Thank you to the organizers for hosting this exciting competition and congratulations to all the winners! This is my first time participating in a competition like this, and below is my solution and approach.
\nwe use 3D Resnet(https://github.com/kenshohara/3D-ResNets-PyTorch/blob/master/models/resnet.py) for our encoder and CNN decoder.
\nWe tried different decoders, such as Unet, Uperhead, CNN, and ResCNN. In the end, the CNN decoder, while being the simplest, achieved the best results.
\n| backbone | \ndecoder | \nZ-DIM | \n
|---|---|---|
| 3DResnet-18 | \nCNN | \n22 | \n
| 3DResnet-34 | \nCNN | \n22 | \n
Z_DIM:21->43
\nimg_size:224
\nstride_size:56
A.Rotate(limit=90, p=0.5),
\nA.HorizontalFlip(p=0.5),
\nA.VerticalFlip(p=0.5),
\nA.RandomBrightnessContrast(p=0.75),
\nA.ShiftScaleRotate(p=0.75),
\nA.OneOf([
\n A.GaussNoise(),
\n A.GaussianBlur(),
\n A.MotionBlur(),
\n ], p=0.5),
\nA.GridDistortion(num_steps=5, distort_limit=0.3, p=0.5),
\nA.CoarseDropout(max_holes=1, max_width=int(size * 0.3), max_height=int(size * 0.3),
\n mask_fill_value=0, p=0.5),
\nA.Normalize(
\n mean= [0] * in_chans,
\n std= [1] * in_chans
\n),
\nToTensorV2(transpose_mask=True)
We used two different training methods. The first method involved selecting one folder from 1, 2, and 3 as the validation set and training three models. However, due to poor performance on validation set 2, we only used models trained with validation sets 1 and 3 for our 3DRsnet-18 model. The second method involved randomly selecting image blocks from 1, 2, and 3 as the validation set. We used this method to train 4 models: two based on 3DResnet-18 with a stride of 56 and 37, and another two based on 3DResnet-34 with a stride of 56 and 37.
\nBCELoss
\nAdamw
\nGradualWarmupSchedulerV2(https://www.kaggle.com/code/underwearfitting/single-fold-training-of-resnet200d-lb0-965)
img_size:224
\nstride_size:224
\nth:0.5
\nTTA: 4 rotates, h/v flips
\nlast model:
\n(3DResnet-18 + CNN) * 4 + (3DResnet-34 ＋CNN) * 2
\npublic:0.786897
\nprivate:0.652892
training code:https://github.com/tkz24589/Deep_Learning_work_space/tree/main/convolution/segmentation/vesuvius-challenge-ink-detection-tutorial
\nnotebook example: https://www.kaggle.com/code/kongzhangtang/resnet3d-cnn
Thank you to the organizers for hosting this exciting competition and congratulations to all the winners! This is my first time participating in a competition like this, and below is my solution and approach.
\nwe use 3D Resnet(https://github.com/kenshohara/3D-ResNets-PyTorch/blob/master/models/resnet.py) for our encoder and CNN decoder.
\nWe tried different decoders, such as Unet, Uperhead, CNN, and ResCNN. In the end, the CNN decoder, while being the simplest, achieved the best results.
\n| backbone | \ndecoder | \nZ-DIM | \n
|---|---|---|
| 3DResnet-18 | \nCNN | \n22 | \n
| 3DResnet-34 | \nCNN | \n22 | \n
Z_DIM:21->43
\nimg_size:224
\nstride_size:56
A.Rotate(limit=90, p=0.5),
\nA.HorizontalFlip(p=0.5),
\nA.VerticalFlip(p=0.5),
\nA.RandomBrightnessContrast(p=0.75),
\nA.ShiftScaleRotate(p=0.75),
\nA.OneOf([
\n A.GaussNoise(),
\n A.GaussianBlur(),
\n A.MotionBlur(),
\n ], p=0.5),
\nA.GridDistortion(num_steps=5, distort_limit=0.3, p=0.5),
\nA.CoarseDropout(max_holes=1, max_width=int(size * 0.3), max_height=int(size * 0.3),
\n mask_fill_value=0, p=0.5),
\nA.Normalize(
\n mean= [0] * in_chans,
\n std= [1] * in_chans
\n),
\nToTensorV2(transpose_mask=True)
We used two different training methods. The first method involved selecting one folder from 1, 2, and 3 as the validation set and training three models. However, due to poor performance on validation set 2, we only used models trained with validation sets 1 and 3 for our 3DRsnet-18 model. The second method involved randomly selecting image blocks from 1, 2, and 3 as the validation set. We used this method to train 4 models: two based on 3DResnet-18 with a stride of 56 and 37, and another two based on 3DResnet-34 with a stride of 56 and 37.
