{
  "id": 49291,
  "title": "Some shakeup , congratulations and thank yoh Gleb",
  "url": "/competitions/sp-society-camera-model-identification/discussion/49291",
  "author_name": "",
  "post_date": "2018-02-09T00:27:42.421768500Z",
  "votes": 6,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Quite interesting the shake up. The last 2 days my rank on the public LB was dropping from rank 70 to rank 104 and now I jumped up to rank 59. On the public LB my best model (resnet 50 with 256 * 256 random crops) achieved 0.960 and on the private LB 0.966 - which matches quite well my own accuracy metric. Seems that quite a number of people overfitted the public LB.  </p>\n\n<p>Thanks to Gleb for his extra dataset - it helped a lot.  </p>",
  "messages": [
    {
      "id": "279924",
      "postDate": "02/09/2018 00:27:42",
      "content": "<p>Quite interesting the shake up. The last 2 days my rank on the public LB was dropping from rank 70 to rank 104 and now I jumped up to rank 59. On the public LB my best model (resnet 50 with 256 * 256 random crops) achieved 0.960 and on the private LB 0.966 - which matches quite well my own accuracy metric. Seems that quite a number of people overfitted the public LB.  </p>\n\n<p>Thanks to Gleb for his extra dataset - it helped a lot.  </p>",
      "rawMarkdown": "Quite interesting the shake up. The last 2 days my rank on the public LB was dropping from rank 70 to rank 104 and now I jumped up to rank 59. On the public LB my best model (resnet 50 with 256 * 256 random crops) achieved 0.960 and on the private LB 0.966 - which matches quite well my own accuracy metric. Seems that quite a number of people overfitted the public LB.  \n\nThanks to Gleb for his extra dataset - it helped a lot.",
      "votes": null
    },
    {
      "id": "279988",
      "postDate": "02/09/2018 03:14:36",
      "content": "<p>Well played!</p>\n\n<blockquote>\n  <p>On the public LB my best model (resnet 50 with 256 * 256 random crops)\n  achieved 0.960 and on the private LB 0.966</p>\n</blockquote>\n\n<p>Did you use data augmentation? If so, what about TTA?</p>\n\n<p>Did you use any learning rate schedulers?</p>\n\n<p>Did you use random search, or anything to search the hyperparameter space?</p>\n\n<p>(The thought on my mind, \"how did he do it with just a resnet50 and 256x256 random crops?! Did he center crop at all? Isn't the information outside of the 512x512 center outside of the test distribution (ignoring resizing)?\")</p>",
      "rawMarkdown": "Well played!\n\n&gt; On the public LB my best model (resnet 50 with 256 * 256 random crops)\n&gt; achieved 0.960 and on the private LB 0.966\n\nDid you use data augmentation? If so, what about TTA?\n\nDid you use any learning rate schedulers?\n\nDid you use random search, or anything to search the hyperparameter space?\n\n(The thought on my mind, \"how did he do it with just a resnet50 and 256x256 random crops?! Did he center crop at all? Isn't the information outside of the 512x512 center outside of the test distribution (ignoring resizing)?\")",
      "votes": null
    },
    {
      "id": "280266",
      "postDate": "02/09/2018 17:07:34",
      "content": "<p>random crops for training, data augmentation,  test : crop in the middle. I used lowering learning rate bei 30% when plataueing for 10 epocs.  TTA with flipping.  Batch size 64, initial Lr 1.0e-4.  Hardware: 4 Titan X GPUs, Dual Xeon server with SSDs and 512GB RAM.</p>\n\n<p>I played a lot with learning rates and other networks. DenseNet201, DenseNet169 and small own networks.  I tried ensembling too but my ensembling code has a bug and I was able to fix it at the last day  - but to late:( But I got a flu and it was quite hard to work on this project for the last 2 weeks. </p>\n\n<p>I m quite happy to achieve 0.966 accuracy with a single model with only 5270 training images and I m sure I could have acchieved an even higher accuracy with a single model and much higher with ensembling.</p>",
