Hi,
\nI wonder if and how compute resources available are related to our results. Would be greate if you share what machines are available for you. The more answers we can get from different ranking groups, the better information we would have on compute resources importance.
\nI can start:
\nGoogle colab pro V100 16gb.
Share please, I think it could be interesting for many of us.
\nCheers.
", "messages": [ { "id": "1229193", "postDate": "03/07/2021 06:42:23", "content": "Hi,
\nI wonder if and how compute resources available are related to our results. Would be greate if you share what machines are available for you. The more answers we can get from different ranking groups, the better information we would have on compute resources importance.
\nI can start:
\nGoogle colab pro V100 16gb.
Share please, I think it could be interesting for many of us.
\nCheers.
", "rawMarkdown": "Hi,\nI wonder if and how compute resources available are related to our results. Would be greate if you share what machines are available for you. The more answers we can get from different ranking groups, the better information we would have on compute resources importance.\nI can start:\nGoogle colab pro V100 16gb.\n\nShare please, I think it could be interesting for many of us.\n\nCheers.", "votes": null }, { "id": "1229482", "postDate": "03/07/2021 11:31:44", "content": "I have a single RTX 2080ti
", "rawMarkdown": "I have a single RTX 2080ti", "votes": null }, { "id": "1229710", "postDate": "03/07/2021 14:59:11", "content": "A single GTX 1080 Ti (11 Gbps GDDR5X memory).
", "rawMarkdown": "A single GTX 1080 Ti (11 Gbps GDDR5X memory).", "votes": null }, { "id": "1229832", "postDate": "03/07/2021 16:05:46", "content": "Thank you for sharing :) Lets wait for more answers!
", "rawMarkdown": "Thank you for sharing :) Lets wait for more answers!", "votes": null }, { "id": "1229835", "postDate": "03/07/2021 16:07:26", "content": "this one is 11 gb, right?
", "rawMarkdown": "this one is 11 gb, right?", "votes": null }, { "id": "1230100", "postDate": "03/07/2021 19:24:30", "content": "a single Quadro RTX 8000 (48GB) . GPU memory is definitely important in this competition.
", "rawMarkdown": "a single Quadro RTX 8000 (48GB) . GPU memory is definitely important in this competition.", "votes": null }, { "id": "1230297", "postDate": "03/08/2021 01:52:35", "content": "May I ask why
", "rawMarkdown": "May I ask why", "votes": null }, { "id": "1230309", "postDate": "03/08/2021 02:21:37", "content": "Why is GPU memory important? I suppose there are more elegant ways of reducing memory usage (freezing batch norm layers and using gradient accumulation), but if you are simply increasing performance via higher image resolutions, then you need a decent amount of memory. But correct me if I am mistaken (I am not nearly as qualified as you are).
", "rawMarkdown": "Why is GPU memory important? I suppose there are more elegant ways of reducing memory usage (freezing batch norm layers and using gradient accumulation), but if you are simply increasing performance via higher image resolutions, then you need a decent amount of memory. But correct me if I am mistaken (I am not nearly as qualified as you are).", "votes": null }, { "id": "1230442", "postDate": "03/08/2021 06:37:18", "content": "Thanks for info. My hypothesis behind this small survey attempt is the idea of how important is larger batch size and higher resolution. Would be great to collect more data points here :)
", "rawMarkdown": "Thanks for info. My hypothesis behind this small survey attempt is the idea of how important is larger batch size and higher resolution. Would be great to collect more data points here :)", "votes": null }, { "id": "1230481", "postDate": "03/08/2021 07:31:43", "content": "\n\nMay I ask why
\n
This is scary that someone in such high place did not think memory is important. I would love to know why as well
", "rawMarkdown": "> May I ask why\n\nThis is scary that someone in such high place did not think memory is important. I would love to know why as well", "votes": null }, { "id": "1230503", "postDate": "03/08/2021 07:43:54", "content": "@nyleve Hi, I think it's important as well. I am trying to get a second opinion on why.
