Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Inference Accelerated PDF batch parsing #106

Open
wants to merge 85 commits into
base: main
Choose a base branch
from

Conversation

veya2ztn
Copy link

@veya2ztn veya2ztn commented Sep 3, 2024

Inference Accelerated PDF parsing

This fold include a series infra-accelerate modules for origin PDF parsing, including:

  • intergrated preprocessing into dataloader
  • fast_postprocessing
  • torch.compile + bf16
  • tensorRT
  • Torch-TensorRT[https://pytorch.org/TensorRT/]

Those engine is tested on a 80,000,000 pdf dataset and get a 5-10x speedup compared with the origin pdf parsing engine. Basicly, it can reach 6-10 pages per second on a single A100 GPU.

This is not a pipline framework but seperated into three task-wise batch processing engine. But it can be easily integrated into your own pipline framework.

  • Detection (Bounding Boxing)
  • Recognition (OCR)
  • Math formula recognition (MFR)

Detection (Bounding Boxing)

Check the unit case:1000pdf takes around 20-30min

python batch_running_task/task_layout/rough_layout.py

LayoutLM

The layoutLM is based on the detectron2. The main Vision Engine(ViT) is implemented via huggingface, the postprocess is based on detectron.
There is a tensorRT version of the detectron model https://github.com/NVIDIA/TensorRT/tree/main/samples/python/detectron2 , but it is only for Mask R-CNN backbone.
The tensorRT author manuelly develop the CUDA NMS and ROIAlign such as DET2GraphSurgeon (see https://github.com/NVIDIA/TensorRT/blob/main/samples/python/detectron2/create_onnx.py) to convert the detectron2 model to tensorRT model.
For layoutLM, there is no such tool to convert whole model into a tensorRT engine.
There are serveral ways to accelerate the layoutLM model:

  • accelerate part by part, such as the tensorRT ViT backbone with detectron ROIAlign, NMS.
  • use torch.compile
  • use bf16

In this repo, I use the torch.compile(1.5x) and bf16(2x) to accelerate the layoutLM model. The tensorRT version is not implemented yet.

Another way to accelerate the layoutLM is avoid .numpy() large GPU tensor. Origin code will use

boxes = outputs["instances"].to("cpu")._fields["pred_boxes"].tensor.numpy()
labels = outputs["instances"].to("cpu")._fields["pred_classes"].numpy()
scores = outputs["instances"].to("cpu")._fields["scores"].numpy()

This will copy the large tensor from GPU to CPU. (Since we later only gather part of data via mask, full tensor copy is unnecessary).
The better way is to do slicing on GPU tensor and then copy the sliced tensor to CPU. (2x) (see batch_running_task/task_layout/get_batch_layout_model.py)

MFD

MFD(Math Formula Detection) is a simple YOLO model build through ultralytics. It has a good tensorRT convert tool chain. See https://docs.ultralytics.com/modes/export/ and convension/MDF/convert.py
Download the engine via huggingface-cli download --resume-download --local-dir-use-symlinks False LLM4SCIENCE/ultralytics-YOLO-MFD --local-dir models/MFD. The batchsize and tensorRT version==10.3.0 must match! if you want to use the trt_engine directly.

PaddleOCR-Det

PaddleOCR-Det is the best text detecter around the world. But original paddle det only support one image per batch. In our detection task, every image is normlized into same size, so the original paddle det does not fit our task. Refer to https://github.com/WenmuZhou/PytorchOCR, Zhou has convert the paddleOCR into pytorch. It allow us use batch detection in pytorch now.

There is a big speed up possiblity for the postprocessing for the paddleOCR-Det module. Currently, we use the DB postprocessing. See https://github.com/PaddlePaddle/PaddleOCR/blob/main/ppocr/postprocess/db_postprocess.py. The DB postprocessing is the slow part compare to whole detection process. Currently, there is no any speedup solution for the DB postprocessing.

