# 场景文本识别算法-NRTR
- [1. 算法简介](#1)
- [2. 环境配置](#2)
- [3. 模型训练、评估、预测](#3)
- [3.1 训练](#3-1)
- [3.2 评估](#3-2)
- [3.3 预测](#3-3)
- [4. 推理部署](#4)
- [4.1 Python推理](#4-1)
- [4.2 C++推理](#4-2)
- [4.3 Serving服务化部署](#4-3)
- [4.4 更多推理部署](#4-4)
- [5. FAQ](#5)
- [6. 发行公告](#6)
## 1. 算法简介
论文信息:
> [NRTR: A No-Recurrence Sequence-to-Sequence Model For Scene Text Recognition](https://arxiv.org/abs/1806.00926)
> Fenfen Sheng and Zhineng Chen and Bo Xu
> ICDAR, 2019
`NRTR`使用MJSynth和SynthText两个文字识别数据集训练,在IIIT, SVT, IC03, IC13, IC15, SVTP, CUTE数据集上进行评估,算法复现效果如下:
|模型|骨干网络|配置文件|Acc|下载链接|
| --- | --- | --- | --- | --- |
|NRTR|MTB|[rec_mtb_nrtr.yml](../../configs/rec/rec_mtb_nrtr.yml)|84.21%|[训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mtb_nrtr_train.tar)|
## 2. 环境配置
请先参考[《运行环境准备》](./environment.md)配置PaddleOCR运行环境,参考[《项目克隆》](./clone.md)克隆项目代码。
## 3. 模型训练、评估、预测
### 3.1 模型训练
请参考[文本识别训练教程](./recognition.md)。PaddleOCR对代码进行了模块化,训练`NRTR`识别模型时需要**更换配置文件**为`NRTR`的[配置文件](../../configs/rec/rec_mtb_nrtr.yml)。
#### 启动训练
具体地,在完成数据准备后,便可以启动训练,训练命令如下:
```shell
#单卡训练(训练周期长,不建议)
python3 tools/train.py -c configs/rec/rec_mtb_nrtr.yml
#多卡训练,通过--gpus参数指定卡号
python3 -m paddle.distributed.launch --gpus '0,1,2,3' tools/train.py -c configs/rec/rec_mtb_nrtr.yml
```
### 3.2 评估
可下载已训练完成的[模型文件](#model),使用如下命令进行评估:
```shell
# 注意将pretrained_model的路径设置为本地路径。
python3 -m paddle.distributed.launch --gpus '0' tools/eval.py -c configs/rec/rec_mtb_nrtr.yml -o Global.pretrained_model=./rec_mtb_nrtr_train/best_accuracy
```
### 3.3 预测
使用如下命令进行单张图片预测:
```shell
# 注意将pretrained_model的路径设置为本地路径。
python3 tools/infer_rec.py -c configs/rec/rec_mtb_nrtr.yml -o Global.infer_img='./doc/imgs_words_en/word_10.png' Global.pretrained_model=./rec_mtb_nrtr_train/best_accuracy
# 预测文件夹下所有图像时,可修改infer_img为文件夹,如 Global.infer_img='./doc/imgs_words_en/'。
```
## 4. 推理部署
### 4.1 Python推理
首先将训练得到best模型,转换成inference model。这里以训练完成的模型为例([模型下载地址](https://paddleocr.bj.bcebos.com/dygraph_v2.0/en/rec_mtb_nrtr_train.tar) ),可以使用如下命令进行转换:
```shell
# 注意将pretrained_model的路径设置为本地路径。
python3 tools/export_model.py -c configs/rec/rec_mtb_nrtr.yml -o Global.pretrained_model=./rec_mtb_nrtr_train/best_accuracy Global.save_inference_dir=./inference/rec_mtb_nrtr/
```
**注意:**
- 如果您是在自己的数据集上训练的模型,并且调整了字典文件,请注意修改配置文件中的`character_dict_path`是否是所需要的字典文件。
- 如果您修改了训练时的输入大小,请修改`tools/export_model.py`文件中的对应NRTR的`infer_shape`。
转换成功后,在目录下有三个文件:
```
/inference/rec_mtb_nrtr/
├── inference.pdiparams # 识别inference模型的参数文件
├── inference.pdiparams.info # 识别inference模型的参数信息,可忽略
└── inference.pdmodel # 识别inference模型的program文件
```
执行如下命令进行模型推理:
```shell
python3 tools/infer/predict_rec.py --image_dir='./doc/imgs_words_en/word_10.png' --rec_model_dir='./inference/rec_mtb_nrtr/' --rec_algorithm='NRTR' --rec_image_shape='1,32,100' --rec_char_dict_path='./ppocr/utils/EN_symbol_dict.txt'