", "rawMarkdown": "nyleve Hi, I think it's important as well. I am trying to get a second opinion on why.", "votes": null }, { "id": "1230715", "postDate": "03/08/2021 11:41:52", "content": "If it helps, I always see an increase in local CV / public leaderboard whenever I bump up the image resolution. I have tested 512 through 736 in increments of 16 (with batch size 32) and the highest I’ve tested was 896 with batch size 24 (it is my best single model).
I'd like to share the following paper. It is a study done on the NIH Chest XRay dataset and the importance of image resolution / batch size. It is not directly relevant, as they use the original NIH labels, and not line catheters, but perhaps it is useful to you.
\n", "rawMarkdown": "If it helps, I always see an increase in local CV / public leaderboard whenever I bump up the image resolution. I have tested `512` through `736` in increments of 16 (with `batch size 32`) and the highest I’ve tested was `896` with `batch size 24` (it is my best single model). \n\nI'd like to share the following paper. It is a study done on the NIH Chest XRay dataset and the importance of image resolution / batch size. It is not directly relevant, as they use the original NIH labels, and not line catheters, but perhaps it is useful to you. \n* [The Effect of Image Resolution on Deep Learning in Radiography](https://pubs.rsna.org/doi/full/10.1148/ryai.2019190015)", "votes": null }, { "id": "1230724", "postDate": "03/08/2021 11:52:17", "content": "The increase in local AUC is nearly uniform across all classes, with the exception of CVC - Borderline, which seems to get a more impressive bump when using larger image resolution. In my experiments, this trend is consistent for both ResNet200D and SEResNet152D.
If you have OOF predictions for the same model trained on different image resolutions, create a correlation heatmap and you will see surprisingly low correlation between these models. I think the model starts to learn more interesting / intricate features when exposed to higher resolution images during training, which is to be expected, but the extent to which this happens is a bit shocking.
", "rawMarkdown": "The increase in local AUC is nearly uniform across all classes, with the exception of `CVC - Borderline`, which seems to get a more impressive bump when using larger image resolution. In my experiments, this trend is consistent for both ResNet200D and SEResNet152D. \n\nIf you have OOF predictions for the same model trained on different image resolutions, create a correlation heatmap and you will see surprisingly low correlation between these models. I think the model starts to learn more interesting / intricate features when exposed to higher resolution images during training, which is to be expected, but the extent to which this happens is a bit shocking.", "votes": null }, { "id": "1230772", "postDate": "03/08/2021 12:53:27", "content": "Thank you very much for such a clear and valuable answers! I will definitely read a paper :)
", "rawMarkdown": "Thank you very much for such a clear and valuable answers! I will definitely read a paper :)", "votes": null }, { "id": "1230800", "postDate": "03/08/2021 13:15:02", "content": "I very much would like to see you win or get top places, so I can learn from you on your approach
", "rawMarkdown": "I very much would like to see you win or get top places, so I can learn from you on your approach", "votes": null }, { "id": "1234461", "postDate": "03/11/2021 09:21:18", "content": "a single Titan RTX (24GB) for training and a single GTX1080Ti (11GB) for inference
", "rawMarkdown": "a single Titan RTX (24GB) for training and a single GTX1080Ti (11GB) for inference", "votes": null }, { "id": "1234479", "postDate": "03/11/2021 09:38:20", "content": "I'm looking forward to see how far the multihead model can go 😃 One of them is inside my blend (thank you for sharing the experiment), but I can't train anymore without GPU quota.
", "rawMarkdown": "I'm looking forward to see how far the multihead model can go 😃 One of them is inside my blend (thank you for sharing the experiment), but I can't train anymore without GPU quota.", "votes": null } ], "comments": [ { "id": 1229482, "author_name": "arc144", "author_url": "", "post_date": "03/07/2021 11:31:44", "content": "I have a single RTX 2080ti
", "votes": null, "replies": [ { "id": 1229835, "author_name": "ademyanchuk", "author_url": "", "post_date": "03/07/2021 16:07:26", "content": "this one is 11 gb, right?
