{
  "id": 359497,
  "title": "12th Solution",
  "content": "<p>Thank to host for hosting a great competiton, hard work with my team member ( <a href=\"https://www.kaggle.com/hycloud\" target=\"_blank\">@hycloud</a> ) and competitors.<br>\nIt's great competition. I have learned many things from this competition and solutions. that's very happy.  </p>\n<p>We write our 12th solution as follows.</p>\n<h2>Dataset</h2>\n<p>Data 1: </p>\n<ul>\n<li>GLDv2-Clean</li>\n<li>Met-Art</li>\n<li>Products-10k</li>\n<li>Shopee-Products</li>\n<li>Food-1k</li>\n<li>DeepFashion</li>\n<li>Standford-Cars</li>\n<li>Storefronts (shared public kaggle dataset)</li>\n<li>Furniture (shared public kaggle dataset).</li>\n</ul>\n<p>Data 2:</p>\n<ul>\n<li>GLDv2 Full</li>\n<li>Met-Art</li>\n<li>Products-10k</li>\n<li>Shopee-Products</li>\n<li>Public Dataset</li>\n<li>rp2k</li>\n<li>product10k</li>\n<li>deepfashion</li>\n<li>fashion 200k</li>\n<li>myautoge</li>\n<li>Storefronts (shared public kaggle dataset)</li>\n<li>ifuniture</li>\n<li>stanford_online_shop</li>\n</ul>\n<h2>Model</h2>\n<table>\n<thead>\n<tr>\n<th>Model</th>\n<th>Loss</th>\n<th>Neck</th>\n<th>Data</th>\n<th>Score(Private/Public)</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>ViT-Huge-224</td>\n<td>Arcface with s=30</td>\n<td>1024-4096-BN-PReLU-64</td>\n<td>Data1</td>\n<td>0.629/0.609</td>\n</tr>\n<tr>\n<td>ViT-Huge-224</td>\n<td>Multi Similarity Loss + CrossBatchMemory</td>\n<td>1024-4096-BN-PReLU-64</td>\n<td>Data2</td>\n<td>0.624/0.622</td>\n</tr>\n<tr>\n<td>Swin-Large-384</td>\n<td>Arcface with s=30</td>\n<td>1536-6144-BN-PReLU-64-BN-PReLU</td>\n<td>Data1</td>\n<td>0.612/0.602</td>\n</tr>\n<tr>\n<td>Swin-Large-224</td>\n<td>Arcface with s=30</td>\n<td>1536-6144-BN-PReLU-64-BN-PReLU</td>\n<td>Data1</td>\n<td>0.576/0.562</td>\n</tr>\n</tbody>\n</table>\n<p>Freeze backbone and train only projector for ViT-H; Two-stage for Swin, first stage freeze backbone second stage fine-tune whole model; 10 to 20 epochs for first stage, 1 to 5 epochs for second stage.<br>\nOptimizer: SGD, lr 0.1 for first stage, 0.001 for second stage.(ViT-H used Adam)</p>\n<h1>Ensemble</h1>\n<p>We used ensemble for a final layer using PCA.<br>\nWe pickup competition distributed dataset for PCA dataset. that's important point. <br>\nfinal result is pickup PCA dataset &gt; full dataset.</p>\n<p>Final Score is 0.667/0.654(Private/Public).</p>\n<h1>Didn't worked</h1>\n<ul>\n<li>Stable Diffusion(storefront generation</li>\n<li>CLIP 2stage training(Backbone freeze-&gt;full)</li>\n</ul>",
  "comments": [
    {
      "id": 1983871,
      "author_name": "hyc",
      "author_url": "",
      "post_date": "2022-10-12T09:58:25.127000",
      "content": "<p>Thanks for the hard work <a href=\"https://www.kaggle.com/tereka\" target=\"_blank\">@tereka</a> , it was fun and a great experience.</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1983951,
          "author_name": "tereka",
          "author_url": "",
          "post_date": "2022-10-12T11:09:44.917000",
          "content": "<p>thanks! it's exciting competition. It's fun too.</p>",
          "votes": 0,
          "replies": []
        }
      ]
    }
  ],
  "comment_count": 2,
  "messages": [
    {
      "id": 1983868,
      "postDate": "2022-10-12T09:55:58.397Z",
      "content": "<p>Thank to host for hosting a great competiton, hard work with my team member ( <a href=\"https://www.kaggle.com/hycloud\" target=\"_blank\">@hycloud</a> ) and competitors.<br>\nIt's great competition. I have learned many things from this competition and solutions. that's very happy.  </p>\n<p>We write our 12th solution as follows.</p>\n<h2>Dataset</h2>\n<p>Data 1: </p>\n<ul>\n<li>GLDv2-Clean</li>\n<li>Met-Art</li>\n<li>Products-10k</li>\n<li>Shopee-Products</li>\n<li>Food-1k</li>\n<li>DeepFashion</li>\n<li>Standford-Cars</li>\n<li>Storefronts (shared public kaggle dataset)</li>\n<li>Furniture (shared public kaggle dataset).