Some stream of thoughts wrt ConvNets and ViTs.
The paper "Patches Are All You Need?" contends that patchification has a special place into why ViTs are so effective, while others have said it's the expressiveness of the attention mechanism, or the expressiveness of multiplicative interactions, or the fact that many operations are data dependent (like, in a spectrum with weights-dependent on the other end) 1/n
#deeplearning #vit #attention #cnn

The #ViT journey:

Vision transformers were easier than I expected. I planned to spend three weeks learning them, but I'm already done :ablobcatrave: .

The #ViT journey:

'Natural Language Processing Demystified' is my all-time favorite course in #NLP, it also covers transformers, check it out:

https://www.nlpdemystified.org/course/transformers

The #ViT journey:

Let's visit Jay Alammar's blog once again and check out his amazing Illustrated #Transformer post. Even if you're already familiar with transformers, it's definitely worth taking a look: https://jalammar.github.io/illustrated-transformer/

The #ViT journey:

I need to spend some time studying the multi-head self-attention mechanism. I'm going to start with this fantastic blog post by Jay Alammar: 'Visualizing A Neural Machine Translation Model (Mechanics of Seq2seq Models With Attention)'.

Check it out here: https://jalammar.github.io/visualizing-neural-machine-translation-mechanics-of-seq2seq-models-with-attention/

The #ViT journey:

This is one of the best step-by-step paper implementation tutorials I've ever encountered. The video version on Udemy, which is over eight hours long :ablobwhee: , was also a real treat to watch. If you want to learn how to replicate Vision Transformers in PyTorch, I highly recommend it:

https://www.learnpytorch.io/08_pytorch_paper_replicating/"

The #ViT journey:

"The key contributions from [AN IMAGE IS WORTH 16X16 WORDS] paper were not in terms of a new architecture, but rather the application of an existing architecture (Transformers), to the field of Computer Vision. It is the training method and the dataset used to pretrain the network, that were key for ViT to get excellent results compared to SOTA (State of the Art) on ImageNet."
https://amaarora.github.io/2021/01/18/ViT.html

So I have to take a step back and learn about #Transformers :woohoo:

I'll spend the next three weeks fully dedicated to learning about Vision Transformers and implementing the #ViT paper in PyTorch from scratch. Let's hope that this will be sufficient time to complete the task! :thisisfine:

Which B/16 reigns supreme? I've recently fine-tuned quite a few new ViT models and wanted to compare them. With new multi-weight support on the way I realized timm will soon have ~20 different B/16 (or close to). B/16 is the most common ViT model and easiest to compare across wide range of pretrain datasets and methods. In the lead is BEiT v2, but hot on its heals are fine-tuned LAION2B and OpenAI CLIP image towers. Check out a notebook at https://colab.research.google.com/drive/12u1csH7_Uun78lGti35zvi5-S6FX4ZKu?usp=sharing #CV #machinelearning #vit #AI

மாணவர்கள் கண்டுபிடித்த ரோபோ மூலம் ஆயுத பூஜை கொண்டாடிய விஐடி – வீடியோ
https://patrikai.com/vit-celebrates-ayudha-puja-with-student-invention-robot-video/ via @patrikaidotcom@twitter.com

#VIT #Robotics #robots #AyudhaPooja2022 #AyudhaPuja #AUTOMATION #AutomateUpdates #ViralVideo @VIT_univ@twitter.com @GraVITas_VIT@twitter.com

வண்டலூர் விஐடி பல்கலைக்கழகத்தில் மேலும் 45 பேருக்கு கொரோனா உறுதி… https://patrikai.com/corona-confirms-45-more-students-at-vandalur-vit-university/ via @patrikaidotcom@twitter.com

#Vandalur #VIT #University #COVID19 #COVID

@uint8_t

By urgency as a calculation of importance, impact, benefit, dependencies, due date and age.

I use #Taskwarrior
https://taskwarrior.org/
A command line task manager.

Recommendable Vim-alike frontend: #vit
https://github.com/scottkosty/vit

Loads of other tools available
https://taskwarrior.org/tools/

Along with #Timewarrior to track time spent.
https://timewarrior.net/