{
  "id": 186369,
  "title": "TPU Baseline: 1% Pubic LB now",
  "url": "/competitions/lyft-motion-prediction-autonomous-vehicles/discussion/186369",
  "author_name": "",
  "post_date": "2020-09-24T08:46:53.382426600Z",
  "votes": 11,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I public the baseline resnet50 TPU that gets 23.xx LB now. <a href=\"https://www.kaggle.com/doanquanvietnamca/tpu-resnet50-faster-better\" target=\"_blank\">https://www.kaggle.com/doanquanvietnamca/tpu-resnet50-faster-better</a></p>\n<p>I use the optimizer here: <a href=\"https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/186256\" target=\"_blank\">https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/186256</a><br>\nwith 10M sample. TPU helps the training process faster and better. Hope it helps someone!!<br>\nwith the right training strategy, you can achieve high LB.</p>",
  "messages": [
    {
      "id": "1024995",
      "postDate": "09/24/2020 08:46:53",
      "content": "<p>I public the baseline resnet50 TPU that gets 23.xx LB now. <a href=\"https://www.kaggle.com/doanquanvietnamca/tpu-resnet50-faster-better\" target=\"_blank\">https://www.kaggle.com/doanquanvietnamca/tpu-resnet50-faster-better</a></p>\n<p>I use the optimizer here: <a href=\"https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/186256\" target=\"_blank\">https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/186256</a><br>\nwith 10M sample. TPU helps the training process faster and better. Hope it helps someone!!<br>\nwith the right training strategy, you can achieve high LB.</p>",
      "rawMarkdown": "I public the baseline resnet50 TPU that gets 23.xx LB now. https://www.kaggle.com/doanquanvietnamca/tpu-resnet50-faster-better\n\nI use the optimizer here: https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/186256\nwith 10M sample. TPU helps the training process faster and better. Hope it helps someone!!\nwith the right training strategy, you can achieve high LB.",
      "votes": null
    },
    {
      "id": "1025085",
      "postDate": "09/24/2020 09:52:04",
      "content": "<p>Can u write something about training information? (only if its okay with you). how much time it took for 10M samples, image size, train/train_full. Is it possible to reach 10M within kaggle kernal limit? thanks.</p>",
      "rawMarkdown": "Can u write something about training information? (only if its okay with you). how much time it took for 10M samples, image size, train/train_full. Is it possible to reach 10M within kaggle kernal limit? thanks.",
      "votes": null
    },
    {
      "id": "1025103",
      "postDate": "09/24/2020 10:12:44",
      "content": "<p>history: 15 I try 30 but it's CUDA Error<br>\ntrain_size and config like in the notebook<br>\nI train much time, because kernel limits, train and retrain. <br>\nMaybe 3-4 times train. So can say that is 3-4epochs</p>",
      "rawMarkdown": "history: 15 I try 30 but it's CUDA Error\ntrain_size and config like in the notebook\nI train much time, because kernel limits, train and retrain. \nMaybe 3-4 times train. So can say that is 3-4epochs",
      "votes": null
    },
    {
      "id": "1032682",
      "postDate": "09/30/2020 11:16:05",
      "content": "<p>How many iterations/sec do you achieve with using TPU?</p>",
      "rawMarkdown": "How many iterations/sec do you achieve with using TPU?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1025085,
      "author_name": "deepakrajpurushothaman",
      "author_url": "",
      "post_date": "09/24/2020 09:52:04",
      "content": "<p>Can u write something about training information? (only if its okay with you). how much time it took for 10M samples, image size, train/train_full. Is it possible to reach 10M within kaggle kernal limit? thanks.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 1025103,
      "author_name": "doanquanvietnamca",
      "author_url": "",
      "post_date": "09/24/2020 10:12:44",
      "content": "<p>history: 15 I try 30 but it's CUDA Error<br>\ntrain_size and config like in the notebook<br>\nI train much time, because kernel limits, train and retrain. <br>\nMaybe 3-4 times train. So can say that is 3-4epochs</p>",
      "votes": null,
      "replies": [
        {
          "id": 1032682,
          "author_name": "aliabdin1",
          "author_url": "",
          "post_date": "09/30/2020 11:16:05",
          "content": "<p>How many iterations/sec do you achieve with using TPU?</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "1024995": "I public the baseline resnet50 TPU that gets 23.xx LB now. https://www.kaggle.com/doanquanvietnamca/tpu-resnet50-faster-better\n\nI use the optimizer here: https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/186256\nwith 10M sample. TPU helps the training process faster and better. Hope it helps someone!!\nwith the right training strategy, you can achieve high LB.",
    "1025085": "Can u write something about training information? (only if its okay with you). how much time it took for 10M samples, image size, train/train_full. Is it possible to reach 10M within kaggle kernal limit? thanks.",
    "1025103": "history: 15 I try 30 but it's CUDA Error\ntrain_size and config like in the notebook\nI train much time, because kernel limits, train and retrain. \nMaybe 3-4 times train. So can say that is 3-4epochs",
    "1032682": "How many iterations/sec do you achieve with using TPU?"
  },
  "source": "meta"
}