TexTeller v2
This commit is contained in:
11
README.md
11
README.md
@@ -7,9 +7,12 @@
|
||||
<p align="center">
|
||||
English | <a href="./assets/README_zh.md">中文</a>
|
||||
</p>
|
||||
<p align="center">
|
||||
<!-- <p align="center">
|
||||
<img src="./assets/web_demo.gif" alt="TexTeller_demo" width=800>
|
||||
</p>
|
||||
</p> -->
|
||||
<video width="800" controls>
|
||||
<source src="./assets/test.mp4" type="video/mp4">
|
||||
</video>
|
||||
</div>
|
||||
|
||||
TexTeller is an end-to-end formula recognition model based on ViT, capable of converting images into corresponding LaTeX formulas.
|
||||
@@ -21,6 +24,8 @@ TexTeller was trained with ~~550K~~7.5M image-formula pairs (dataset available [
|
||||
## 🔄 Change Log
|
||||
|
||||
* 📮[2024-03-24] TexTeller 2.0 released! The training data for TexTeller 2.0 has been increased to 7.5M (about **15 times more** than TexTeller 1.0 and also improved in data quality). The trained TexTeller 2.0 demonstrated **superior performance** in the test set, especially in recognizing rare symbols, complex multi-line formulas, and matrices.
|
||||
> [!INFO]
|
||||
> [There](./assets/test.pdf) are more test images here and a horizontal comparison of recognition models from different companies.
|
||||
|
||||
## 🔑 Prerequisites
|
||||
|
||||
@@ -138,8 +143,6 @@ In `TexTeller/src/globals.py` and `TexTeller/src/models/ocr_model/train/train_ar
|
||||
|
||||
## 🚧 Limitations
|
||||
|
||||
* Some complex multi-line scenarios are not well handled (e.g., long formulas mixed with matrices)
|
||||
|
||||
* Does not support scanned images and PDF document recognition
|
||||
|
||||
* Does not support handwritten formulas
|
||||
|
||||
@@ -7,9 +7,12 @@
|
||||
<p align="center">
|
||||
<a href="../README.md">English</a> | 中文
|
||||
</p>
|
||||
<p align="center">
|
||||
<!-- <p align="center">
|
||||
<img src="./web_demo.gif" alt="TexTeller_demo" width=800>
|
||||
</p>
|
||||
</p> -->
|
||||
<video width="800" controls>
|
||||
<source src="./test.mp4" type="video/mp4">
|
||||
</video>
|
||||
</div>
|
||||
|
||||
TexTeller是一个基于ViT的端到端公式识别模型,可以把图片转换为对应的latex公式
|
||||
@@ -21,6 +24,8 @@ TexTeller用了~~550K~~7.5M的图片-公式对进行训练(数据集可以在[
|
||||
## 🔄 变更信息
|
||||
|
||||
* 📮[2024-03-24] TexTeller2.0发布!TexTeller2.0的训练数据增大到了7.5M(相较于TexTeller1.0**增加了~15倍**并且数据质量也有所改善)。训练后的TexTeller2.0在测试集中展现出了**更加优越的性能**,尤其在生僻符号、复杂多行、矩阵的识别场景中。
|
||||
> [!INFO]
|
||||
> 在[这里](./test.pdf)有更多的测试图片以及各家识别模型的横向对比。
|
||||
|
||||
## 🔑 前置条件
|
||||
|
||||
@@ -28,7 +33,8 @@ python=3.10
|
||||
|
||||
pytorch
|
||||
|
||||
> 注意: 只有CUDA版本>= 12.0被完全测试过,所以最好使用>= 12.0的CUDA版本
|
||||
> [!WARNING]
|
||||
> 只有CUDA版本>= 12.0被完全测试过,所以最好使用>= 12.0的CUDA版本
|
||||
|
||||
## 🖼 关于把latex渲染成图片
|
||||
|
||||
@@ -62,6 +68,7 @@ pytorch
|
||||
#+e.g. python inference.py -img "./img.jpg" -cuda
|
||||
```
|
||||
|
||||
> [!NOTE]
|
||||
> 第一次运行时会在hugging face上下载所需要的checkpoints
|
||||
|
||||
## ❓ 常见问题:无法连接到Hugging Face
|
||||
@@ -102,9 +109,11 @@ pytorch
|
||||
|
||||
然后在浏览器里输入`http://localhost:8501`就可以看到web demo
|
||||
|
||||
> [!TIP]
|
||||
> 你可以改变`start_web.sh`的默认配置, 例如使用GPU进行推理(e.g. `USE_CUDA=True`) 或者增加beams的数量(e.g. `NUM_BEAM=3`)来获得更高的精确度
|
||||
|
||||
**NOTE:** 如果你想直接把预测结果在网页上渲染成图片(比如为了检查预测结果是否正确)你需要确保[xelatex被正确安装](https://github.com/OleehyO/TexTeller?tab=readme-ov-file#Rendering-Predicted-Results)
|
||||
> [!IMPORTANT]
|
||||
