TensorFlow Extend with OpenVINO pre-trained models

I am trying to develop a pipeline with TFX for a computer vision problem. I am using OpenVINO pre-trained models for this task. Does TFX supports to use these models and can TFX used for computer vision tasks?

OpenVINO is used to optimize trained models for inference on Intel hardware. If you’re planning to do transfer learning and fine tune the model with new data you’ll need to unfreeze the model parameters, but it looks to me like the OpenVINO IR does not support that.

For a computer vision problem I might suggest that you take a look at the CIFAR-10 example. Another great resource is KerasCV.

KerasCV, in its current stage, would be the most useful for augmentation (given it’s an image classification problem). Cc: @Luke_Wood

Regarding using OpenVINO pre-trained models with TFX, I am assuming it’s just for inference? If so, you could either write a custom component that allows you to load and persist the OpenVINO model throughout the pipeline lifetime. Or you could use a Beam pipeline to perform batch inference if that’s what you’re looking for.

Again pinging @Robert_Crowe to just gauge the possibility of writing the TFX component I mentioned. I think it should be possible, WDYT?

Thank you very much. I’ll refer them @Robert_Crowe

Yes it is for inference. Thank you very much. @Sayak_Paul

AFAIK yes, it should be possible, especially in a container, but I don’t know much about OpenVINO.