![\"log\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2523201%2Fa68774c1d4ac186c62deb87253620843%2FScreenshot%20from%202019-10-25%2021-11-47.png?generation=1572005834655874&alt=media\")
The loss of validation is strange.
\n\nI use\n- U-Net(efficientnet-b3, imagenet(weight))\n- This kernel almost the same
\n\nThe validation loss is unusually high.
\n\n
What are the possible reasons?
\n\nMaybe, because there was a bug in the normalization code.
\n\nNOW\npython\ntrain_transform = [\n albu.HorizontalFlip(p=0.5),\n albu.ShiftScaleRotate(scale_limit=0.5, rotate_limit=0, shift_limit=0.1, p=0.5, border_mode=0),\n albu.GridDistortion(p=0.5),\n albu.OpticalDistortion(p=0.5, distort_limit=2, shift_limit=0.5),\n # albu.Normalize(mean=0.45, std=0.23),\n albu.Resize(320, 640)\n]\n\nHow to normalize?
I'm sorry for the poor English.
", "messages": [ { "id": "657783", "postDate": "10/25/2019 12:21:21", "content": "The loss of validation is strange.
\n\nI use\n- U-Net(efficientnet-b3, imagenet(weight))\n- This kernel almost the same
\n\nThe validation loss is unusually high.
\n\n
What are the possible reasons?
\n\nMaybe, because there was a bug in the normalization code.
\n\nNOW\npython\ntrain_transform = [\n albu.HorizontalFlip(p=0.5),\n albu.ShiftScaleRotate(scale_limit=0.5, rotate_limit=0, shift_limit=0.1, p=0.5, border_mode=0),\n albu.GridDistortion(p=0.5),\n albu.OpticalDistortion(p=0.5, distort_limit=2, shift_limit=0.5),\n # albu.Normalize(mean=0.45, std=0.23),\n albu.Resize(320, 640)\n]\n\nHow to normalize?
I'm sorry for the poor English.
", "rawMarkdown": "The loss of validation is strange.\n\nI use\n- U-Net(efficientnet-b3, imagenet(weight))\n- [This kernel](https://www.kaggle.com/artgor/segmentation-in-pytorch-using-convenient-tools#Model-training) almost the same\n\nThe validation loss is unusually high.\n\n\n\n\n\nWhat are the possible reasons?\n\nMaybe, because there was a bug in the normalization code.\n\nNOW\n```python\ntrain_transform = [\n albu.HorizontalFlip(p=0.5),\n albu.ShiftScaleRotate(scale_limit=0.5, rotate_limit=0, shift_limit=0.1, p=0.5, border_mode=0),\n albu.GridDistortion(p=0.5),\n albu.OpticalDistortion(p=0.5, distort_limit=2, shift_limit=0.5),\n # albu.Normalize(mean=0.45, std=0.23),\n albu.Resize(320, 640)\n]\n```\nHow to normalize?\n\nI'm sorry for the poor English.", "votes": null }, { "id": "658497", "postDate": "10/26/2019 04:21:10", "content": "How many channels of images do you use ?\nI'm not sure, but you should probably remove the parameters or use famous parameters.\nFor example, albu.Normalize(mean=0.45, std=0.23)→albu.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225))
\n\n", "rawMarkdown": "How many channels of images do you use ?\nI'm not sure, but you should probably remove the parameters or use famous parameters.\nFor example, albu.Normalize(mean=0.45, std=0.23)→albu.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225))\n\nhttps://stackoverflow.com/questions/58151507/why-pytorch-officially-use-mean-0-485-0-456-0-406-and-std-0-229-0-224-0-2", "votes": null }, { "id": "658514", "postDate": "10/26/2019 05:04:56", "content": "Thank you.\nAfter posting this discussion, I guessed it that way too, but it didn't work.
\n\nAfter reviewing the code, I found that I did not normalize the validation data.
\n\nThis is just my mistake.
\n\n![\"code\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2523201%2Fcecb435903c22098728a64ccd18fb5f1%2FScreenshot%20from%202019-10-26%2013-59-23.png?generation=1572065983086950&alt=media\")
It has been improved.
\n\n![\"log\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2523201%2F9c50ab7fbda98d94b71de5ae7ed8ada7%2FScreenshot%20from%202019-10-26%2014-02-57.png?generation=1572066194381644&alt=media\")
Glad to hear so.\nTo Japanese young treasure, Good luck !!
", "rawMarkdown": "Glad to hear so.\nTo Japanese young treasure, Good luck !!", "votes": null } ], "comments": [ { "id": 658497, "author_name": "yasagure", "author_url": "", "post_date": "10/26/2019 04:21:10", "content": "How many channels of images do you use ?\nI'm not sure, but you should probably remove the parameters or use famous parameters.\nFor example, albu.Normalize(mean=0.45, std=0.23)→albu.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225))
\n\n", "votes": null, "replies": [ { "id": 658514, "author_name": "chizuchizu", "author_url": "", "post_date": "10/26/2019 05:04:56", "content": "Thank you.\nAfter posting this discussion, I guessed it that way too, but it didn't work.
\n\nAfter reviewing the code, I found that I did not normalize the validation data.
\n\nThis is just my mistake.
\n\n
It has been improved.
\n\n
Glad to hear so.\nTo Japanese young treasure, Good luck !!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "657783": "The loss of validation is strange.\n\nI use\n- U-Net(efficientnet-b3, imagenet(weight))\n- [This kernel](https://www.kaggle.com/artgor/segmentation-in-pytorch-using-convenient-tools#Model-training) almost the same\n\nThe validation loss is unusually high.\n\n\n\n\n\nWhat are the possible reasons?\n\nMaybe, because there was a bug in the normalization code.\n\nNOW\n```python\ntrain_transform = [\n albu.HorizontalFlip(p=0.5),\n albu.ShiftScaleRotate(scale_limit=0.5, rotate_limit=0, shift_limit=0.1, p=0.5, border_mode=0),\n albu.GridDistortion(p=0.5),\n albu.OpticalDistortion(p=0.5, distort_limit=2, shift_limit=0.5),\n # albu.Normalize(mean=0.45, std=0.23),\n albu.Resize(320, 640)\n]\n```\nHow to normalize?\n\nI'm sorry for the poor English.", "658497": "How many channels of images do you use ?\nI'm not sure, but you should probably remove the parameters or use famous parameters.\nFor example, albu.Normalize(mean=0.45, std=0.23)→albu.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225))\n\nhttps://stackoverflow.com/questions/58151507/why-pytorch-officially-use-mean-0-485-0-456-0-406-and-std-0-229-0-224-0-2", "658514": "Thank you.\nAfter posting this discussion, I guessed it that way too, but it didn't work.\n\nAfter reviewing the code, I found that I did not normalize the validation data.\n\nThis is just my mistake.\n\n\n\n\nIt has been improved.\n\n", "658520": "Glad to hear so.\nTo Japanese young treasure, Good luck !!" }, "source": "meta" }