{
  "id": 37377,
  "title": "How long does BN take to converge?",
  "url": "/competitions/carvana-image-masking-challenge/discussion/37377",
  "author_name": "",
  "post_date": "2017-08-01T10:28:12.664418300Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I am currently training segNet. The training loss is around 0.008 after about 15 epochs. But when validating, the loss is not stable like below :</p>\n\n<p>echpo :13, avg_loss: 0.759422 avg_iou: 0.457123</p>\n\n<p>echpo :14, avg_loss: 0.045098 avg_iou: 0.931947</p>\n\n<p>echpo :15, avg_loss: 0.011147 avg_iou: 0.980389</p>\n\n<p>echpo :16, avg_loss: 0.077130 avg_iou: 0.874013</p>\n\n<p>echpo :17, avg_loss: 0.117785 avg_iou: 0.823593</p>\n\n<p>echpo :18, avg_loss: 0.025768 avg_iou: 0.955402</p>\n\n<p>echpo :19, avg_loss: 0.056577 avg_iou: 0.906639</p>\n\n<p>With the batch size of 16 and input image size of( 480, 320) and one epoch of around 10000 images (after augment) , how long does BN generally need to get converged?</p>\n\n<p>Looking for help~ </p>",
  "messages": [
    {
      "id": "209132",
      "postDate": "08/01/2017 10:28:12",
      "content": "<p>I am currently training segNet. The training loss is around 0.008 after about 15 epochs. But when validating, the loss is not stable like below :</p>\n\n<p>echpo :13, avg_loss: 0.759422 avg_iou: 0.457123</p>\n\n<p>echpo :14, avg_loss: 0.045098 avg_iou: 0.931947</p>\n\n<p>echpo :15, avg_loss: 0.011147 avg_iou: 0.980389</p>\n\n<p>echpo :16, avg_loss: 0.077130 avg_iou: 0.874013</p>\n\n<p>echpo :17, avg_loss: 0.117785 avg_iou: 0.823593</p>\n\n<p>echpo :18, avg_loss: 0.025768 avg_iou: 0.955402</p>\n\n<p>echpo :19, avg_loss: 0.056577 avg_iou: 0.906639</p>\n\n<p>With the batch size of 16 and input image size of( 480, 320) and one epoch of around 10000 images (after augment) , how long does BN generally need to get converged?</p>\n\n<p>Looking for help~ </p>",
      "rawMarkdown": "I am currently training segNet. The training loss is around 0.008 after about 15 epochs. But when validating, the loss is not stable like below :\n\nechpo :13, avg_loss: 0.759422 avg_iou: 0.457123\n\nechpo :14, avg_loss: 0.045098 avg_iou: 0.931947\n\nechpo :15, avg_loss: 0.011147 avg_iou: 0.980389\n\nechpo :16, avg_loss: 0.077130 avg_iou: 0.874013\n\nechpo :17, avg_loss: 0.117785 avg_iou: 0.823593\n\nechpo :18, avg_loss: 0.025768 avg_iou: 0.955402\n\nechpo :19, avg_loss: 0.056577 avg_iou: 0.906639\n\nWith the batch size of 16 and input image size of( 480, 320) and one epoch of around 10000 images (after augment) , how long does BN generally need to get converged?\n\nLooking for help~",
      "votes": null
    },
    {
      "id": "209411",
      "postDate": "08/02/2017 06:29:29",
      "content": "<p>i use BN for my unet. Refer to my post for the train curve for the loss. The train log files can be downloaded from the google drive in my post too.</p>\n\n<p><a href=\"https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/37208\">https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/37208</a></p>",
      "rawMarkdown": "i use BN for my unet. Refer to my post for the train curve for the loss. The train log files can be downloaded from the google drive in my post too.\n \nhttps://www.kaggle.com/c/carvana-image-masking-challenge/discussion/37208",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 209411,
      "author_name": "hengck23",
      "author_url": "",
      "post_date": "08/02/2017 06:29:29",
      "content": "<p>i use BN for my unet. Refer to my post for the train curve for the loss. The train log files can be downloaded from the google drive in my post too.</p>\n\n<p><a href=\"https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/37208\">https://www.kaggle.com/c/carvana-image-masking-challenge/discussion/37208</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "209132": "I am currently training segNet. The training loss is around 0.008 after about 15 epochs. But when validating, the loss is not stable like below :\n\nechpo :13, avg_loss: 0.759422 avg_iou: 0.457123\n\nechpo :14, avg_loss: 0.045098 avg_iou: 0.931947\n\nechpo :15, avg_loss: 0.011147 avg_iou: 0.980389\n\nechpo :16, avg_loss: 0.077130 avg_iou: 0.874013\n\nechpo :17, avg_loss: 0.117785 avg_iou: 0.823593\n\nechpo :18, avg_loss: 0.025768 avg_iou: 0.955402\n\nechpo :19, avg_loss: 0.056577 avg_iou: 0.906639\n\nWith the batch size of 16 and input image size of( 480, 320) and one epoch of around 10000 images (after augment) , how long does BN generally need to get converged?\n\nLooking for help~",
    "209411": "i use BN for my unet. Refer to my post for the train curve for the loss. The train log files can be downloaded from the google drive in my post too.\n \nhttps://www.kaggle.com/c/carvana-image-masking-challenge/discussion/37208"
  },
  "source": "meta"
}