vit-pytorch VS memory-efficient-attention-pytorch

I also need to learn Pyspark so right now I am going to download the Fashion Mnist dataset, use Pyspark to downsize each image and put the into separate folders according to their labels (just to show employers I can do some basic ETL with Pyspark, not sure how I am going to load for training in Pytorch yet though). Then I am going to write the simplest Le Net to try to categorize the fashion MNIST dataset (results will most likely be bad but it's okay). Next, try to learn transfer learning in Pytorch for both CNN or maybe skip ahead to ViT. Ideally at this point I want to study the Attention mechanism a bit more and try to implement Simple Vit which I saw here: https://github.com/lucidrains/vit-pytorch/blob/main/vit_pytorch/simple_vit.py

For code implementation, check out this git repo. It contains fairly straightforward PyTorch implementations for various ViT papers with references.

For ViT’s specifically, there’s been a good amount of research trying to extend ViT’s to work on small datasets without a large amount of pre-training (which comes with its own host of issues such as the best way to fine tune such a huge model). One paper which comes to mind is ViT’s for small datasets (https://arxiv.org/abs/2112.13492), which has an implementation in lucidrain’s repo here: https://github.com/lucidrains/vit-pytorch

Here's a pytorch implementation of ViT https://github.com/lucidrains/vit-pytorch

looks great Junho! i've linked to it from https://github.com/lucidrains/vit-pytorch like you asked :)

Sure thing. Also if you're getting into transformers I'd recommend lucidrains's GitHub[0] since it has a large collection of them with links to papers. It's nice that things are consolidated.

[0] https://github.com/lucidrains/vit-pytorch

nah, it is really simple. here is the code https://github.com/lucidrains/vit-pytorch/blob/main/vit_pytorch/mae.py

lucid rains VI has all of what you may need in a clean API

I've attempted it here https://github.com/lucidrains/vit-pytorch/blob/main/vit_pytorch/rvt.py but those who have tried it haven't seen knock out results as 1d. Perhaps the axial lengths are too small to see a benefit

Check these efficient attention mechanism which are almost a drop in replacement: efficient attention flash attention

I would recommend Routing Transformer https://github.com/lucidrains/routing-transformer but the real truth is nothing beats full attention. Luckily, someone recently figured out how to get past the memory bottleneck. https://github.com/lucidrains/memory-efficient-attention-pyt...