Phase 1 Completed

Ayush Agrawal
GSoC ’19
Published in
3 min readJul 1, 2019

Hi there, welcome back! Let us continue with Tutorial 5 and see what’s going on with the project right now. I’m sorry for the delay in the post. I was waiting for Phase 1 Evals to get completed so that I can conclude Phase 1 with this post.

In last blog post I mentioned that I was facing some problem with my Image Classifier — The training loop seemed to be unresponsive and always gave 50% accuracy in every epoch.

Turned out that the error was because of the images I was passing to the classifier for training and validation purposes. I had naively converted the RGB Images to Greyscale Images which resulted in that performance loss. The original tutorial only divided the initial RGB values by 255 to make them appropriate for the network. I misunderstood that they were converting the images to greyscale and then passing them onto the network. Once I corrected it, everything started working absolutely fine.

Tutorial 5 is almost complete now— code review is going on currently.

Now moving on to Tutorial 6, this is where things get a little more interesting. Right now, there’s a hole in the Swift for TensorFlow ecosystem around Image Augmentation Techniques along with some other things that I need for upcoming tutorials.

Brad created an issue to track this: https://github.com/Ayush517/S4TF-Tutorials/issues/7 so that we can have an active discussion with everyone about what might be needed and how best to add it or work around these needs.

So, until then I am mostly looking for ways to work this out before diving into porting the remaining tutorials.

Swift for Tensorflow v0.4 is now available. Feel free to try it out in Colab or download a pre-built package from here.

I updated the converted tutorials to make them compatible with the release.

And finally, I passed First Evaluation of Google Summer of Code 2019.

First evaluation was basically a survey where we had to answer questions about our GSoC experience so far and the progress on our proposed project. Mentors had to give their evaluation of the students and how well they had worked, and whether or not they should pass or fail the student in the first evaluation.

I filled the evaluation and I knew well before that I’ll be passing since I had already talked to Paige regarding it.

To summarise, Phase I was more about getting to know the S4TF Ecosystem, how it works, Swift code style and how to proceed if you get stuck in something. There was no well-defined plan like I had to complete this tutorial in this much time.

Things are very flexible and my mentors have helped me in every way I could have imagined. They have made things accessible to me which I never even knew existed. I used to watch these developers online and in conferences, always dreaming about how they do it and what all it takes. Today, I am working with them, learning from them, interacting with them and living the best summer I never even knew I wanted.

Brad’s feedback has motivated me even further to give everything that I have to this project and make it successful.

I’ve just mailed my mentors to talk about how we further plan to move with the project. And I’m eagerly waiting for their reply…

It’s 3:15 AM here right now and I am feeling sleepy. I think I should go to sleep now.

Bye 👋, see you next week! 😴

--

--

Ayush Agrawal
GSoC ’19

Deep Learning Enthusiast | TensorFlow GSoC ’20 Mentor, GSoC ’19 Dev & GCI ’19 Mentor | BITS Pilani, India