Articles on Eshan's Blog

Finale

thisiseshan@gmail.com (Eshan) — Fri, 20 Aug 2021 23:31:23 +0000

Hey friends!
Thank you for being a part of this wonderful journey. The past 10 weeks were some of the best I have ever had and I am glad to have shared them with you.

Let's get on with the finale check-in with these simple questions.

What did I do this week?

Following up on last week's blog, All 3 PRs were merged. I sent out another PR to fix an ingestion summary bug and Colab fixes.
Additionally, I had presented my final project on Friday to the Activeloop team. They're quite happy with it and I can't wait for everyone to start using Hub auto asap!

What will I do next week?

I plan to continue contributing to Hub as much as I can. I love the package and I'm grateful to be working with such smart people.

Did I get stuck anywhere?

I found myself stuck at returning the correct number of ingested files in a directory, I have sent out a new PR to temporarily remove this feature and instead keep the ingestion summary simple.

Thank you for reading this, I am grateful for your time and hope you were able to learn something from my journey.

I'm gonna miss this a lot.
E

What a week! 3 PRs in one week, loving this! 🔥

thisiseshan@gmail.com (Eshan) — Sun, 15 Aug 2021 20:09:54 +0000

Hey friends!
Welcome to my blog post on my Google Summer of Code '21 Journey. The last week is about to begin and we're gonna be in the endgame soon enough, however this blog post is about the 3 new features I have been working on to get them into Hub auto quickly before we wrap up GsoC.

3 PRs for 3 features:

kaggle fixture

auto-compression

ingestion summary

Without further ado, lets talk about these features

Kaggle fixture [merged]

Right now, all kaggle tests run as regular tests. However with this PR, I have added a —kaggle fixture to specifically run pytests. This makes hub auto testing for kaggle datasets efficient and convienient.

Auto-compression [in review]

Hub auto did everything automatically except figuring out the compression type for the dataset. So naturally this made that possible. As long as the dataset is one of the accepted dataset types (.jpeg, .png, .jpg) setting 'compression' argument as "auto" will automatically figure out the compression type of the entire dataset.

Ingestion summary [in review]

Now once the ingestion process is complete, it's pretty neat to give the user an idea of all files skipped due to errors in the dataset.

Check this out:

When all files under the directory "images_classification" are ingested.

When "images_classification/lol.json' and "image_classification/class0/lol.json" are skipped during ingestion.

To conclude, 3 amazing feature. 1 merged, 2 in review, Coming to Hub in a few days.

with ♥️, E

Week 8-9: Hub auto PR merged 🥰 + New stuff 🐳

thisiseshan@gmail.com (Eshan) — Sat, 07 Aug 2021 16:53:46 +0000

Without further ado, Let's take a look into what I've been doing for the past week!

What did I do this week?

Following up on last week's blog, I fixed a few issues and implemented the suggestions that were provided by the activeloopai community (Dyllan, Abhinav, Fariz and Ivo). Well then, my code passes all test cases and I received 4 PR approves!
The PR got merged instantly!

Without wasting any time, I started working on the next steps for Hub auto because we've got a lot planned!

Couple of things that are coming up!

Auto detect compression type

Hub auto docs

kaggle pytest fixture

Ingestion summary

and more..

I specifically targeted an issue where the code fails if the kaggle credentials aren't stored as environment variables in the system.
And then worked on a script that would auto detect the compression type of the dataset. I submitted a PR for that here. But i think I'm gonna submit another one that includes the auto detect script and the kaggle pytest fixture.

What will I do next week?

I'm going to finish up the automatic compression detection, this would enable the user to just mention the dataset he / she wishes to ingest and boom! the code will figure out the most optimal image compression type and ingest it into a Hub dataset.

Did I get stuck anywhere?

I wouldn't say I was "stuck" somewhere. I just went ahead and learned what I didn't know. I figured out the auto detection of compression and that use a fraction of my brain power hehe.

Blog Post #5: PR to the main branch 🌱

thisiseshan@gmail.com (Eshan) — Sat, 31 Jul 2021 10:32:19 +0000

Hey friends!
I have always considered working on open-source projects to be closely related to watering my plants, to watch them burgeon brings pride and happiness to my heart. This is my first PR for hub auto and I can't wait for the world to use it.

The past 2 weeks have been all about:

Fine tuning the API code

Writing quality tests

Solving bugs

Providing well-defined docstrings

My PR has been reviewed and is about to get merged real soon, I am VERY excited for people to try out Hub auto and use it as the primary method of ingestion for image classification datasets.
Currently, I update my code with the suggestions I receive from the community in the form of PR reviews and would love to summarise my work in the blog here!

