search

sayakpaul / portfolio

Personal site of Sayak Paul. Deployed here 👉
https://sayak.dev/
24 stars 16 forks source link

Optimizing MobileDet for Mobile Deployments | Sayak Paul #4

Open utterances-bot opened 4 years ago

utterances-bot commented 4 years ago

Optimizing MobileDet for Mobile Deployments | Sayak Paul

Learn about the criticalities of effectively optimizing MobileDet object detectors for mobile deployments.

https://sayak.dev/mobiledet-optimization/

RodolfoFerro commented 4 years ago

An amazing post.

I see that you also use Netron for visualization!

sayakpaul commented 4 years ago

Thanks @RodolfoFerro!

I LOVE Netron.

karuneshpalekar commented 3 years ago

Great job . Can you also provide some resource on merging tflite model with metadata . The documentation provides script for image classifier but there is nothing much about object detection .

sayakpaul commented 3 years ago

@karuneshpalekar as per my knowledge for image data, image classification, semantic segmentation, and style transfer are supported for Metadata integration. Object detection is not supported currently I think.

@khanhlvg can maybe provide additional details.

khanhlvg commented 3 years ago

Actually object detection metadata is already supported for the Task library but we haven't published any sample on how to append metadata to a TFLite object detection model yet. It's on the roadmap so stay tuned :)

Mukulareddy commented 2 years ago

Thanks for the post. It will be helpful if you could explain in detail the QAT steps for mobileDet for TF2. I could see in some other blogs that QAT is not supported for MobileDet. Can someone please confirm this.

sayakpaul commented 2 years ago

@khanhlvg could you help out @Mukulareddy regarding her doubt here?

alexander-sony commented 2 years ago

Great post! I also found that a fully quantized cpu model using repr dataset provides very bad detection results. Can you comment of why the results are so bad in this case?

sayakpaul commented 2 years ago

I didn't have the bandwidth to dive deeper into that unfortunately. I'd suggest directly reaching out to the authors of MobileDet to clarify this doubt.