{
  "id": 337570,
  "title": "Previous competition specifications ",
  "url": "/competitions/hubmap-organ-segmentation/discussion/337570",
  "author_name": "",
  "post_date": "2022-07-16T17:03:28.950254600Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I complied some import parameters from the 2021 competition <a href=\"https://github.com/cns-iu/ccf-research-kaggle-2021/tree/main/models\" target=\"_blank\">source</a> </p>\n<table>\n<thead>\n<tr>\n<th>competitor</th>\n<th>architecture</th>\n<th>encoder</th>\n<th>batch_size</th>\n<th>image_size</th>\n<th>epoch</th>\n<th>loss</th>\n<th>optim</th>\n<th>max_lr</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Tom</td>\n<td>unet</td>\n<td>seresnext101</td>\n<td>8</td>\n<td>1024</td>\n<td>20</td>\n<td>Binary cross-entropy loss and Lovász Hinge loss</td>\n<td>SGD</td>\n<td>1e-4 - 1e-6</td>\n</tr>\n<tr>\n<td>Gleb</td>\n<td>unet</td>\n<td>regnety16 with scse attention decoder</td>\n<td>8</td>\n<td>1024</td>\n<td>80</td>\n<td>Dice coefficient loss</td>\n<td>AdamW</td>\n<td>1e-4 - 1e-6</td>\n</tr>\n<tr>\n<td>Whats goin on</td>\n<td>unet</td>\n<td>resnet101_32x4d64</td>\n<td>16</td>\n<td>1024</td>\n<td>50</td>\n<td>Binary cross-entropy</td>\n<td>Adam</td>\n<td>1e-4</td>\n</tr>\n<tr>\n<td>Deeplive.exe</td>\n<td>unet</td>\n<td>efficientnet-b1</td>\n<td>32</td>\n<td>512</td>\n<td>-</td>\n<td>Binary cross-entropy</td>\n<td>Adam</td>\n<td>1e-3</td>\n</tr>\n<tr>\n<td>Deepflas2</td>\n<td>unet</td>\n<td>efficientnet-b2</td>\n<td>16</td>\n<td>512</td>\n<td>-</td>\n<td>Dice-Crossentropy loss</td>\n<td>Ranger</td>\n<td>le-3</td>\n</tr>\n</tbody>\n</table>\n<p>The one similarity for the top 5 is Unet.</p>",
  "messages": [
    {
      "id": "1858109",
      "postDate": "07/16/2022 17:03:28",
      "content": "<p>I complied some import parameters from the 2021 competition <a href=\"https://github.com/cns-iu/ccf-research-kaggle-2021/tree/main/models\" target=\"_blank\">source</a> </p>\n<table>\n<thead>\n<tr>\n<th>competitor</th>\n<th>architecture</th>\n<th>encoder</th>\n<th>batch_size</th>\n<th>image_size</th>\n<th>epoch</th>\n<th>loss</th>\n<th>optim</th>\n<th>max_lr</th>\n</tr>\n</thead>\n<tbody>\n<tr>\n<td>Tom</td>\n<td>unet</td>\n<td>seresnext101</td>\n<td>8</td>\n<td>1024</td>\n<td>20</td>\n<td>Binary cross-entropy loss and Lovász Hinge loss</td>\n<td>SGD</td>\n<td>1e-4 - 1e-6</td>\n</tr>\n<tr>\n<td>Gleb</td>\n<td>unet</td>\n<td>regnety16 with scse attention decoder</td>\n<td>8</td>\n<td>1024</td>\n<td>80</td>\n<td>Dice coefficient loss</td>\n<td>AdamW</td>\n<td>1e-4 - 1e-6</td>\n</tr>\n<tr>\n<td>Whats goin on</td>\n<td>unet</td>\n<td>resnet101_32x4d64</td>\n<td>16</td>\n<td>1024</td>\n<td>50</td>\n<td>Binary cross-entropy</td>\n<td>Adam</td>\n<td>1e-4</td>\n</tr>\n<tr>\n<td>Deeplive.exe</td>\n<td>unet</td>\n<td>efficientnet-b1</td>\n<td>32</td>\n<td>512</td>\n<td>-</td>\n<td>Binary cross-entropy</td>\n<td>Adam</td>\n<td>1e-3</td>\n</tr>\n<tr>\n<td>Deepflas2</td>\n<td>unet</td>\n<td>efficientnet-b2</td>\n<td>16</td>\n<td>512</td>\n<td>-</td>\n<td>Dice-Crossentropy loss</td>\n<td>Ranger</td>\n<td>le-3</td>\n</tr>\n</tbody>\n</table>\n<p>The one similarity for the top 5 is Unet.</p>",
