Retrain fastseg model - performance

[Mauricio]

Hi, I have working on EII for Intel, and we are using OpenVINO model in our use cases.

I have tried to retrain the semantic segmentation model, but I have not been able to.
semantic segemtation links:

https://docs.openvino.ai/2021.4/omz_models_model_semantic_segmentation_adas_0001.html

https://docs.openvino.ai/2021.4/omz_models_model_fastseg_small.html

My idea is to retrain the semantic segmentation model to detect only traffic signals, in order to improve the performance (FPS), because the current performance is only 1 or 2 fps. Is this idea correct? Will the performance increase if I decrease the number of detected classes?

I have used an aws ec2 instance with an nvidia Tesla card and another with a T3 card, following the steps in the guide:

https://github.com/ekzhang/semantic-segmentation

which comes from:

https://github.com/ekzhang/fastseg

Do you have a guide to train this model?, because the official guide is not very clear.

Best regards,

Mauricio.

[Zulkifli_Intel · Accepted solution]

Hi Mauricio,

Greetings to you.

To retrain an Intel pre-trained model, you can use the Training Extensions to do so. Please refer to the quick-start guide to get started.

Reducing the number of class objects may have an impact on the inference speed, but it's not significant as it will only

change a few layers and make it simpler. Another way to increase the precision time is you can use a smaller precision model (such as FP16/INT8).

Sincerely,

Zulkifli 

[Mauricio]

Hi, I have been doing some research on OTE.

I have followed all the steps to retrain a model, from the quick start guide.

But, we have a problem because we are using EII 3.0 which uses OpenVINO 2021.4.1.

Can we use the OTE tool for OpenVINO 2021.4.1?

What is the github branch and documentation about it?

Best regards,

Mauricio.

[Mauricio]

Thank you for your quick and accurate response. I will investigate the OTE tool.

Regards!

Mauricio.

[Zulkifli_Intel]

Hello Mauricio,

This master branch of the Training Extension repository is based on OpenVINO 2021.4, you can refer to Prerequisites.

Moreover, Training Extension supported limited models. You can find the list of all supported segmentation models here in OTE Algorithms. For Intel Pre-Trained Models that are available in OMZ, all are in Intermediate (IR) format, the non-IR versions of Intel’s Pre-Trained Models are not publicly provided, hence they cannot be used for retraining purposes, it was also mentioned in this article.

Sincerely,

Zulkifli 

[Zulkifli_Intel]

Hi Mauricio,

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.

Sincerely,

Zulkifli