# 9th place solution: finetune CLIP ViT-H/14

First of all, I thank the organizers for this great competition!
My solution is heavily based on
[@motono0223](https://www.kaggle.com/motono0223)
's work. Thank you very much!
[https://www.kaggle.com/code/motono0223/guie-clip-tensorflow-train-example](https://www.kaggle.com/code/motono0223/guie-clip-tensorflow-train-example)
[* My code is now public *](https://www.kaggle.com/code/akihirok/9th-place-guie-fintune-tf-clip-trainable/notebook)
Key improvements over the original work
Input 64D vectors to ArcFace (instead of 256D ones) (~ 0.01 gain)
[Open CLIP ViT-H/14 on LAION-2B](https://laion.ai/blog/large-openclip/)
(~ 0.08 gain)
Reduce LR linearly (~ 0.01 gain)
Finetune backbone CLIP model for 1 epoch w/ low LR w/o DA (~ 0.02 gain)
1 epoch = 10 epochs in the
[@motono0223](https://www.kaggle.com/motono0223)
's work
Dataset
I indeed thought that finding nice datasets was the most important thing to win the competition.
However, I eventually used the same datasets as the original work.
I couldn't find any other dataset which works well. 😭
Products-10K > Landmark Retrieval 2021 > ImageNet-1K > Deep Fashion = Food Recognition > others (in my best knowledge)
Note
I suffered from random OOM errors in submission (with CLIP H/14). I gave up more complex architecture because of this. (why you can do an ensemble or TTA? 🤔)
Thank you for reading!