Welcome to the official code and experiment hub for the DeepLearning MegaThread — a 5-part blog series by the GDGC ML Team that dives into the practical side of deep learning with a perfect blend of theory, code, and good vibes. 🚀
We’re talking hands-on implementation, clear explanations, and models that actually work — all wrapped into weekly drops designed to make deep learning more accessible, exciting, and deployable.
Author: Anaant Raj
Anaant kicked things off with a banger 💥 by walking us through how to fine-tune MobileNetV2 on CIFAR-10.
From loading pre-trained weights to understanding which layers to freeze, this post is a must-read if you’re diving into transfer learning.
Author: Nidhi Rohra
Nidhi broke down the GPT-2 architecture and the entire Transformer revolution with crystal-clear visuals and that golden example:
“In ‘The cat sat on the mat,’ the word ‘sat’ pays attention to both ‘cat’ and ‘mat.’” 🤯
She also included runnable code and a clean walkthrough of why decoder-only models like GPT-2 are so powerful.
Author: Arihant Bhandari
CNNs are great at extracting local features. Transformers are amazing at global reasoning. So... why not both?
MobileViT does just that — blending MobileNet’s efficiency with ViT’s attention mechanism for a hybrid model that punches way above its size.
We experiment with MobileViT-XXS under multiple settings using Tiny ImageNet to test how image resolution, pretraining, and architecture affect performance.
Author: Samyak Waghdare
This one’s all about securing GPT-like models against manipulation. Samyak walks you through how adversarial prompts work, how to generate them using libraries like TextAttack, and how adversarial training can improve model robustness — without sacrificing performance.
Author: Sahil Garje
In this hilarious and thought-provoking finale, Sahil takes on the challenge of humanizing GPT output — testing dozens of prompting techniques and measuring them using GPTZero, a state-of-the-art AI detector. It's prompt engineering meets performance art.
All the code, experiments, and notebooks from the series — neatly organized for you to run, tweak, and learn from.
📂 Folder: VisionGPT/
| Notebook | Description |
|---|---|
mobilevit_xxs_scratch_64.ipynb |
Trains MobileViT-XXS from scratch on 64×64 Tiny ImageNet images. |
mobilevit_xxs_pretrained_64.ipynb |
Fine-tunes pretrained MobileViT-XXS on 64×64 Tiny ImageNet. |
mobilevit_xxs_scratch_96.ipynb |
Trains MobileViT-XXS from scratch on 96×96 resolution input. |
mobilevit_xxs_pretrained_96.ipynb |
Fine-tunes pretrained MobileViT-XXS on 96×96 Tiny ImageNet. |
📂 Folder: GPTAdversarial/
| Notebook | Description |
|---|---|
gdgc-llm-adversarial.ipynb |
Core training + evaluation pipeline for adversarial robustness. |
gdgc-llm-adversarial-training-fgsm.ipynb |
Implements FGSM adversarial training on LLMs. |
gdgc-llm-adversarial-training-text-attack-deepbugword.ipynb |
Uses TextAttack with the DeepBugWord attack method. |
gdgc-llm-adversarial-training-text-attack-textfooler.ipynb |
Uses TextAttack with the TextFooler method. |
Because let’s face it — deep learning can feel like a black box sometimes.
This series is our way of turning that box transparent.
- 🤖 From pre-trained models to attention mechanisms
- 🧠 From “how it works” to “why it matters”
- 💻 From paper to practice
Whether you're a beginner or a practitioner brushing up on fundamentals, we’ve got something here for you.
🧠 Arihant Bhandari
🧠 Sahil Garje
🧠 Anaant Raj
🧠 Nidhi Rohra
🧠 Samyak Waghdare
We're the GDGC ML Team — on a mission to make deep learning more practical, fun, and open to all.
Spotted a bug, want to contribute, or just want to say hi?
Open an issue, fork the repo, or comment on the blog posts — we love hearing from fellow learners and builders.
🧠 Happy Learning from the GDGC ML Team!
🎉 This wraps up our 5-part DeepLearning MegaThread series — thanks for following along!
We hope it sparked ideas, curiosity, and a love for building. Until next time! 💙