      "rawMarkdown": "random crops for training, data augmentation,  test : crop in the middle. I used lowering learning rate bei 30% when plataueing for 10 epocs.  TTA with flipping.  Batch size 64, initial Lr 1.0e-4.  Hardware: 4 Titan X GPUs, Dual Xeon server with SSDs and 512GB RAM.\n\nI played a lot with learning rates and other networks. DenseNet201, DenseNet169 and small own networks.  I tried ensembling too but my ensembling code has a bug and I was able to fix it at the last day  - but to late:( But I got a flu and it was quite hard to work on this project for the last 2 weeks. \n\nI m quite happy to achieve 0.966 accuracy with a single model with only 5270 training images and I m sure I could have acchieved an even higher accuracy with a single model and much higher with ensembling.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 279988,
      "author_name": "kleinsmith",
      "author_url": "",
      "post_date": "02/09/2018 03:14:36",
      "content": "<p>Well played!</p>\n\n<blockquote>\n  <p>On the public LB my best model (resnet 50 with 256 * 256 random crops)\n  achieved 0.960 and on the private LB 0.966</p>\n</blockquote>\n\n<p>Did you use data augmentation? If so, what about TTA?</p>\n\n<p>Did you use any learning rate schedulers?</p>\n\n<p>Did you use random search, or anything to search the hyperparameter space?</p>\n\n<p>(The thought on my mind, \"how did he do it with just a resnet50 and 256x256 random crops?! Did he center crop at all? Isn't the information outside of the 512x512 center outside of the test distribution (ignoring resizing)?\")</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 280266,
      "author_name": "thomastilli",
      "author_url": "",
      "post_date": "02/09/2018 17:07:34",
      "content": "<p>random crops for training, data augmentation,  test : crop in the middle. I used lowering learning rate bei 30% when plataueing for 10 epocs.  TTA with flipping.  Batch size 64, initial Lr 1.0e-4.  Hardware: 4 Titan X GPUs, Dual Xeon server with SSDs and 512GB RAM.</p>\n\n<p>I played a lot with learning rates and other networks. DenseNet201, DenseNet169 and small own networks.  I tried ensembling too but my ensembling code has a bug and I was able to fix it at the last day  - but to late:( But I got a flu and it was quite hard to work on this project for the last 2 weeks. </p>\n\n<p>I m quite happy to achieve 0.966 accuracy with a single model with only 5270 training images and I m sure I could have acchieved an even higher accuracy with a single model and much higher with ensembling.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "279924": "Quite interesting the shake up. The last 2 days my rank on the public LB was dropping from rank 70 to rank 104 and now I jumped up to rank 59. On the public LB my best model (resnet 50 with 256 * 256 random crops) achieved 0.960 and on the private LB 0.966 - which matches quite well my own accuracy metric. Seems that quite a number of people overfitted the public LB.  \n\nThanks to Gleb for his extra dataset - it helped a lot.",
    "279988": "Well played!\n\n&gt; On the public LB my best model (resnet 50 with 256 * 256 random crops)\n&gt; achieved 0.960 and on the private LB 0.966\n\nDid you use data augmentation? If so, what about TTA?\n\nDid you use any learning rate schedulers?\n\nDid you use random search, or anything to search the hyperparameter space?\n\n(The thought on my mind, \"how did he do it with just a resnet50 and 256x256 random crops?! Did he center crop at all? Isn't the information outside of the 512x512 center outside of the test distribution (ignoring resizing)?\")",
    "280266": "random crops for training, data augmentation,  test : crop in the middle. I used lowering learning rate bei 30% when plataueing for 10 epocs.  TTA with flipping.  Batch size 64, initial Lr 1.0e-4.  Hardware: 4 Titan X GPUs, Dual Xeon server with SSDs and 512GB RAM.\n\nI played a lot with learning rates and other networks. DenseNet201, DenseNet169 and small own networks.  I tried ensembling too but my ensembling code has a bug and I was able to fix it at the last day  - but to late:( But I got a flu and it was quite hard to work on this project for the last 2 weeks. \n\nI m quite happy to achieve 0.966 accuracy with a single model with only 5270 training images and I m sure I could have acchieved an even higher accuracy with a single model and much higher with ensembling."
  },
  "source": "meta"
}