# 预测文件夹下所有图像时,可修改image_dir为文件夹,如 --image_dir='./doc/imgs_words_en/'。
```
![](../imgs_words_en/word_10.png)
执行命令后,上面图像的预测结果(识别的文本和得分)会打印到屏幕上,示例如下:
结果如下:
```shell
Predicts of ./doc/imgs_words_en/word_10.png:('pain', 0.9465042352676392)
```
**注意**:
- 训练上述模型采用的图像分辨率是[1,32,100],需要通过参数`rec_image_shape`设置为您训练时的识别图像形状。
- 在推理时需要设置参数`rec_char_dict_path`指定字典,如果您修改了字典,请修改该参数为您的字典文件。
- 如果您修改了预处理方法,需修改`tools/infer/predict_rec.py`中NRTR的预处理为您的预处理方法。
### 4.2 C++推理部署
由于C++预处理后处理还未支持NRTR,所以暂未支持
### 4.3 Serving服务化部署
暂不支持
### 4.4 更多推理部署
暂不支持
## 5. FAQ
1. `NRTR`论文中使用Beam搜索进行解码字符,但是速度较慢,这里默认未使用Beam搜索,以贪婪搜索进行解码字符。
## 6. 发行公告
1. release/2.6更新NRTR代码结构,新版NRTR可加载旧版(release/2.5及之前)模型参数,使用下面示例代码将旧版模型参数转换为新版模型参数:
```python
params = paddle.load('path/' + '.pdparams') # 旧版本参数
state_dict = model.state_dict() # 新版模型参数
new_state_dict = {}
for k1, v1 in state_dict.items():
k = k1
if 'encoder' in k and 'self_attn' in k and 'qkv' in k and 'weight' in k:
k_para = k[:13] + 'layers.' + k[13:]
q = params[k_para.replace('qkv', 'conv1')].transpose((1, 0, 2, 3))
k = params[k_para.replace('qkv', 'conv2')].transpose((1, 0, 2, 3))
v = params[k_para.replace('qkv', 'conv3')].transpose((1, 0, 2, 3))
new_state_dict[k1] = np.concatenate([q[:, :, 0, 0], k[:, :, 0, 0], v[:, :, 0, 0]], -1)
elif 'encoder' in k and 'self_attn' in k and 'qkv' in k and 'bias' in k:
k_para = k[:13] + 'layers.' + k[13:]
q = params[k_para.replace('qkv', 'conv1')]
k = params[k_para.replace('qkv', 'conv2')]
v = params[k_para.replace('qkv', 'conv3')]
new_state_dict[k1] = np.concatenate([q, k, v], -1)
elif 'encoder' in k and 'self_attn' in k and 'out_proj' in k:
k_para = k[:13] + 'layers.' + k[13:]
new_state_dict[k1] = params[k_para]
elif 'encoder' in k and 'norm3' in k:
k_para = k[:13] + 'layers.' + k[13:]
new_state_dict[k1] = params[k_para.replace('norm3', 'norm2')]
elif 'encoder' in k and 'norm1' in k:
k_para = k[:13] + 'layers.' + k[13:]
new_state_dict[k1] = params[k_para]
elif 'decoder' in k and 'self_attn' in k and 'qkv' in k and 'weight' in k:
k_para = k[:13] + 'layers.' + k[13:]
q = params[k_para.replace('qkv', 'conv1')].transpose((1, 0, 2, 3))
k = params[k_para.replace('qkv', 'conv2')].transpose((1, 0, 2, 3))
v = params[k_para.replace('qkv', 'conv3')].transpose((1, 0, 2, 3))
new_state_dict[k1] = np.concatenate([q[:, :, 0, 0], k[:, :, 0, 0], v[:, :, 0, 0]], -1)
elif 'decoder' in k and 'self_attn' in k and 'qkv' in k and 'bias' in k:
k_para = k[:13] + 'layers.' + k[13:]
q = params[k_para.replace('qkv', 'conv1')]
k = params[k_para.replace('qkv', 'conv2')]
v = params[k_para.replace('qkv', 'conv3')]
new_state_dict[k1] = np.concatenate([q, k, v], -1)
