Again, thanks to Kaggle and the hosts for hosting an interesting competition. Congratulations to all the winners. Special thanks to all my teammates @takuok and @tascj0 . Fortunately, we got the gold medal of Recognition competition, too. We show the details of our models in Retrieval solution post.
\nHere, we show the unique part of Recognition challenge.
We used the same two strategies as the last year's 1st team used. While submitting, we used pre-computed embedding of train4.1m data, penalized these embeddings that are similar to the non-landmark images. These methods are again very robust this year and critical for the gold medal.
\nThis approach boost the score for a single model from public/private 0.37843/0.36045 to 0.45724/0.43252.
\nFinal score is ensemble of beit, swin, b4, b6, xcit, v2m, b5. See the details about models in Retrieval post.
takuoko is a member of Z by HP & NVIDIA Data Science Global Ambassadors.
\nSpecial Thanks to Z by HP & NVIDIA for sponsoring me a Z8G4 Workstation with dual RTX6000 GPU and a ZBook with RTX5000 GPU.
\nThis competition has the big dataset.
\nSo I tried pytorch's DDP parallel training on my dual RTX6000 GPUs and it helped a lot.
Again, thanks to Kaggle and the hosts for hosting an interesting competition. Congratulations to all the winners. Special thanks to all my teammates @takuok and @tascj0 . Fortunately, we got the gold medal of Recognition competition, too. We show the details of our models in Retrieval solution post.
\nHere, we show the unique part of Recognition challenge.
We used the same two strategies as the last year's 1st team used. While submitting, we used pre-computed embedding of train4.1m data, penalized these embeddings that are similar to the non-landmark images. These methods are again very robust this year and critical for the gold medal.
\nThis approach boost the score for a single model from public/private 0.37843/0.36045 to 0.45724/0.43252.
\nFinal score is ensemble of beit, swin, b4, b6, xcit, v2m, b5. See the details about models in Retrieval post.
takuoko is a member of Z by HP & NVIDIA Data Science Global Ambassadors.
\nSpecial Thanks to Z by HP & NVIDIA for sponsoring me a Z8G4 Workstation with dual RTX6000 GPU and a ZBook with RTX5000 GPU.
\nThis competition has the big dataset.
\nSo I tried pytorch's DDP parallel training on my dual RTX6000 GPUs and it helped a lot.
Congrats and thanks for the write-up. I also used DDP, but on AWS instance with 4xT4 GPU
", "rawMarkdown": "Congrats and thanks for the write-up. I also used DDP, but on AWS instance with 4xT4 GPU", "votes": null }, { "id": "1538198", "postDate": "10/08/2021 06:48:16", "content": "@inoueu1 Thanks for the tips. Congratulations.
", "rawMarkdown": "inoueu1 Thanks for the tips. Congratulations.", "votes": null }, { "id": "1540016", "postDate": "10/10/2021 04:45:06", "content": "Thanks for sharing.
", "rawMarkdown": "Thanks for sharing.", "votes": null } ], "comments": [ { "id": 1535855, "author_name": "narsil", "author_url": "", "post_date": "10/06/2021 08:47:56", "content": "Congrats and thanks for the write-up. I also used DDP, but on AWS instance with 4xT4 GPU
", "votes": null, "replies": [] }, { "id": 1538198, "author_name": "towhidultonmoy", "author_url": "", "post_date": "10/08/2021 06:48:16", "content": "@inoueu1 Thanks for the tips. Congratulations.
", "votes": null, "replies": [] }, { "id": 1540016, "author_name": "saleesh", "author_url": "", "post_date": "10/10/2021 04:45:06", "content": "Thanks for sharing.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1535237": "Again, thanks to Kaggle and the hosts for hosting an interesting competition. Congratulations to all the winners. Special thanks to all my teammates @takuok and @tascj0 . Fortunately, we got the gold medal of Recognition competition, too. We show the details of our models in [Retrieval solution post](https://www.kaggle.com/c/landmark-retrieval-2021/discussion/276632).\nHere, we show the unique part of Recognition challenge.\n\n# Tips in Recognition.\n\nWe used the same two strategies as the last year's 1st team used. While submitting, we used pre-computed embedding of train4.1m data, penalized these embeddings that are similar to the non-landmark images. These methods are again very robust this year and critical for the gold medal.\nThis approach boost the score for a single model from **public/private 0.37843/0.36045** to **0.45724/0.43252**.\nFinal score is **ensemble of beit, swin, b4, b6, xcit, v2m, b5**. See the details about models in [Retrieval post](https://www.kaggle.com/c/landmark-retrieval-2021/discussion/276632).\n\n# Acknowledge\ntakuoko is a member of Z by HP & NVIDIA Data Science Global Ambassadors.\nSpecial Thanks to Z by HP & NVIDIA for sponsoring me a Z8G4 Workstation with dual RTX6000 GPU and a ZBook with RTX5000 GPU.\nThis competition has the big dataset.\nSo I tried pytorch's DDP parallel training on my dual RTX6000 GPUs and it helped a lot.", "1535855": "Congrats and thanks for the write-up. I also used DDP, but on AWS instance with 4xT4 GPU", "1538198": "inoueu1 Thanks for the tips. Congratulations.", "1540016": "Thanks for sharing." }, "source": "meta" }