Regarding E2E Object Detection in TFLite

[Sayak Paul]

Hi folks,

A few days ago I open-sourced this repository that shows how to train a custom detection model with the TFOD API (TF2 and TF1), optimize it with TFLite, and perform inference with the optimized model. It contains the following notebooks (pardon the naming) - 

• Training_a_pets_detector_model_within_minutes_with_TFOD_API.ipynb: Shows how to train a custom object detection model on the Pets dataset (non-eager mode) with Cloud TPUs. Note that it does not use TPUs offered by Colab.
• Running_inference_with_a_custom_TFOD_API_model.ipynb: Shows how to export a SavedModel graph from the trained checkpoint files, and run inference.
• Object_Detection_in_TFLite.ipynb: Shows how to quantize the original model, generate a TFLite model, and run inference.
• Training_MobileDet_Custom_Dataset.ipynb: Shows how to train a custom object detection model on the Pets dataset (non-eager mode) on Colab (GPU), optimize the fine-tuned model with TFLite, and perform inference with the optimized model. 

I hope you'd find them to be useful. I would like to thank the GDE program for providing me with GCP support which made the first notebook possible :)

Sayak Paul | sayak.dev

[Kshitiz Rimal]

Thanks for sharing this, Sayak! Looks amazing!

Regards,

Kshitiz 

--
You received this message because you are subscribed to the Google Groups "ML on Mobile OS Working Group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mlwg-mobile...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/mlwg-mobile/CAGa_XGExtViSJOLCBPxWjwEnDz8Q%2BkdeXx28AAorRQn6%3DSNhpw%40mail.gmail.com.
For more options, visit https://groups.google.com/d/optout.

[Soonson Kwon]

+Karl Weinmeister +Martin Gorner FYI for Cloud TPU

Thank you for sharing, Sayak!

[Srivatsan Srinivasan]

Sayak,

This is really amazing 👏

[George Soloupis]

You have done a great job Sayak!

Really useful! 👏

[Khanh LeViet]

This is great! Thanks Sayak.

--

    

Le Viet Gia Khanh (カン)

TensorFlow Developer Advocate

mail: khan...@google.com

tel: +81 3-6384-6867 

Shibuya Stream 

3-21-3 Shibuya, Shibuya-ku, Tokyo

150-0002, Japan

[Sayak Paul]

Hi folks,

Thanks for the kind words. Today I updated the Colab Training notebook and also added a notebook to build a dataset for TFOD API training (non-eager mode). 

So, here's two complete notebooks that demonstrate an end-to-end detection pipeline solely on Colab:

https://github.com/sayakpaul/E2E-Object-Detection-in-TFLite/tree/master/colab_training. 

Thanks to Khanh for the dataset idea. 

Sayak Paul | sayak.dev

[Kshitiz Rimal]

Hi Sayak,

Thanks for sharing the notebook. Looks amazing. Alternatively, have you tried converting directly from PASCAL or COCO type format to TFRecords using the scripts provided in dataset tools?

I found it quite useful when your dataset is in any one of those popular format:

https://github.com/tensorflow/models/tree/master/research/object_detection/dataset_tools

There is also a service called RoboFlow which allows you to directly convert any of those format to TFRecord with data-preprocessing, data split and augmentation features for free.

Here is a blog to that:

https://blog.roboflow.com/how-to-create-to-a-tfrecord-file-for-computer-vision/

Enjoying your E2E notebooks on object detection, it was something I was also looking into for a while. Again, thanks for sharing.

Regards,

Kshitiz 

-- 
You received this message because you are subscribed to the Google Groups "ML on Mobile OS Working Group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to mlwg-mobile...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/mlwg-mobile/CAGa_XGFrwcpbohED3dx%3Dtym-vuEUR%3DC-FufMcPE7m80JLcrjCw%40mail.gmail.com.

[Sayak Paul]

Hey Kshitiz,

Thanks for the kind words :)

I am aware of the dataset scripts provided by the TFOD API. But I wanted to have a bit more autonomy in the process, hence :) 

RoboFlow is from another galaxy! Thank you so much for sharing the resources. I will check them out in details in a few days. 

Sayak Paul | sayak.dev