Need guidance with using openvino inference on custom data and classes.
Please suggest semantic segmentation training with custom data and openvino inference capability.
First and foremost, you'll need to create your own custom model if you don't have any, or use Transfer Learning with your custom data/dataset.
you could use OpenVINO pre-trained model: Semantic Segmentation. Bear in mind that all pre-trained models have layers that already been frozen to sustain whatever they already learned. If you want to use your custom data/dataset with these model, you'll need to unfreeze those layers or in another word re-train them. This process won't be straightforward and would be helpful if you already have development background. You may refer to this GitHub repo.
Once you have that ready, you can proceed to infer the model with OpenVINO.
There are 3 ways to get OpenVINO to work in your env:
- Binary installation (choose the correct OS as yours): Install Intel® Distribution of OpenVINO™ toolkit for Windows* 10
- Pip installation: PyPi OpenVINO-dev installation
- Build from source from OpenVINO GitHub repo
Next, you'll need to convert your model's native format into OV IR. See this documentation
Infer that model with OpenVINO sample app, e.g.: benchmark_app -m model.xml -d CPU
Thank you for the information.
Some more questions have come up:
1. The OpenVINO pre-trained model: Semantic Segmentation
it says that the above version is old.
Is there a 2022.1 version?
2. Is there a "train" example code for re-train or transfer learning ?
It really helps when there is an example code I could modify.
If you search up you'll find this semantic-segmentation-adas-0001
The official sample app that you could try also mentioned there.
Generally, Transfer Learning in AI means leveraging an existing neural network architecture proven to work on problems similar to your own. You'll need to do your own research as this is out of OpenVINO scope.
Maybe you could research from here if you intend to implement this method:
From OpenVINO perspective, you could re-train the pre-trained model using OpenVINO Training Extension. You'll need to read the instruction on how to use it on the link I provided before (especially the Quick Start Guide): Training Extension
Rule thumb in any coding is to explore & experiment, if in doubt, try it first, and it's good not to practice spoon-fed as certain things you'll need to solve on your own. (Not a condemn but simply an advise for your future endeavours which I also practise)
I appreciate the help but do not appreciate the tone or sarcasm.
No one is asking about SPOON-FED! This is a place to freely ask questions. If you can't answer then have someone else do it.
I will make an official complaint about you to whoever is managing these.
Actually, I don't intend of sarcasm or sarcastic tone or anything and this is a platform where people could ask questions freely.
What I'm trying to convey is certain methodology might require you to build up your own code with only documentations as reference without the code to modify (this is what I meant with the spoon fed) and your choice is to ask around forum/discord.
I apologize if that sentence has mislead to different meaning.
Sorry for the bad experience.
I am afraid that you overlooked this Quick Start Guide as mentioned before.
You can use the OTE CLI commands to train models with model template. For segmentation task, the example model templates can be found in respective model folder.
Thank you for your question. If you need any additional information from Intel, please submit a new question as this thread is no longer being monitored.