      "rawMarkdown": "I complied some import parameters from the 2021 competition [source](https://github.com/cns-iu/ccf-research-kaggle-2021/tree/main/models) \n|competitor|architecture|encoder|batch_size|image_size|epoch|loss|optim|max_lr|\n| --- | --- |--- | --- |--- | --- |--- | --- |--- |\n|Tom|unet|seresnext101|8|1024|20| Binary cross-entropy loss and Lovász Hinge loss|SGD|1e-4 - 1e-6|\n|Gleb|unet|regnety16 with scse attention decoder|8|1024|80|Dice coefficient loss|AdamW|1e-4 - 1e-6|\n|Whats goin on|unet|resnet101_32x4d64|16|1024|50|Binary cross-entropy|Adam|1e-4|\n|Deeplive.exe|unet|efficientnet-b1|32|512|-|Binary cross-entropy|Adam|1e-3|\n|Deepflas2|unet|efficientnet-b2|16|512|-|Dice-Crossentropy loss|Ranger|le-3|\n\nThe one similarity for the top 5 is Unet.",
      "votes": null
    },
    {
      "id": "1858276",
      "postDate": "07/16/2022 19:24:51",
      "content": "<p>In the last hubmap challenge there was little difference in the architectures of the different solutions we (deepflash) tried a couple of different encoders and decoders, but never saw consistent improvements.<br>\nMost Important adjustments have been the way the data was used for training and which additional annotations have been used in the training process. </p>",
      "rawMarkdown": "In the last hubmap challenge there was little difference in the architectures of the different solutions we (deepflash) tried a couple of different encoders and decoders, but never saw consistent improvements.\nMost Important adjustments have been the way the data was used for training and which additional annotations have been used in the training process.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1858276,
      "author_name": "theudas",
      "author_url": "",
      "post_date": "07/16/2022 19:24:51",
      "content": "<p>In the last hubmap challenge there was little difference in the architectures of the different solutions we (deepflash) tried a couple of different encoders and decoders, but never saw consistent improvements.<br>\nMost Important adjustments have been the way the data was used for training and which additional annotations have been used in the training process. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1858109": "I complied some import parameters from the 2021 competition [source](https://github.com/cns-iu/ccf-research-kaggle-2021/tree/main/models) \n|competitor|architecture|encoder|batch_size|image_size|epoch|loss|optim|max_lr|\n| --- | --- |--- | --- |--- | --- |--- | --- |--- |\n|Tom|unet|seresnext101|8|1024|20| Binary cross-entropy loss and Lovász Hinge loss|SGD|1e-4 - 1e-6|\n|Gleb|unet|regnety16 with scse attention decoder|8|1024|80|Dice coefficient loss|AdamW|1e-4 - 1e-6|\n|Whats goin on|unet|resnet101_32x4d64|16|1024|50|Binary cross-entropy|Adam|1e-4|\n|Deeplive.exe|unet|efficientnet-b1|32|512|-|Binary cross-entropy|Adam|1e-3|\n|Deepflas2|unet|efficientnet-b2|16|512|-|Dice-Crossentropy loss|Ranger|le-3|\n\nThe one similarity for the top 5 is Unet.",
    "1858276": "In the last hubmap challenge there was little difference in the architectures of the different solutions we (deepflash) tried a couple of different encoders and decoders, but never saw consistent improvements.\nMost Important adjustments have been the way the data was used for training and which additional annotations have been used in the training process."
  },
  "source": "meta"
}