you need to setnum_workersto 0; -
macOS GPU mode is not supported, you need to set
use_gputo False in the configuration file, and the rest of the training evaluation prediction commands are exactly the same as Linux GPU. -
Linux DCU Running on a DCU device requires setting the environment variable
export HIP_VISIBLE_DEVICES=0,1,2,3, and the rest of the training and evaluation prediction commands are exactly the same as the Linux GPU.
In actual use, it is recommended to load the official pre-trained model and fine-tune it in your own data set. For the fine-tuning method of the recognition model, please refer to: Model Fine-tuning Tutorial.
The model parameters during training are saved in the Global.save_model_dir directory by default. When evaluating indicators, you need to set Global.checkpoints to point to the saved parameter file. The evaluation dataset can be set by modifying the Eval.dataset.label_file_list field in the configs/rec/PP-OCRv3/en_PP-OCRv3_rec.yml file.
# GPU evaluation, Global.checkpoints is the weight to be tested
python3 -m paddle.distributed.launch --gpus '0' tools/eval.py -c configs/rec/PP-OCRv3/en_PP-OCRv3_rec.yml -o Global.checkpoints={path/to/weights}/best_accuracy
Using the model trained by paddleocr, you can quickly get prediction through the following script.
The default prediction picture is stored in infer_img, and the trained weight is specified via -o Global.checkpoints:
According to the save_model_dir and save_epoch_step fields set in the configuration file, the following parameters will be saved:
output/rec/
├── best_accuracy.pdopt
├── best_accuracy.pdparams
├── best_accuracy.states
├── config.yml
├── iter_epoch_3.pdopt
├── iter_epoch_3.pdparams
├── iter_epoch_3.states
├── latest.pdopt
├── latest.pdparams
├── latest.states
└── train.log
Among them, best_accuracy.* is the best model on the evaluation set; iter_epoch_x.* is the model saved at intervals of save_epoch_step; latest.* is the model of the last epoch.
# Predict English results
python3 tools/infer_rec.py -c configs/rec/PP-OCRv3/en_PP-OCRv3_rec.yml -o Global.pretrained_model={path/to/weights}/best_accuracy Global.infer_img=doc/imgs_words/en/word_1.png
Input image:
Get the prediction result of the input image:
infer_img: doc/imgs_words/en/word_1.png
result: ('joint', 0.9998967)
The configuration file used for prediction must be consistent with the training. For example, you completed the training of the Chinese model with python3 tools/train.py -c configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml, you can use the following command to predict the Chinese model:
# Predict Chinese results
python3 tools/infer_rec.py -c configs/rec/ch_ppocr_v2.0/rec_chinese_lite_train_v2.0.yml -o Global.pretrained_model={path/to/weights}/best_accuracy Global.infer_img=doc/imgs_words/ch/word_1.jpg
Input image:
Get the prediction result of the input image:
infer_img: doc/imgs_words/ch/word_1.jpg
result: ('韩国小馆', 0.997218)
The inference model (the model saved by paddle.jit.save) is generally a solidified model saved after the model training is completed, and is mostly used to give prediction in deployment.
The model saved during the training process is the checkpoints model, which saves the parameters of the model and is mostly used to resume training.
Compared with the checkpoints model, the inference model will additionally save the structural information of the model. Therefore, it is easier to deploy because the model structure and model parameters are already solidified in the inference model file, and is suitable for integration with actual systems.
The recognition model is converted to the inference model in the same way as the detection, as follows:
# -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.save_inference_dir Set the address where the converted model will be saved.
python3 tools/export_model.py -c configs/rec/PP-OCRv3/en_PP-OCRv3_rec.yml -o Global.pretrained_model=en_PP-OCRv3_rec_train/best_accuracy Global.save_inference_dir=./inference/en_PP-OCRv3_rec/
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.
After the conversion is successful, there are three files in the model save directory:
inference/en_PP-OCRv3_rec/
├── inference.pdiparams # The parameter file of recognition inference model
├── inference.pdiparams.info # The parameter information of recognition inference model, which can be ignored
└── inference.pdmodel # The program file of recognition model
-
Text recognition model Inference using custom characters dictionary
If the text dictionary is modified during training, when using the inference model to predict, you need to specify the dictionary path used by
--rec_char_dict_pathpython3 tools/infer/predict_rec.py --image_dir="./doc/imgs_words_en/word_336.png" --rec_model_dir="./your inference model" --rec_image_shape="3, 32, 100" --rec_char_dict_path="your text dict path"
Q1: After the training model is transferred to the inference model, the prediction effect is inconsistent?
A: There are many such problems, and the problems are mostly caused by inconsistent preprocessing and postprocessing parameters when the trained model predicts and the preprocessing and postprocessing parameters when the inference model predicts. You can compare whether there are differences in preprocessing, postprocessing, and prediction in the configuration files used for training.