Deployment code

[jey07]

Is there any different version or method of doc ocr which has production deployment code/env which runs fast ?

[felixdittrich92]

Hi @jey07 👋 ,
Do you mean something like a optimized graph inference pipeline (for example Onnxruntime) ? If so we work currently on ONNX support for each model and integrating a pipeline for this 👍
For the moment the fastest solution would be to use the detection and recognition models which uses a mobilenet backbone and run this on a Gpu

[jey07]

Hello @felixdittrich92 , may i know how much time does it take to detect and predict together using Onnex or in gpu.

[felixdittrich92]

@jey07 👋
This question is really hard to answer without knowledge about your environment and files 😅

• For ONNX as explained ONNX support is currently under development, if it is done it will speed up the latency on CPU.

• We have some benchmarks in the documentation :)

• To get the fastest latency i would suggest to use db_mobilenet_v3_large as detection model and crnn_mobilenet_v3_small / large as recognition model you can also take an look at this discussion: Bad latency on CPU #876