", "votes": null, "replies": [] } ] }, { "id": 1229710, "author_name": "bjoernholzhauer", "author_url": "", "post_date": "03/07/2021 14:59:11", "content": "A single GTX 1080 Ti (11 Gbps GDDR5X memory).
", "votes": null, "replies": [] }, { "id": 1229832, "author_name": "ademyanchuk", "author_url": "", "post_date": "03/07/2021 16:05:46", "content": "Thank you for sharing :) Lets wait for more answers!
", "votes": null, "replies": [] }, { "id": 1230100, "author_name": "tuckerarrants", "author_url": "", "post_date": "03/07/2021 19:24:30", "content": "a single Quadro RTX 8000 (48GB) . GPU memory is definitely important in this competition.
", "votes": null, "replies": [ { "id": 1230297, "author_name": "underwearfitting", "author_url": "", "post_date": "03/08/2021 01:52:35", "content": "May I ask why
", "votes": null, "replies": [] }, { "id": 1230309, "author_name": "tuckerarrants", "author_url": "", "post_date": "03/08/2021 02:21:37", "content": "Why is GPU memory important? I suppose there are more elegant ways of reducing memory usage (freezing batch norm layers and using gradient accumulation), but if you are simply increasing performance via higher image resolutions, then you need a decent amount of memory. But correct me if I am mistaken (I am not nearly as qualified as you are).
", "votes": null, "replies": [] }, { "id": 1230442, "author_name": "ademyanchuk", "author_url": "", "post_date": "03/08/2021 06:37:18", "content": "Thanks for info. My hypothesis behind this small survey attempt is the idea of how important is larger batch size and higher resolution. Would be great to collect more data points here :)
", "votes": null, "replies": [] }, { "id": 1230481, "author_name": "nyleve", "author_url": "", "post_date": "03/08/2021 07:31:43", "content": "\n\nMay I ask why
\n
This is scary that someone in such high place did not think memory is important. I would love to know why as well
", "votes": null, "replies": [] }, { "id": 1230503, "author_name": "underwearfitting", "author_url": "", "post_date": "03/08/2021 07:43:54", "content": "@nyleve Hi, I think it's important as well. I am trying to get a second opinion on why.
", "votes": null, "replies": [] }, { "id": 1230715, "author_name": "tuckerarrants", "author_url": "", "post_date": "03/08/2021 11:41:52", "content": "If it helps, I always see an increase in local CV / public leaderboard whenever I bump up the image resolution. I have tested 512 through 736 in increments of 16 (with batch size 32) and the highest I’ve tested was 896 with batch size 24 (it is my best single model).
I'd like to share the following paper. It is a study done on the NIH Chest XRay dataset and the importance of image resolution / batch size. It is not directly relevant, as they use the original NIH labels, and not line catheters, but perhaps it is useful to you.
\n", "votes": null, "replies": [] }, { "id": 1230724, "author_name": "tuckerarrants", "author_url": "", "post_date": "03/08/2021 11:52:17", "content": "The increase in local AUC is nearly uniform across all classes, with the exception of CVC - Borderline, which seems to get a more impressive bump when using larger image resolution. In my experiments, this trend is consistent for both ResNet200D and SEResNet152D.
If you have OOF predictions for the same model trained on different image resolutions, create a correlation heatmap and you will see surprisingly low correlation between these models. I think the model starts to learn more interesting / intricate features when exposed to higher resolution images during training, which is to be expected, but the extent to which this happens is a bit shocking.