\nBCELoss
\nAdamw
\nGradualWarmupSchedulerV2(https://www.kaggle.com/code/underwearfitting/single-fold-training-of-resnet200d-lb0-965)
img_size:224
\nstride_size:224
\nth:0.5
\nTTA: 4 rotates, h/v flips
\nlast model:
\n(3DResnet-18 + CNN) * 4 + (3DResnet-34 ＋CNN) * 2
\npublic:0.786897
\nprivate:0.652892
training code:https://github.com/tkz24589/Deep_Learning_work_space/tree/main/convolution/segmentation/vesuvius-challenge-ink-detection-tutorial
\nnotebook example: https://www.kaggle.com/code/kongzhangtang/resnet3d-cnn
Congratulations on your gold medal on your first competition! It looks like you used a very sound approach to trying a number of options and figuring out what worked best. Thank you for sharing your summary and your code.
", "rawMarkdown": "Congratulations on your gold medal on your first competition! It looks like you used a very sound approach to trying a number of options and figuring out what worked best. Thank you for sharing your summary and your code.", "votes": null }, { "id": "2319847", "postDate": "06/27/2023 10:36:05", "content": "Thank you, and congratulations on winning first place!
", "rawMarkdown": "Thank you, and congratulations on winning first place!", "votes": null } ], "comments": [ { "id": 2319678, "author_name": "socated", "author_url": "", "post_date": "06/27/2023 08:50:39", "content": "Congratulations on your gold medal on your first competition! It looks like you used a very sound approach to trying a number of options and figuring out what worked best. Thank you for sharing your summary and your code.
", "votes": null, "replies": [ { "id": 2319847, "author_name": "kongzhangtang", "author_url": "", "post_date": "06/27/2023 10:36:05", "content": "Thank you, and congratulations on winning first place!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2314145": "# **12th place solution**\nThank you to the organizers for hosting this exciting competition and congratulations to all the winners! This is my first time participating in a competition like this, and below is my solution and approach.\n## **overview**\nwe use 3D Resnet(https://github.com/kenshohara/3D-ResNets-PyTorch/blob/master/models/resnet.py) for our encoder and CNN decoder.\n\nWe tried different decoders, such as Unet, Uperhead, CNN, and ResCNN. In the end, the CNN decoder, while being the simplest, achieved the best results.\n##**model architecture**\n| backbone |decoder| Z-DIM |\n| --- | --- | --- |\n| 3DResnet-18 |CNN| 22 |\n| 3DResnet-34 |CNN| 22 |\n\n## **training**\n\nZ_DIM:21->43\nimg_size:224\nstride_size:56\n\n###augmentation:\nA.Rotate(limit=90, p=0.5),\nA.HorizontalFlip(p=0.5),\nA.VerticalFlip(p=0.5),\nA.RandomBrightnessContrast(p=0.75),\nA.ShiftScaleRotate(p=0.75),\nA.OneOf([\n A.GaussNoise(),\n A.GaussianBlur(),\n A.MotionBlur(),\n ], p=0.5),\nA.GridDistortion(num_steps=5, distort_limit=0.3, p=0.5),\nA.CoarseDropout(max_holes=1, max_width=int(size * 0.3), max_height=int(size * 0.3), \n mask_fill_value=0, p=0.5),\nA.Normalize(\n mean= [0] * in_chans,\n std= [1] * in_chans\n),\nToTensorV2(transpose_mask=True)\n\n###training details\nWe used two different training methods. The first method involved selecting one folder from 1, 2, and 3 as the validation set and training three models. However, due to poor performance on validation set 2, we only used models trained with validation sets 1 and 3 for our 3DRsnet-18 model. The second method involved randomly selecting image blocks from 1, 2, and 3 as the validation set. We used this method to train 4 models: two based on 3DResnet-18 with a stride of 56 and 37, and another two based on 3DResnet-34 with a stride of 56 and 37.\n\n##others\nBCELoss\nAdamw\nGradualWarmupSchedulerV2(https://www.kaggle.com/code/underwearfitting/single-fold-training-of-resnet200d-lb0-965)\n\n## **inference**\nimg_size:224\nstride_size:224\nth:0.5\nTTA: 4 rotates, h/v flips\nlast model:\n(3DResnet-18 + CNN) * 4 + (3DResnet-34 ＋CNN) * 2\npublic:0.786897\nprivate:0.652892\n\ntraining code:https://github.com/tkz24589/Deep_Learning_work_space/tree/main/convolution/segmentation/vesuvius-challenge-ink-detection-tutorial\nnotebook example: https://www.kaggle.com/code/kongzhangtang/resnet3d-cnn", "2319678": "Congratulations on your gold medal on your first competition! It looks like you used a very sound approach to trying a number of options and figuring out what worked best. Thank you for sharing your summary and your code.", "2319847": "Thank you, and congratulations on winning first place!" }, "source": "meta" }