{ "id": 436943, "title": "My 3D model learns nothing", "url": "/competitions/rsna-2023-abdominal-trauma-detection/discussion/436943", "author_name": "JanGlinko2", "post_date": "2023-09-04T17:53:03.919000", "votes": 0, "comment_count": 7, "views": 0, "content": "

Hello,
\nI'm creating 3D stacks with spacing 5mm, then resizing it to [1, 80, 256, 256] with scipy.ndimage.zoom. Then I use 3D ResNet50 model from pytorchvideo. I wanted to start from predicting any_injuries, because class imbalance is not as high as in other classes. The problem is my model learns nothing. Training loss is fluctuating.

\n

What I've tried:

\n\n

Using all classes, model is learning class imbalance.

\n

Can you provide me with any tips how to continue? I'm stuck.

", "messages": [ { "id": 2425581, "postDate": "2023-09-06T03:48:22.520Z", "content": "

this should help:
\nhttps://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/435053#2425570

", "rawMarkdown": "this should help:\nhttps://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/435053#2425570", "votes": 1, "replies": [ { "id": 2426276, "postDate": "2023-09-06T14:21:53.307Z", "content": "

Have you changed (increased) number of input/output channels for the whole ResNet50d or have you only modified number of input channels for the first layer?

", "rawMarkdown": "Have you changed (increased) number of input/output channels for the whole ResNet50d or have you only modified number of input channels for the first layer?", "replies": [ { "id": 2426285, "postDate": "2023-09-06T14:28:35.513Z", "content": "
from timm.models.resnet  *\n\n\n        .encoder_in_dim =  \n\n        .encoder = resnet50d(pretrained=, in_chans=.encoder_in_dim)  \n        .decoder = nn.Sequential(\n            nn.Conv3d(, , kernel_size=(,,),stride=(,,), padding=(,,)),\n            nn.BatchNorm3d(),\n\n
", "rawMarkdown": "```\nfrom timm.models.resnet import *\n\n...\n\t\tself.encoder_in_dim = 4 \n \n\t\tself.encoder = resnet50d(pretrained=True, in_chans=self.encoder_in_dim) \n\t\tself.decoder = nn.Sequential(\n\t\t\tnn.Conv3d(2048, 256, kernel_size=(3,3,3),stride=(2,1,1), padding=(1,1,1)),\n\t\t\tnn.BatchNorm3d(256),\n...\n\n```" }, { "id": 2426441, "postDate": "2023-09-06T15:54:10.857Z", "content": "

So you use 4 slices to extract features and probably you perform multiple runs to stack features from many slices?

", "rawMarkdown": "So you use 4 slices to extract features and probably you perform multiple runs to stack features from many slices?" } ] } ] }, { "id": 2429424, "postDate": "2023-09-08T15:35:58.483Z", "content": "

Update: My 3D model still learns nothing, I can't find way data preprocessing to have a signal in 3D. Moreover, its awfully slow.
\n2.5D works better and 2D much much better.

", "rawMarkdown": "Update: My 3D model still learns nothing, I can't find way data preprocessing to have a signal in 3D. Moreover, its awfully slow.\n2.5D works better and 2D much much better." }, { "id": 2423730, "postDate": "2023-09-04T18:29:15.597Z", "content": "

You should work on your data pipeline. Model isn't that important at this stage. I'm still experimenting with EfficientNet b0.

", "rawMarkdown": "You should work on your data pipeline. Model isn't that important at this stage. I'm still experimenting with EfficientNet b0.", "replies": [ { "id": 2424068, "postDate": "2023-09-05T03:17:36.640Z", "content": "

Thank you for the reminder. I just discovered this. The prediction score of the VIT model I used dummy is actually higher than the LB score I trained on the patient-level data. Data quality determines the upper limit of the model.

", "rawMarkdown": "Thank you for the reminder. I just discovered this. The prediction score of the VIT model I used dummy is actually higher than the LB score I trained on the patient-level data. Data quality determines the upper limit of the model.\n\n", "isDeleted": true } ] }, { "id": 2423666, "postDate": "2023-09-04T17:53:03.920Z", "content": "

Hello,
\nI'm creating 3D stacks with spacing 5mm, then resizing it to [1, 80, 256, 256] with scipy.ndimage.zoom. Then I use 3D ResNet50 model from pytorchvideo. I wanted to start from predicting any_injuries, because class imbalance is not as high as in other classes. The problem is my model learns nothing. Training loss is fluctuating.

