← sources

How are image classification models actually deployed to end users?

source: reddit (reddit:MLOps)published 2026-10-13 14:48ingested 2026-10-14 02:39

retained text

I am a software engineer researching how image-classification models are handed-off to end users, those who actually run the model in a production environment. 

  
I am working on a Windows-based, batch image-classification application, and I am interested in understanding how teams currently handle deployment, rather than assume a particular approach is needed. 

  
For any of you who are in charge of deploying image classifiers:

1.  **Where does the inference run?** On a local workstation, an on-premises server, an edge device, or a cloud-hosted service?
2. **What software stack do you use?** E.g., PyTorch, ONNX Runtime, TensorRT, Triton, a custom Python application, or something else? Is it packaged as a desktop application, container, API, or scheduled batch job?
3. **Who actually operates the model?** Does the ML developer run the inference, or is the model handed off to another engineer, technician, analyst, or end user? How does that person submit images and obtain results?

I'm particularly interested in **batch classification of existing image collections**, as distinct from real-time camera inspection.

For example, a QA engineer wants to do a quality analysis of the manufacturing production runs over the last year. You have the model, in some form for him or her, but how do they run it?

In addition, how are the preprocessing requirements and class labels transferred to whomever operates the inference workflow?

Are there aspects of deployment or handoff that create recurring work for each deployment, or do your existing tools handle this adequately?

I'd appreciate hearing about actual workflows, including cases where a simple Python script is entirely sufficient.  

document fetched 2026-10-14 00:28 · canonical https://reddit.com/r/mlops/comments/1x4ynuh/how_are_image_classification_models_actually