Skip to content

Instantly share code, notes, and snippets.

@djdongjin
Last active September 4, 2020 16:49
Show Gist options
  • Save djdongjin/bdd20e0020b0bf48544a0906c98f98c2 to your computer and use it in GitHub Desktop.
Save djdongjin/bdd20e0020b0bf48544a0906c98f98c2 to your computer and use it in GitHub Desktop.
GSoC 2020 Summary

Google Summer of Code with TensorFlow

Mentors:

  • Elizabeth Kemp
  • Jaeyoun Kim

This is a summary of my GSoC project: New Tools for TensorFlow Hub. In this project, I extended and created tools that build models from command line directly. At the end, I reproduced a RoBERTa model using TF2.

A detailed blog post about the project can be found at TBD.

Milestone 1

The first milestone aims at updating the make_image_classifier tool of TF Hub, including dataset pipeline, colab demo, new tricks, etc.

Related PRs and commits:

  • PR595: update argument parsing.
  • PR612: update dataset pipeline and add distribute strategy.
  • PR609: update corresponding colab demo.
  • PR602: address a TF Hub bug.
  • new tricks: add new tricks for training.

The new tricks branch will be PRed after PR609 and PR612 are merged.

Milestone 2

Milestone 2 is to create a make_object_detector tool for TF Hub. It builds an object detection model finetuned on customized datasets, by using pretrained models from tfhub.dev.

Related PRs and commits:

  • An fine-tunable object detection colab.

Milestone 3

The last milestone of the project is to implement the RoBERTa model for TF Model Garden.

Related PRs and commits:

  • TBD.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment