SK-TDA4VM: Edge AI Studio Modeling

Hi,

The engineer is OoO for regional holidays until 11/20 so please expect delay in response.
Apologies for the inconvenience and thank you for your patience.

Regards,
Parth

Dear Parth Nagpal 

No problem!

Thank you for your patience.

Regards,
Parth

Hi Vic,

As you have queries are related specifically to model composer tool let me answer them.

Vic Chang said:

My understanding of EDGE-AI-STUDIO is that TI provides a method similar to online model training at

Yes, you can train the model by feeding images for different supported tasks like object detection and classification.

Vic Chang said:

1. Analysis and optimization/model compression/quantization acceleration?

You will get different graph level analysis info on epoch stage accuracy of current model and many more when training is on going. You can opt for supported quantization rage option 8/16 bits. Not sure what you mean by compression here, can you elaborate more on this term ?

Vic Chang said:

2. Use non-default Image Classification/Feature Extraction/Object Detection/Semantic Segmentation/Human Pose Estimation/3D Object Detection models?

Vic Chang said:

Can I design and create my own models, possibly not following a conventional architecture, and run them on the platform?

Per my understanding currently custom models are not supported as part of this tool.

However you can look at our edgeai tild tools repos for custom model training here : https://github.com/TexasInstruments/edgeai-tidl-tools

Hi Parth Nagpal 

The things I'm discussing are not problems, so there's no need to feel pressured or worried. The main focus is on understanding the situation and feasibility of this platform.

Q1
Model Compression:
Quantization:
Basically, it involves the conversion of floating point to integer for compression and acceleration.

Other Techniques:
Pruning, Distillation, Structural Optimization.

Q2
Custom Models:
In line with the Q1 question, other techniques involve modifications to the model structure (model compression). Q2 is more like a discussion on the portability of models.

If I have a trained model, can it be directly transformed for hardware acceleration?
Not for Classification/Feature Extraction/Object Detection/Semantic Segmentation/Human Pose Estimation/3D Object Detection models.
It's about a generic model structure.

Hi Parth Nagpal 

Thank you!

I'll give it a try.