\n

What I've tried:

\n\n

Using all classes, model is learning class imbalance.

\n

Can you provide me with any tips how to continue? I'm stuck.

", "rawMarkdown": "Hello, \nI'm creating 3D stacks with spacing 5mm, then resizing it to [1, 80, 256, 256] with `scipy.ndimage.zoom`. Then I use 3D ResNet50 model from `pytorchvideo`. I wanted to start from predicting `any_injuries`, because class imbalance is not as high as in other classes. The problem is my model learns nothing. Training loss is fluctuating. \n\nWhat I've tried:\n* class weighting\n* standarization (with mean and std)\n* different learning rates\n* different optimizers\n\nUsing all classes, model is learning class imbalance.\n\nCan you provide me with any tips how to continue? I'm stuck." } ], "comments": [ { "id": 2425581, "author_name": "hengck23", "author_url": "", "post_date": "2023-09-06T03:48:22.520000", "content": "

this should help:
\nhttps://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/435053#2425570

", "votes": 1, "replies": [ { "id": 2426276, "author_name": "JanGlinko2", "author_url": "", "post_date": "2023-09-06T14:21:53.307000", "content": "

Have you changed (increased) number of input/output channels for the whole ResNet50d or have you only modified number of input channels for the first layer?

", "votes": 0, "replies": [ { "id": 2426285, "author_name": "hengck23", "author_url": "", "post_date": "2023-09-06T14:28:35.513000", "content": "
from timm.models.resnet  *\n\n\n        .encoder_in_dim =  \n\n        .encoder = resnet50d(pretrained=, in_chans=.encoder_in_dim)  \n        .decoder = nn.Sequential(\n            nn.Conv3d(, , kernel_size=(,,),stride=(,,), padding=(,,)),\n            nn.BatchNorm3d(),\n\n
", "votes": 0, "replies": [] }, { "id": 2426441, "author_name": "JanGlinko2", "author_url": "", "post_date": "2023-09-06T15:54:10.857000", "content": "

So you use 4 slices to extract features and probably you perform multiple runs to stack features from many slices?

", "votes": 0, "replies": [] } ] } ] }, { "id": 2429424, "author_name": "JanGlinko2", "author_url": "", "post_date": "2023-09-08T15:35:58.483000", "content": "

Update: My 3D model still learns nothing, I can't find way data preprocessing to have a signal in 3D. Moreover, its awfully slow.
\n2.5D works better and 2D much much better.

", "votes": 0, "replies": [] }, { "id": 2423730, "author_name": "Gunes Evitan", "author_url": "", "post_date": "2023-09-04T18:29:15.597000", "content": "

You should work on your data pipeline. Model isn't that important at this stage. I'm still experimenting with EfficientNet b0.

", "votes": 0, "replies": [ { "id": 2424068, "author_name": "", "author_url": "", "post_date": "2023-09-05T03:17:36.640000", "content": "

Thank you for the reminder. I just discovered this. The prediction score of the VIT model I used dummy is actually higher than the LB score I trained on the patient-level data. Data quality determines the upper limit of the model.

", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "2425581": "this should help:\nhttps://www.kaggle.com/competitions/rsna-2023-abdominal-trauma-detection/discussion/435053#2425570", "2429424": "Update: My 3D model still learns nothing, I can't find way data preprocessing to have a signal in 3D. Moreover, its awfully slow.\n2.5D works better and 2D much much better.", "2423730": "You should work on your data pipeline. Model isn't that important at this stage. I'm still experimenting with EfficientNet b0.", "2423666": "Hello, \nI'm creating 3D stacks with spacing 5mm, then resizing it to [1, 80, 256, 256] with `scipy.ndimage.zoom`. Then I use 3D ResNet50 model from `pytorchvideo`. I wanted to start from predicting `any_injuries`, because class imbalance is not as high as in other classes. The problem is my model learns nothing. Training loss is fluctuating. \n\nWhat I've tried:\n* class weighting\n* standarization (with mean and std)\n* different learning rates\n* different optimizers\n\nUsing all classes, model is learning class imbalance.\n\nCan you provide me with any tips how to continue? I'm stuck." } }