This PR provides 2 APIs:

Ingest()

The ingest API provides a one line solution to get your locally stored Image dataset on Hub.

Ingest_kaggle()

The ingest API provides a one line solution to get your Kaggle dataset to Hub.

I believe this is going to open doors to a lot of new Hub use cases and make onboarding of new users to Hub a breeze 🏖
For anyone interested, my PR is available here!

Week 6-7 check-in: Submit draft PR for hub.auto 🚀

thisiseshan@gmail.com (Eshan) — Fri, 23 Jul 2021 22:44:20 +0000

Let's take a look into what I've been doing for the past week!

What did I do this week?

Following up on last week's blog, I have been working consistently to write good test cases for hub.auto and I am glad to share that I finished up writing the tests this week (yesterday to be more precise). My mentor, Dyllan asked me to create another create another branch feature/2.0/hub-auto2 as the branch from last week has serious merge conflicts that couldn't be resolved.
Additionally, I wrote tests for increasing the coverage of the code using pytest.raises() and I made hub.auto more robust by handling few important edge cases.

What will I do next week?

I plan to fix anything that comes up as an issue and try to get the branch merged into main asap! That's the main goal.
Apart form that I will be volunteering to upload more datasets to Hub 2.0's cloud!

Did I get stuck anywhere?

I found myself stuck at few bugs after creating a draft PR. One issue was the tensor_names and label_names needed to replace every occurrence of ('\\') to ('/') using a string.replace() function as the circle tests failed on Windows. However, setting fixtures for Kaggle and Ingestion tests is where i spent most of the time understanding foreign code and then later integrated them with my tests.
All in all this week was exciting and a lot of fun! 🐳

Blog Post #4: Write Tests ✅

thisiseshan@gmail.com (Eshan) — Thu, 15 Jul 2021 15:53:06 +0000

Hey friends!
We are mid-way through the GSOC journey and we have put considerable amount of work into it. The project is turning out good and we are very close to the finish line. This is where we enable DRS and move quickly!
Recently another branch was merged into Hub's main that optimized the code and replaced the npz format. So naturally, after last week's checkin I updated the code to run Hub auto smoothly.
Later last week I spent time on 2 major things:

Mid season presentation

Writing tests for Hub auto

Mid-season presentation

This week I presented my project's progress to Activeloop members, highlighting my progress and laying out a plan to wrap up this project in the second half of GSOC.
I think it went great and I received important tips from the members that would enhance the working of Hub auto.
I'll take this opportunity to share my progress on the blog. Dyllan and I have been through multiple iterations of hub auto and we have settled at feature/2.0/image-classification-auto.
I have tested hub.auto on about 200 datasets and the success rate is around 50%, soon things will turn around when I'll integrate handling of multiple file extensions in the coming weeks and this number will shoot to 80-90% 🚀

Writing tests for Hub auto

I wrote tests for hub-auto that makes the testing of hub-auto convenient. Additionally, I have added the "kaggle" functionality to hub-auto that was earlier removed. This api would allow users to download Kaggle datasets, convert them to structured datasets with just one line of code.
To wrap up, this week was super productive and by next week we are hoping to increase the success rate of hub-auto!

Week 4-5 check-in! 🐳

thisiseshan@gmail.com (Eshan) — Thu, 08 Jul 2021 12:47:03 +0000

Hey friends!
Welcome to a new weekly checkin, we have made big progress in our journey of GSOC and this week was pretty significant!

What Did I do this week?

Before we get started, it's story time! Few months ago, my mentor Dyllan had developed a version of hub-auto which worked on an early release of Hub 2.0. This version of hub-auto did the job of structuring datasets and worked locally.
In the past few weeks, Davit, Abhinav, Kristina and Dyllan (all part of Activeloop.ai) have worked on Wasabi Integration, Compression, new index meta and a complete refactor of chunk-engine. This broke the hub-auto code which has not been changed in months. So naturally, this week I upgraded hub-auto to work with the latest release of Hub 2.0! I encountered several bugs but it took me few days to get the hang of what hub-auto is supposed to be doing in the latest release.

What will I do in the coming week?

Now that I have upgraded hub-auto to work with the latest release of Hub 2.0. I am going to use that to upload datasets directly to Hub cloud ☁ What earlier took hours, will now be done in seconds. I had prepared a Notion doc with over 200 datasets and their response to hub-auto. Dyllan asked me to create a spreadsheet with that data. I'm currently working on the spreadsheet part. Once that is ready, we can essentially use all that data to improve hub-auto in the coming weeks.