Detection Async(experimental)

See batch_running_task/task_layout/rough_layout_with_aync.py
The async detection is a way to async postprocess and GPU inference. It works perfectly. But in slurm system, there is exit error when run the script, this will make your machine CPU soft lock. So, I do not recommend to use this script in slurm system.

Recognition (OCR)

Check the unit case:1000 pdf takes around 2-5 min

    python batch_running_task/task_rec/rough_rec.py

PaddleOCR-Rec is the best text recognizer around the world. The original paddle rec support batch image processing. And the origin paddleOCR "is already very fast".
However, you can see I still use PytorchOCR in this part. Just want to provide a non-paddle solution.
Download the engine via huggingface-cli download --resume-download --local-dir-use-symlinks False LLM4SCIENCE/pytorch_paddle_weight --local-dir models/pytorch_paddle_weight . The batchsize and tensorRT version==10.3.0 must match! if you want to use the trt_engine directly.

Math formula recognition (MFR)

Check the unit case: 1000 pdf takes around 2-5min

Update: I find for large formula like those in \begin{align} and \end{align} the recognition always quit unexcepted.

    python batch_running_task/task_mfr/rough_mfr.py

MFR model is nougat based model named UniMERNet. I tried to use Huggingface tensorRT convert tool chain to convert the model into tensorRT. But it failed. (The reshape module is not set properly). One way is using the TensorRT-LLM, see https://github.com/NVIDIA/TensorRT-LLM/tree/main/examples/multimodal and convension/unimernet.

  • Notice TensorRT-LLM will default install mpi4py=4.*.* which will require mpi.so40. The conda conda install -c conda-forge openmpi can only support openmpi==3.*.*'. So you need to install openmpifrom source. Or, you can justpip install mpi4py==3.*`.
  • Notice you should srun --mpi=pmi2 when run script in slurm.

Download the engine via huggingface-cli download --resume-download --local-dir-use-symlinks False LLM4SCIENCE/unimernet --local-dir models/MFR/unimernet. The batchsize and tensorRT version==10.3.0 must match! if you want to use the trt_engine directly.

The different between LLM4SCIENCE/unimernet and wanderkid/unimernet is we delete the counting module in weight file. (it only works in training). And it is a pure nougat model.

Batch run the task

Each task has a "batch_deal_with_xxx" module which will automatively schedule task. For example, your can prepare a .jsonl file named test.filelist with each line is

{"track_id":"e8824f5a-9fcb-4ee5-b2d4-6bf2c67019dc","path":"10.1017/cbo9780511770425.012.pdf","file_type":"pdf","content_type":"application/pdf","content_length":80078,"title":"German Idealism and the Concept of Punishment || Conclusion","remark":{"file_id":"cbo9780511770425.012","file_source_type":"paper","original_file_id":"10.1017/cbo9780511770425.012","file_name":"10.1017/cbo9780511770425.012.pdf","author":"Merle, Jean-Christophe"}}
{"track_id":"64d182ba-21bf-478f-bb65-6a276aab3f4d","path":"10.1111/j.1365-2559.2006.02442.x.pdf","file_type":"pdf","content_type":"application/pdf","content_length":493629,"title":"Sensitivity and specificity of immunohistochemical antibodies used to distinguish between benign and malignant pleural disease: a systematic review of published reports","remark":{"file_id":"j.1365-2559.2006.02442.x","file_source_type":"paper","original_file_id":"10.1111/j.1365-2559.2006.02442.x","file_name":"10.1111/j.1365-2559.2006.02442.x.pdf","author":"J King; N Thatcher; C Pickering; P Hasleton"}}

and then run

python batch_running_task/task_layout/batch_deal_with_layout.py --root test.filelist
python batch_running_task/task_layout/batch_deal_with_rec.py --root test.filelist
python batch_running_task/task_layout/batch_deal_with_mfr.py --root test.filelist

Copy link


Thank you for your submission, we really appreciate it. Like many open-source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution. You can sign the CLA by just posting a Pull Request Comment same as the below format.


I have read the CLA Document and I hereby sign the CLA


You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

1 participant