First of all, I thank the organizers for this great competition!
\nMy solution is heavily based on @motono0223's work. Thank you very much!
\nhttps://www.kaggle.com/code/motono0223/guie-clip-tensorflow-train-example
I indeed thought that finding nice datasets was the most important thing to win the competition.
\nHowever, I eventually used the same datasets as the original work.
\nI couldn't find any other dataset which works well. ðŸ˜
Products-10K > Landmark Retrieval 2021 > ImageNet-1K > Deep Fashion = Food Recognition > others (in my best knowledge)
\nThank you for reading!
", "comments": [ { "id": 1983846, "author_name": "RocketScience", "author_url": "", "post_date": "2022-10-12T09:25:57.683000", "content": "What I found is that ensembling multiple CLIP backbones togther via tensorflow may throw exactly same error as more complicated CLIP structures, while top ensemble-based solutions provided in PyTorch can be smoothly evaluated and now I wonder what is actually happening.
", "votes": 1, "replies": [ { "id": 1983849, "author_name": "Akihiro Katsura", "author_url": "", "post_date": "2022-10-12T09:34:52.300000", "content": "PyTorch was all I needed 😂
", "votes": 1, "replies": [] } ] }, { "id": 1983030, "author_name": "Nanashi", "author_url": "", "post_date": "2022-10-11T18:38:20.300000", "content": "@akihirok congrats! we have a similar setup with more tricks :) Interesting that all your code is apparently tensorflow-based
", "votes": 1, "replies": [] }, { "id": 1983501, "author_name": "Wenxuan Ye", "author_url": "", "post_date": "2022-10-12T05:09:07.590000", "content": "How do you solve the problem:\"ResourceExhaustedError: received trailing metadata size exceeds limit\"
", "votes": 0, "replies": [ { "id": 1983825, "author_name": "Akihiro Katsura", "author_url": "", "post_date": "2022-10-12T09:04:14.860000", "content": "I didn't experience that actually.
\nTry the notebook I shared 😃
First of all, I thank the organizers for this great competition!
\nMy solution is heavily based on @motono0223's work. Thank you very much!
\nhttps://www.kaggle.com/code/motono0223/guie-clip-tensorflow-train-example
I indeed thought that finding nice datasets was the most important thing to win the competition.
\nHowever, I eventually used the same datasets as the original work.
\nI couldn't find any other dataset which works well. ðŸ˜
Products-10K > Landmark Retrieval 2021 > ImageNet-1K > Deep Fashion = Food Recognition > others (in my best knowledge)
\nThank you for reading!
", "rawMarkdown": "First of all, I thank the organizers for this great competition!\n\nMy solution is heavily based on @motono0223's work. Thank you very much!\nhttps://www.kaggle.com/code/motono0223/guie-clip-tensorflow-train-example\n\n[* My code is now public *](https://www.kaggle.com/code/akihirok/9th-place-guie-fintune-tf-clip-trainable/notebook)\n\n## Key improvements over the original work\n\n- Input 64D vectors to ArcFace (instead of 256D ones) (~ 0.01 gain)\n- [Open CLIP ViT-H/14 on LAION-2B](https://laion.ai/blog/large-openclip/) (~ 0.08 gain)\n- Reduce LR linearly (~ 0.01 gain)\n- Finetune backbone CLIP model for 1 epoch w/ low LR w/o DA (~ 0.02 gain)\n - 1 epoch = 10 epochs in the @motono0223's work\n\n## Dataset\n\nI indeed thought that finding nice datasets was the most important thing to win the competition.\nHowever, I eventually used the same datasets as the original work.\nI couldn't find any other dataset which works well. ðŸ˜\n\nProducts-10K > Landmark Retrieval 2021 > ImageNet-1K > Deep Fashion = Food Recognition > others (in my best knowledge)\n\n## Note\n\n- 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? 🤔)\n\nThank you for reading!", "votes": 20 }, { "id": 1983846, "postDate": "2022-10-12T09:25:57.683Z", "content": "What I found is that ensembling multiple CLIP backbones togther via tensorflow may throw exactly same error as more complicated CLIP structures, while top ensemble-based solutions provided in PyTorch can be smoothly evaluated and now I wonder what is actually happening.
", "rawMarkdown": "What I found is that ensembling multiple CLIP backbones togther via tensorflow may throw exactly same error as more complicated CLIP structures, while top ensemble-based solutions provided in PyTorch can be smoothly evaluated and now I wonder what is actually happening.", "votes": 1, "replies": [ { "id": 1983849, "postDate": "2022-10-12T09:34:52.300Z", "content": "PyTorch was all I needed 😂
", "rawMarkdown": "PyTorch was all I needed 😂", "votes": 1 } ] }, { "id": 1983030, "postDate": "2022-10-11T18:38:20.300Z", "content": "@akihirok congrats! we have a similar setup with more tricks :) Interesting that all your code is apparently tensorflow-based
", "rawMarkdown": "@akihirok congrats! we have a similar setup with more tricks :) Interesting that all your code is apparently tensorflow-based", "votes": 1 }, { "id": 1983501, "postDate": "2022-10-12T05:09:07.590Z", "content": "How do you solve the problem:\"ResourceExhaustedError: received trailing metadata size exceeds limit\"
", "rawMarkdown": "How do you solve the problem:\"ResourceExhaustedError: received trailing metadata size exceeds limit\"", "replies": [ { "id": 1983825, "postDate": "2022-10-12T09:04:14.860Z", "content": "I didn't experience that actually.
\nTry the notebook I shared 😃