Where did I get stuck?

While upgrading hub-auto, I fixed several issues with the code. There were many libraries to be dropped and chunks of code to be removed. Particularly to make the code work with the refactored chunk-engine. Fortunately, I wasn't very stuck with bugs like I was used to during the earlier days of contributing to Hub. Experience is the best teacher after all.

Special thanks to my mentor Dyllan, who has been super proactive in helping me improve hub-auto and fixing bugs that were out of my scope. Much respect and love to you man! 💜

Blog Post #3: Update Hub auto to work with compression 🤖

thisiseshan@gmail.com (Eshan) — Thu, 01 Jul 2021 05:47:50 +0000

Hey friends!
Welcome to the 3rd blog post on my Google Summer of Code '21 Journey. This is a follow-up to my last week's blog post. Things have taken an interesting turn with Wasabi integration and Compression being merged into the repo release/2.0.

Previously, I was working on updating feature/2.0/hub-auto branch to work with release/2.0. Now with the latest integration, this process has become as little complex as a lot of changes have been made, for good.

I tried merging feature/2.0/hub-auto into release/2.0 locally. I fixed several merge conflicts however, I got stuck at this error.

I spent few days solving these errors however, my efforts proved futile. I brought this up to my mentor Dyllan and he stated an alternate approach. Since neither Compression/Wasabi nor Hub auto was written by me. It was given that I would have face trouble merging two pieces of code. He also kindly allowed me to take my time with it.

He suggested that I should start with release/2.0 and start from there to work towards Hub 2.0. This approach might take time but would allow me to gain a deep sense of the working of Hub auto. Thus I have begun working on the building Hub auto from the ground up.

Shh! I am also giving myself another day to work on the integration of hub 2.0 if it works great! else I will resolve back to start building it from the ground up.

I will be mainly working on 2 functions:

from_kaggle
from_path

I have begun working on these functions and things are turning out satisfactory. Hope to get in a lot of progress this week! 🚀

Blog Post #2: Research on Datasets 📕

thisiseshan@gmail.com (Eshan) — Thu, 01 Jul 2021 04:43:18 +0000

Hey friends!
Welcome to the 2nd blog post on my Google Summer of Code '21 Journey. It's exciting times, a lot of work is being done on Hub 2.0 and things are looking better than ever! 🚀

How it started

This week, I put myself in the users' shoes. I started using Hub 2.0 to upload some datasets, I was using Hub 1.3.7 however, I encountered this error:

Some fiddling around with my datasets and I couldn't figure out what the problem is. I approached the community members on the Slack channel and they asked me to roll back to Hub version from 1.3.7 to 1.3.5, the issue persisted. This gave me an opportunity to work with the alpha release of Hub 2.0 to solve my error, specifically Hub auto! Hub auto is a feature I am working on for Hub 2.0. It currently works on Image classification tasks thanks to my super mentor Dyllan.

How its going

With the issues I encountered, Dyllan asked me to test Hub auto with datasets. I took this opportunity to document the errors on Notion. Currently, Hub auto works on 3 file extensions [.jpeg, jpg, png] I created a list of all file extensions that do not work and throw in the error
hub.auto does not support the " " extension. Available extensions: ['.jpeg', '.jpg', '.png'].
I went ahead and created a color-coded toggle list that would have the dataset up-top and it would reveal the datasets under the hood. I tested Hub auto with around 200 datasets and made a detail list of all errors encountered as such:

File Extensions

Hub auto errors

Summing up
It's been a productive week, I have utilised my time testing hub auto and I am looking forward to to bring it up to the release/2.0 state next week.

Week 3-4 Research 💭

thisiseshan@gmail.com (Eshan) — Tue, 22 Jun 2021 11:37:02 +0000

Let's take a deep dive into what I've been doing for the past week followed by some tasks for next week!

What did I do this week?

Following up on last week's blog, I was tasked with research on dataset structures and file extensions. I have been through 200 datasets in the past week and documented every bit of it on Notion to later present it to my mentor Dyllan McCreary. This task will help us get super equipped to work on hub-auto.
This week Activeloop organised few CVPR panels on Clubhouse that I actively took part in as an audience. I thoroughly enjoyed the presentation by the speakers and I'm looking forward to more discussions of such kind on ML and Computer Vision discussions in the future.

What will I do next week?

Next week, I have put myself to 2 tasks:

Going through more datasets
Currently, the hub auto branch doesn't work with the compression branch. I am going to experiment and try to make them work together from the ground up!

Did I get stuck anywhere?