", "votes": null, "replies": [] }, { "id": 1230772, "author_name": "ademyanchuk", "author_url": "", "post_date": "03/08/2021 12:53:27", "content": "Thank you very much for such a clear and valuable answers! I will definitely read a paper :)
", "votes": null, "replies": [] }, { "id": 1230800, "author_name": "reighns", "author_url": "", "post_date": "03/08/2021 13:15:02", "content": "I very much would like to see you win or get top places, so I can learn from you on your approach
", "votes": null, "replies": [] } ] }, { "id": 1234461, "author_name": "ttahara", "author_url": "", "post_date": "03/11/2021 09:21:18", "content": "a single Titan RTX (24GB) for training and a single GTX1080Ti (11GB) for inference
", "votes": null, "replies": [ { "id": 1234479, "author_name": "nyleve", "author_url": "", "post_date": "03/11/2021 09:38:20", "content": "I'm looking forward to see how far the multihead model can go 😃 One of them is inside my blend (thank you for sharing the experiment), but I can't train anymore without GPU quota.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1229193": "Hi,\nI wonder if and how compute resources available are related to our results. Would be greate if you share what machines are available for you. The more answers we can get from different ranking groups, the better information we would have on compute resources importance.\nI can start:\nGoogle colab pro V100 16gb.\n\nShare please, I think it could be interesting for many of us.\n\nCheers.", "1229482": "I have a single RTX 2080ti", "1229710": "A single GTX 1080 Ti (11 Gbps GDDR5X memory).", "1229832": "Thank you for sharing :) Lets wait for more answers!", "1229835": "this one is 11 gb, right?", "1230100": "a single Quadro RTX 8000 (48GB) . GPU memory is definitely important in this competition.", "1230297": "May I ask why", "1230309": "Why is GPU memory important? I suppose there are more elegant ways of reducing memory usage (freezing batch norm layers and using gradient accumulation), but if you are simply increasing performance via higher image resolutions, then you need a decent amount of memory. But correct me if I am mistaken (I am not nearly as qualified as you are).", "1230442": "Thanks for info. My hypothesis behind this small survey attempt is the idea of how important is larger batch size and higher resolution. Would be great to collect more data points here :)", "1230481": "> May I ask why\n\nThis is scary that someone in such high place did not think memory is important. I would love to know why as well", "1230503": "nyleve Hi, I think it's important as well. I am trying to get a second opinion on why.", "1230715": "If it helps, I always see an increase in local CV / public leaderboard whenever I bump up the image resolution. I have tested `512` through `736` in increments of 16 (with `batch size 32`) and the highest I’ve tested was `896` with `batch size 24` (it is my best single model). \n\nI'd like to share the following paper. It is a study done on the NIH Chest XRay dataset and the importance of image resolution / batch size. It is not directly relevant, as they use the original NIH labels, and not line catheters, but perhaps it is useful to you. \n* [The Effect of Image Resolution on Deep Learning in Radiography](https://pubs.rsna.org/doi/full/10.1148/ryai.2019190015)", "1230724": "The increase in local AUC is nearly uniform across all classes, with the exception of `CVC - Borderline`, which seems to get a more impressive bump when using larger image resolution. In my experiments, this trend is consistent for both ResNet200D and SEResNet152D. \n\nIf you have OOF predictions for the same model trained on different image resolutions, create a correlation heatmap and you will see surprisingly low correlation between these models. I think the model starts to learn more interesting / intricate features when exposed to higher resolution images during training, which is to be expected, but the extent to which this happens is a bit shocking.", "1230772": "Thank you very much for such a clear and valuable answers! I will definitely read a paper :)", "1230800": "I very much would like to see you win or get top places, so I can learn from you on your approach", "1234461": "a single Titan RTX (24GB) for training and a single GTX1080Ti (11GB) for inference", "1234479": "I'm looking forward to see how far the multihead model can go 😃 One of them is inside my blend (thank you for sharing the experiment), but I can't train anymore without GPU quota." }, "source": "meta" }