Adam optimizer; 40 epochs; learning rate 5e-4, 1215 (1456) training (full-training) samples per epoch
\n
LR-Scheduler: OneCycleLR
\n
\n
What did not work:
\n
\n
TverskyLoss, DiceFocalLoss, weights in dice loss
\n
Other models like AttentionUnet
\n
RandCoarseDropoutd, RandAdjustContrastd with lower/higher values, RandRotated around x/y
\n
Removing smaller or larger connected-components based on threshold
\n
",
"rawMarkdown": "# Approach: Single 3D-UNet to detect segmentation masks & run connected-components-3d to localize detections\nFull code: https://github.com/Oliver-Busemann/CryoET/tree/main\n\n- 7-fold-CV: Train on patches from 6 samples\n- Validate on patches from hold-out sample\n- Save predictions from hold-out fold to disc\n- After 7-folds run connected-components-3D & competition metric on all predictions at once to get final score\n\n### CV-score: 0.7781; Public-LB: 0.72782; Private-LB: 0.71783 \n\n### What improved CV: \n\n- Larger patches (128 > 96 > 48) \n- Loss: 0.75 cross-entropy (without weights) + 0.25 dice loss (excluding background) \n- Masked loss: use only the inner (96, 96, 96) for loss as partially visible targets at the border lack context \n- Use the same mask for assigning predictions to reduce border artifacts \n- WeightedRandomSampler: upsample patches with targets such that each target is present in patches as many times as patches with only background (most) \n- Augmentations: RandFlipd (x, y, z), RandRotated (only z), RandGaussianNoised (mean=0.0, std=0.075), RandAdjustContrastd (gamma=(0.9, 1.1)) \n- Lower train stride to get more patches to sample from (32)\n- Individual radius for creating segmentation masks (apo-ferritin: 5, beta-galactosidase: 6, ribosome: 10, thyroglobulin: 6, virus-like-particle: 12) \n- Simple UNet: channels=(32, 64, 128, 256, 512), strides=(2, 2, 1, 1), dropout=0.2, num_res_units=1) \n- Adam optimizer; 40 epochs; learning rate 5e-4, 1215 (1456) training (full-training) samples per epoch\n- LR-Scheduler: OneCycleLR \n\n### What did not work: \n\n- TverskyLoss, DiceFocalLoss, weights in dice loss \n- Other models like AttentionUnet\n- RandCoarseDropoutd, RandAdjustContrastd with lower/higher values, RandRotated around x/y\n- Removing smaller or larger connected-components based on threshold \n",
"votes": 5
}
],
"comments": [],
"raw_markdown_by_id": {
"3118753": "# Approach: Single 3D-UNet to detect segmentation masks & run connected-components-3d to localize detections\nFull code: https://github.com/Oliver-Busemann/CryoET/tree/main\n\n- 7-fold-CV: Train on patches from 6 samples\n- Validate on patches from hold-out sample\n- Save predictions from hold-out fold to disc\n- After 7-folds run connected-components-3D & competition metric on all predictions at once to get final score\n\n### CV-score: 0.7781; Public-LB: 0.72782; Private-LB: 0.71783 \n\n### What improved CV: \n\n- Larger patches (128 > 96 > 48) \n- Loss: 0.75 cross-entropy (without weights) + 0.25 dice loss (excluding background) \n- Masked loss: use only the inner (96, 96, 96) for loss as partially visible targets at the border lack context \n- Use the same mask for assigning predictions to reduce border artifacts \n- WeightedRandomSampler: upsample patches with targets such that each target is present in patches as many times as patches with only background (most) \n- Augmentations: RandFlipd (x, y, z), RandRotated (only z), RandGaussianNoised (mean=0.0, std=0.075), RandAdjustContrastd (gamma=(0.9, 1.1)) \n- Lower train stride to get more patches to sample from (32)\n- Individual radius for creating segmentation masks (apo-ferritin: 5, beta-galactosidase: 6, ribosome: 10, thyroglobulin: 6, virus-like-particle: 12) \n- Simple UNet: channels=(32, 64, 128, 256, 512), strides=(2, 2, 1, 1), dropout=0.2, num_res_units=1) \n- Adam optimizer; 40 epochs; learning rate 5e-4, 1215 (1456) training (full-training) samples per epoch\n- LR-Scheduler: OneCycleLR \n\n### What did not work: \n\n- TverskyLoss, DiceFocalLoss, weights in dice loss \n- Other models like AttentionUnet\n- RandCoarseDropoutd, RandAdjustContrastd with lower/higher values, RandRotated around x/y\n- Removing smaller or larger connected-components based on threshold \n"
}
}
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8876667%2F9e8f3d93e5bc4779f812616ecdb1ff8e%2F75.png?generation=1739019029401530&alt=media\")