I was trying to resolve merge conflicts between Hub release/2.0 branch and feature/2.0/auto. Unfortunately, I wasn't able to find common ground between these features. I might be able to get it done if I start building it from the ground up this week!

Week 2-3 Coding and Research 🧐

thisiseshan@gmail.com (Eshan) — Mon, 14 Jun 2021 13:28:35 +0000

The Python Software Foundation provides a pretty template, so I have decided to use that to answer 3 basic questions about my week's progress.

What did I do this week?

I have been in touch with my mentor Dyllan, we have begun brainstorming on how to approach our project. Last week's work on Index Map was taken over by Dyllan, as it was a little ambiguous and complex for me. Well, I am glad I gave it a shot because no effort hasn't ever paid off.
Next up, I am going through a tonne of Kaggle datasets and figuring out how the structure co-relation. I am using Notability and Notion for highlighting and creating a decent layout of my work.
Here's some Behind the Scenes: 🐳

Notion:

Notability:

What will I do next week?

I plan to go through a lot more datasets to cover as many edge cases as possible. In parallel, I am working on uploading datasets to Hub using its latest Hub 2.0 alpha version which released last week.

Did I get stuck anywhere?

I was unable to upload datasets to Hub, turns out there was a bug which is actively been worked on and I was recommended by the proactive community to use Hub 2.0 to proceed further.

Well, that's all for now. I will write a comprehensive one next week! ✨

Week 1-2 (Community Bonding Period) 🌎

thisiseshan@gmail.com (Eshan) — Mon, 07 Jun 2021 19:05:47 +0000

Hey friends!
Happy to say that I'm off to a great start to Google Summer of Code! The first 2 weeks we have the Community Bonding Period, the idea behind this is to get a good grasp of the codebase and connect with the wonderful people at the organisation.
My community bonding period started with an insightful discussion with my mentor Dyllan McCreary, who walked me through the new codebase Activeloop has been working on for Hub 2.0. He was kind enough to get deep into the important parts of Hub. Dyllan assigned me a few (3) tasks that would help me get familiarised with the codebase.
This session was quite helpful and now that I have tasks assigned to me I thought it was great that I could get a head-start to the summer of code!

Task 1: Depickle the code!

As a general practice it is advised against to use pickle() in your code as it is susceptible to security vulnerabilities. Hence, I was required to replace all occurances of pickle.dumps() and pickle.loads() to something else. (json did the trick!) For more info on why pickle shouldn't be used check this out.

Task 2: Convert IndexMap to a list of IndexMapEntry

This task took the longest to complete. We have been through multiple ways we could do this and have decided to stick to the class method. I was required to convert Indexmap from a list of dictionaries to a list of IndexMapEntry (new class). Thus creating new classes for both IndexMap and IndexMapEntry, followed by writing tests for the same. In the course of this task, Dyllan introduced me to a datatype namedTuples and also helped me get started with parametrising tests. I have implemented the classes and tests however, this broke some part of the other code the team has tirelessly put together :(
I am currently working on it and it should be done by today 🤞🏻

Task 3: Modify Read/Write fixtures for cache in Storage tests

This was a 2-minute task but it helped me get familiar with a few storage things that happen under the hood in Hub 2.0.

All these tasks were designed to give me a solid head start, in hope of making me feel at home when working on my project this summer!

Apart from these tasks we witnessed the launch of the Alpha version of Hub 2.0 on 3rd June 2021. It was a wonderful launch where the team showcased all the shiny features coming to Hub 2.0 with a progress report on what is implemented. The results displayed more than a 6x increase in performance compared to Hub 1.0. Activeloop's CEO Davit Buniatyan's presentation showed what the team has achieved with Hub 2.0 is truly remarkable!
Things are just getting started and I am beyond excited for what is to come! ♥️

Welcome to my GSoC journey ✨

thisiseshan@gmail.com (Eshan) — Mon, 07 Jun 2021 18:44:39 +0000

Hello there!
I am Eshan Arora (@thisiseshan everywhere on the internet), a senior year computer engineering undergrad at NIT Surat, India.
Welcome to my blog, I am thrilled to share that I will be working on Hub 2.0 by Activeloop.ai this summer!

Hub aims to reduce time spent by the researchers figuring out data by enabling dataset streaming, thus allowing data scientists to spend more of their time on building epic Machine Learning models.

My project this summer is to implement Automatic generation of Schema at Hub 2.0, which would enable any kind of dataset to be seamlessly stored at Hub with just a single line of code.
I am excited to be given this opportunity and aim to bring the best out of it this summer!

You can checkout Hub here
For Hub 2.0 checkout the release/2.0 branch! 🚀