Advanced Triton Inference Pipeline for CRAFT (Character-Region Awareness For Text detection)

• High performance and advanced inference for text detection with Triton Inference Server. Model format included: TensorRT, ONNX, Torchscript.
  
• CRAFT: (forked from https://github.com/clovaai/CRAFT-pytorch) Official Pytorch implementation of CRAFT text detector | Paper | Pretrained Model | Supplementary

Overview

Implementation new inference pipeline using NVIDIA Triton Inference Server for CRAFT text detector in Pytorch.

Author

k9ele7en. Give 1 star if you find some value in this repo.
Thank you.

License

[BSD-3-Clause License] The BSD 3-clause license allows you almost unlimited freedom with the software so long as you include the BSD copyright and license notice in it (found in Fulltext).

Updates

13 Jul, 2021: Initial update, preparation script run well.

14 Jul, 2021: Inference on Triton server run well (single request), TensorRT format give advance performance.

Getting started

1. Install dependencies

Requirements

$ pip install -r requirements.txt

2. Install required environment for inference using Triton server

Check ./README_ENV.md for details. Install tools/packages included:

• TensorRT
• Docker
• nvidia-docker
• PyCUDA ...

3. Training

The code for training is not included in this repository, as ClovaAI provided.

4. Inference instruction using pretrained model

• Download the trained models

Model name	Used datasets	Languages	Purpose	Model Link
General	SynthText, IC13, IC17	Eng + MLT	For general purpose	Click
IC15	SynthText, IC15	Eng	For IC15 only	Click
LinkRefiner	CTW1500	-	Used with the General Model	Click

5. Model preparation before run Triton server:

a. Triton Inference Server inference: see details at ./README_ENV.md
Initially, you need to run a (.sh) script to prepare Model Repo, then, you just need to run Docker image when inferencing. Script get things ready for Triton server, steps covered:

• Convert downloaded pretrain into mutiple formats
• Locate converted model formats into Triton's Model Repository
• Run (Pull first if not exist) Triton Server image from NGC

Check if Server running correctly:

$ curl -v localhost:8000/v2/health/ready
...
< HTTP/1.1 200 OK
< Content-Length: 0
< Content-Type: text/plain

Now everythings ready, start inference by:

• Run docker image of Triton server (replace mount -v path to your full path to model_repository):

$ sudo docker run --gpus all --rm -p8000:8000 -p8001:8001 -p8002:8002 -v /home/maverick911/repo/Triton-TensorRT-Inference-CRAFT-pytorch/model_repository:/models nvcr.io/nvidia/tritonserver:21.05-py3 tritonserver --model-repository=/models
...
+------------+---------+--------+
| Model      | Version | Status |
+------------+---------+--------+
| detec_onnx | 1       | READY  |
| detec_pt   | 1       | READY  |
| detec_trt  | 1       | READY  |
+------------+---------+--------+
I0714 00:37:55.265177 1 grpc_server.cc:4062] Started GRPCInferenceService at 0.0.0.0:8001
I0714 00:37:55.269588 1 http_server.cc:2887] Started HTTPService at 0.0.0.0:8000
I0714 00:37:55.312507 1 http_server.cc:2906] Started Metrics Service at 0.0.0.0:8002

Run infer by cmd:

$ python infer_triton.py -m='detec_trt' -x=1 --test_folder='./images' -i='grpc' -u='localhost:8001'
Request 1, batch size 1s/sample.jpg
elapsed time : 0.9521937370300293s

Output from Triton:

Performance benchmarks: single image (sample.jpg), time in seconds

• Triton server: (gRPC-HTTP):
  

  Model format	gRPC (s)	HTTP (s)
  TensoRT	0.946	0.952
  Torchscript	1.244	1.098
  ONNX	1.052	1.060

• Classic Pytorch: 1.319s

Arguments

• -m: name of model with format
• -x: version of model
• --test_folder: input image/folder
• -i: protocol (HTTP/gRPC)
• -u: URL of corresponding protocol (HTTP-8000, gRPC-8001)
• ... (Details in ./infer_triton.py)

Notes:

• Error below is caused by wrong dynamic input shapes, check if the input image shape is valid to dynamic shapes in config.

inference failed: [StatusCode.INTERNAL] request specifies invalid shape for input 'input' for detec_trt_0_gpu0. Error details: model expected the shape of dimension 2 to be between 256 and 1200 but received 1216

b. Classic Pytorch (.pth) inference:

$ python test.py --trained_model=[weightfile] --test_folder=[folder path to test images]

The result image and socre maps will be saved to ./result by default.