</li>\n</ul>\n<p>Data 2:</p>\n<ul>\n<li>GLDv2 Full</li>\n<li>Met-Art</li>\n<li>Products-10k</li>\n<li>Shopee-Products</li>\n<li>Public Dataset</li>\n<li>rp2k</li>\n<li>product10k</li>\n<li>deepfashion</li>\n<li>fashion 200k</li>\n<li>myautoge</li>\n<li>Storefronts (shared public kaggle dataset)</li>\n<li>ifuniture</li>\n<li>stanford_online_shop</li>\n</ul>\n<h2>Model</h2>\n<table>\n<thead>\n<tr>\n<th>Model</th>\n<th>Loss</th>\n<th>Neck</th>\n<th>Data</th>\n<th>Score(Private/Public)</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>ViT-Huge-224</td>\n<td>Arcface with s=30</td>\n<td>1024-4096-BN-PReLU-64</td>\n<td>Data1</td>\n<td>0.629/0.609</td>\n</tr>\n<tr>\n<td>ViT-Huge-224</td>\n<td>Multi Similarity Loss + CrossBatchMemory</td>\n<td>1024-4096-BN-PReLU-64</td>\n<td>Data2</td>\n<td>0.624/0.622</td>\n</tr>\n<tr>\n<td>Swin-Large-384</td>\n<td>Arcface with s=30</td>\n<td>1536-6144-BN-PReLU-64-BN-PReLU</td>\n<td>Data1</td>\n<td>0.612/0.602</td>\n</tr>\n<tr>\n<td>Swin-Large-224</td>\n<td>Arcface with s=30</td>\n<td>1536-6144-BN-PReLU-64-BN-PReLU</td>\n<td>Data1</td>\n<td>0.576/0.562</td>\n</tr>\n</tbody>\n</table>\n<p>Freeze backbone and train only projector for ViT-H; Two-stage for Swin, first stage freeze backbone second stage fine-tune whole model; 10 to 20 epochs for first stage, 1 to 5 epochs for second stage.<br>\nOptimizer: SGD, lr 0.1 for first stage, 0.001 for second stage.(ViT-H used Adam)</p>\n<h1>Ensemble</h1>\n<p>We used ensemble for a final layer using PCA.<br>\nWe pickup competition distributed dataset for PCA dataset. that's important point. <br>\nfinal result is pickup PCA dataset &gt; full dataset.</p>\n<p>Final Score is 0.667/0.654(Private/Public).</p>\n<h1>Didn't worked</h1>\n<ul>\n<li>Stable Diffusion(storefront generation</li>\n<li>CLIP 2stage training(Backbone freeze-&gt;full)</li>\n</ul>",
      "rawMarkdown": "Thank to host for hosting a great competiton, hard work with my team member ( @hycloud ) and competitors.\nIt's great competition. I have learned many things from this competition and solutions. that's very happy.  \n\nWe write our 12th solution as follows.\n\n## Dataset\nData 1: \n- GLDv2-Clean\n- Met-Art\n- Products-10k\n- Shopee-Products\n- Food-1k\n- DeepFashion\n- Standford-Cars\n- Storefronts (shared public kaggle dataset)\n- Furniture (shared public kaggle dataset).\n\nData 2:\n- GLDv2 Full\n- Met-Art\n- Products-10k\n- Shopee-Products\n- Public Dataset\n- rp2k\n- product10k\n- deepfashion\n- fashion 200k\n- myautoge\n- Storefronts (shared public kaggle dataset)\n- ifuniture\n- stanford_online_shop\n\n## Model\n|Model|Loss|Neck|Data|Score(Private/Public)|\n|----|----|----|----|----|\n|ViT-Huge-224|Arcface with s=30|1024-4096-BN-PReLU-64|Data1|0.629/0.609|\n|ViT-Huge-224|Multi Similarity Loss + CrossBatchMemory|1024-4096-BN-PReLU-64|Data2|0.624/0.622|\n|Swin-Large-384|Arcface with s=30|1536-6144-BN-PReLU-64-BN-PReLU|Data1|0.612/0.602|\n|Swin-Large-224|Arcface with s=30|1536-6144-BN-PReLU-64-BN-PReLU|Data1|0.576/0.562|\n\nFreeze backbone and train only projector for ViT-H; Two-stage for Swin, first stage freeze backbone second stage fine-tune whole model; 10 to 20 epochs for first stage, 1 to 5 epochs for second stage.\nOptimizer: SGD, lr 0.1 for first stage, 0.001 for second stage.(ViT-H used Adam)\n\n# Ensemble\nWe used ensemble for a final layer using PCA.\nWe pickup competition distributed dataset for PCA dataset. that's important point. \nfinal result is pickup PCA dataset > full dataset.\n\nFinal Score is 0.667/0.654(Private/Public).\n\n# Didn't worked\n- Stable Diffusion(storefront generation\n- CLIP 2stage training(Backbone freeze->full)",
      "votes": 6
    },
    {
      "id": 1983871,
      "postDate": "2022-10-12T09:58:25.127Z",
      "content": "<p>Thanks for the hard work <a href=\"https://www.kaggle.com/tereka\" target=\"_blank\">@tereka</a> , it was fun and a great experience.</p>",
      "rawMarkdown": "Thanks for the hard work @tereka , it was fun and a great experience.",
      "votes": 1,
      "replies": [
        {
          "id": 1983951,
          "postDate": "2022-10-12T11:09:44.917Z",
          "content": "<p>thanks! it's exciting competition. It's fun too.</p>",
          "rawMarkdown": "thanks! it's exciting competition. It's fun too."
        }
      ]
    }
  ],
  "complete": true,
  "source": "live",
  "fetched_at": "2026-10-24T13:24:25Z",
  "competition_id": 36414,
  "competition_slug": "google-universal-image-embedding",
  "team_id": 8882478,
  "solution_url": "/competitions/google-universal-image-embedding/writeups/th-12th-solution"
}