elif 'decoder' in k and 'self_attn' in k and 'out_proj' in k:
k_para = k[:13] + 'layers.' + k[13:]
new_state_dict[k1] = params[k_para]
elif 'decoder' in k and 'cross_attn' in k and 'q' in k and 'weight' in k:
k_para = k[:13] + 'layers.' + k[13:]
k_para = k_para.replace('cross_attn', 'multihead_attn')
q = params[k_para.replace('q', 'conv1')].transpose((1, 0, 2, 3))
new_state_dict[k1] = q[:, :, 0, 0]
elif 'decoder' in k and 'cross_attn' in k and 'q' in k and 'bias' in k:
k_para = k[:13] + 'layers.' + k[13:]
k_para = k_para.replace('cross_attn', 'multihead_attn')
q = params[k_para.replace('q', 'conv1')]
new_state_dict[k1] = q
elif 'decoder' in k and 'cross_attn' in k and 'kv' in k and 'weight' in k:
k_para = k[:13] + 'layers.' + k[13:]
k_para = k_para.replace('cross_attn', 'multihead_attn')
k = params[k_para.replace('kv', 'conv2')].transpose((1, 0, 2, 3))
v = params[k_para.replace('kv', 'conv3')].transpose((1, 0, 2, 3))
new_state_dict[k1] = np.concatenate([k[:, :, 0, 0], v[:, :, 0, 0]], -1)
elif 'decoder' in k and 'cross_attn' in k and 'kv' in k and 'bias' in k:
k_para = k[:13] + 'layers.' + k[13:]
k_para = k_para.replace('cross_attn', 'multihead_attn')
k = params[k_para.replace('kv', 'conv2')]
v = params[k_para.replace('kv', 'conv3')]
new_state_dict[k1] = np.concatenate([k, v], -1)
elif 'decoder' in k and 'cross_attn' in k and 'out_proj' in k:
k_para = k[:13] + 'layers.' + k[13:]
k_para = k_para.replace('cross_attn', 'multihead_attn')
new_state_dict[k1] = params[k_para]
elif 'decoder' in k and 'norm' in k:
k_para = k[:13] + 'layers.' + k[13:]
new_state_dict[k1] = params[k_para]
elif 'mlp' in k and 'weight' in k:
k_para = k[:13] + 'layers.' + k[13:]
k_para = k_para.replace('fc', 'conv')
k_para = k_para.replace('mlp.', '')
w = params[k_para].transpose((1, 0, 2, 3))
new_state_dict[k1] = w[:, :, 0, 0]
elif 'mlp' in k and 'bias' in k:
k_para = k[:13] + 'layers.' + k[13:]
k_para = k_para.replace('fc', 'conv')
k_para = k_para.replace('mlp.', '')
w = params[k_para]
new_state_dict[k1] = w
else:
new_state_dict[k1] = params[k1]
if list(new_state_dict[k1].shape) != list(v1.shape):
print(k1)
for k, v1 in state_dict.items():
if k not in new_state_dict.keys():
print(1, k)
elif list(new_state_dict[k].shape) != list(v1.shape):
print(2, k)
model.set_state_dict(new_state_dict)
paddle.save(model.state_dict(), 'nrtrnew_from_old_params.pdparams')
```
2. 新版相比与旧版,代码结构简洁,推理速度有所提高。
## 引用
```bibtex
@article{Sheng2019NRTR,
title = {NRTR: A No-Recurrence Sequence-to-Sequence Model For Scene Text Recognition},
author = {Fenfen Sheng and Zhineng Chen and Bo Xu},
booktitle = {ICDAR},
year = {2019},
url = {http://arxiv.org/abs/1806.00926},
pages = {781-786}
}
```