> 如果你想直接把预测结果在网页上渲染成图片(比如为了检查预测结果是否正确)你需要确保[xelatex被正确安装](https://github.com/OleehyO/TexTeller?tab=readme-ov-file#Rendering-Predicted-Results)
|
||||
|
||||
## 📡 API调用
|
||||
|
||||
@@ -127,6 +136,7 @@ python server.py # default settings
|
||||
| `--ncpu_per_replica` | 每个服务副本所用的CPU核心数,*默认为1*。 |
|
||||
| `--ngpu_per_replica` | 每个服务副本所用的GPU数量,*默认为1*。你可以把这个值设置成 0~1之间的数,这样会在一个GPU上运行多个服务副本来共享GPU,从而提高GPU的利用率。(注意,如果 --num_replicas 2, --ngpu_per_replica 0.7, 那么就必须要有2个GPU可用) |
|
||||
|
||||
> [!NOTE]
|
||||
> 一个客户端demo可以在`TexTeller/client/demo.py`找到,你可以参考`demo.py`来给server发送请求
|
||||
|
||||
## 🏋️♂️ 训练
|
||||
@@ -142,6 +152,7 @@ python server.py # default settings
|
||||
如果你使用了不一样的数据集,你可能需要重新训练tokenizer来得到一个不一样的字典。配置好数据集后,可以通过以下命令来训练自己的tokenizer:
|
||||
|
||||
1. 在`TexTeller/src/models/tokenizer/train.py`中,修改`new_tokenizer.save_pretrained('./your_dir_name')`为你自定义的输出目录
|
||||
> [!IMPORTANT]
|
||||
> 如果要用一个不一样大小的字典(默认1W个token),你需要在 `TexTeller/src/models/globals.py`中修改`VOCAB_SIZE`变量
|
||||
|
||||
2. **在 `TexTeller/src` 目录下**运行以下命令:
|
||||
@@ -162,20 +173,11 @@ python -m models.ocr_model.train.train
|
||||
|
||||
在`TexTeller/src/globals.py`和`TexTeller/src/models/ocr_model/train/train_args.py`中,你可以改变模型的架构以及训练的超参数。
|
||||
|
||||
> [!NOTE]
|
||||
> 我们的训练脚本使用了[Hugging Face Transformers](https://github.com/huggingface/transformers)库, 所以你可以参考他们提供的[文档](https://huggingface.co/docs/transformers/v4.32.1/main_classes/trainer#transformers.TrainingArguments)来获取更多训练参数的细节以及配置。
|
||||
|
||||
## 🚧 不足
|
||||
|
||||
* 部分细节很多的公式无法做到100%的准确率
|
||||
|
||||
<img src="" width=30, height=30>
|
||||
|
||||
* 部分复杂的大型多行公式识别效果不佳(例如长公式与矩阵混合)
|
||||
|
||||
<img src="" width=30, height=30>
|
||||
|
||||
> 如果遇到这种情况,你可以尝试把大型的多行公式分成多个小的子公式来识别。
|
||||
|
||||
* 不支持扫描图片以及PDF文档识别
|
||||
|
||||
* 不支持手写体公式
|
||||
|
||||
BIN
assets/test.mp4
Normal file
BIN
assets/test.mp4
Normal file
Binary file not shown.
BIN
assets/test.pdf
Normal file
BIN
assets/test.pdf
Normal file
Binary file not shown.
@@ -17,7 +17,7 @@ from transformers import (
|
||||
|
||||
|
||||
class TexTeller(VisionEncoderDecoderModel):
|
||||
REPO_NAME = '/home/lhy/code/TexTeller/src/models/ocr_model/train/train_result/TexTellerv2/checkpoint-356000'
|
||||
REPO_NAME = '/home/lhy/code/TexTeller/src/models/ocr_model/train/train_result/TexTellerv2/checkpoint-588000'
|
||||
def __init__(self, decoder_path=None, tokenizer_path=None):
|
||||
encoder = ViTModel(ViTConfig(
|
||||
image_size=FIXED_IMG_SIZE,
|
||||
|
||||
@@ -3,7 +3,8 @@ set -exu
|
||||
|
||||
export CHECKPOINT_DIR="default"
|
||||
export TOKENIZER_DIR="default"
|
||||
export USE_CUDA=False # True or False (case-sensitive)
|
||||
export NUM_BEAM=1
|
||||
# export USE_CUDA=False # True or False (case-sensitive)
|
||||
export USE_CUDA=True # True or False (case-sensitive)
|
||||
export NUM_BEAM=10
|
||||
|
||||
streamlit run web.py
|
||||
|
||||
@@ -95,7 +95,7 @@ def rendering(formula: str, out_img_path: Path) -> bool:
|
||||
return p.returncode == 0
|
||||
|
||||
def pdf_to_pngbytes(pdf_path):
|
||||
images = convert_from_path(pdf_path, first_page=1, last_page=1)
|
||||
images = convert_from_path(pdf_path, dpi=400,first_page=1, last_page=1)
|
||||
trimmed_images = trim(images[0])
|
||||
png_image_bytes = io.BytesIO()
|
||||
trimmed_images.save(png_image_bytes, format='PNG')
|
||||
|
||||
Reference in New Issue
Block a user