![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F8722753%2Febebff60b5fa6262a84b8ec226e3d2ec%2F8825c21d-308d-4322-8878-41a5b3c33439.png?generation=1738803362204041&alt=media\")
References
\n[lb0.748] experiment resuts on cyroET foundation model build with synthetic data
\nif you are stuck below the benchmark.csv, read this!
\nHas Anyone Been Successful With Anything But Denoised Data?
\nDoes anyone have results with the 10441 external dataset works?
\nThanks to autors for their insights.
\n\n7 fold CV with experimental data only (0.808)
\nNormalization between 1 and 99 percentiles
\n64x128x128 voxels with hard masks at .5 particle radius
\nAugmentations: yz Rot90, contrast, brigthness and yz flips
\nLoss: Unweighted Zhang Shuai Dice Loss. It has a small bug on weights implementation (the multiplication should be outside dtype check).
\nThis initially led to 0.741/0.733, one more epoch of fine tune with weights [2,1,2,1,2,1] led to 0.740/0.735 (I'm not sure what to think about that)
\n\nA rotatory scan through the seven folds of 64x640x640 voxels each 16 slices so overlaps get 1,2,3,4,4,3,2,1 different rotation readings.
\nWhat didn't worked
\n\nDenoise synthetic data by training DenoisET that feeds freezed models already trained on experimental data.
\nSynthetic denoised data with real masks
\nHome made data by folding different small voxels into a bigger one masking inter-voxel particles
\nAugmentations: Fliping z or omelette rotation and free yz rotations
\nPseudo relabel some false positives as positives trying to emulate beta at same time than correct possible wrong annotations
\nPost processing: Filter by voxel_counts or discriminate false positves
\nTry to emulate F-Beta metric in loss
\nFinal Curiosity
\nI knew the data was already highly standardized, but that surprised me.
\n
References
\n[lb0.748] experiment resuts on cyroET foundation model build with synthetic data
\nif you are stuck below the benchmark.csv, read this!
\nHas Anyone Been Successful With Anything But Denoised Data?
\nDoes anyone have results with the 10441 external dataset works?
\nThanks to autors for their insights.
\n\n7 fold CV with experimental data only (0.808)
\nNormalization between 1 and 99 percentiles
\n64x128x128 voxels with hard masks at .5 particle radius
\nAugmentations: yz Rot90, contrast, brigthness and yz flips
\nLoss: Unweighted Zhang Shuai Dice Loss. It has a small bug on weights implementation (the multiplication should be outside dtype check).
\nThis initially led to 0.741/0.733, one more epoch of fine tune with weights [2,1,2,1,2,1] led to 0.740/0.735 (I'm not sure what to think about that)
\n\nA rotatory scan through the seven folds of 64x640x640 voxels each 16 slices so overlaps get 1,2,3,4,4,3,2,1 different rotation readings.
\nWhat didn't worked
\n\nDenoise synthetic data by training DenoisET that feeds freezed models already trained on experimental data.
\nSynthetic denoised data with real masks
\nHome made data by folding different small voxels into a bigger one masking inter-voxel particles
\nAugmentations: Fliping z or omelette rotation and free yz rotations
\nPseudo relabel some false positives as positives trying to emulate beta at same time than correct possible wrong annotations
\nPost processing: Filter by voxel_counts or discriminate false positves
\nTry to emulate F-Beta metric in loss
\nFinal Curiosity
\nI knew the data was already highly standardized, but that surprised me.
\n
Congratulations! We couldn't get any benefit from synthetic data too.
", "rawMarkdown": "Congratulations! We couldn't get any benefit from synthetic data too.", "votes": null }, { "id": "3116869", "postDate": "02/06/2025 11:50:01", "content": "Did you used 3D encoders? I've seen many teams that got profit with synthetic with 3D UNets May was that the difference.
", "rawMarkdown": "Did you used 3D encoders? I've seen many teams that got profit with synthetic with 3D UNets May was that the difference.", "votes": null }, { "id": "3116896", "postDate": "02/06/2025 12:21:20", "content": "Yes, we used 3D unets from monai. I am not sure either but maybe it was due to our setup.
", "rawMarkdown": "Yes, we used 3D unets from monai. I am not sure either but maybe it was due to our setup.", "votes": null } ], "comments": [ { "id": 3116744, "author_name": "snnclsr", "author_url": "", "post_date": "02/06/2025 09:08:20", "content": "Congratulations! We couldn't get any benefit from synthetic data too.
", "votes": null, "replies": [ { "id": 3116869, "author_name": "sacuscreed", "author_url": "", "post_date": "02/06/2025 11:50:01", "content": "Did you used 3D encoders? I've seen many teams that got profit with synthetic with 3D UNets May was that the difference.
", "votes": null, "replies": [ { "id": 3116896, "author_name": "snnclsr", "author_url": "", "post_date": "02/06/2025 12:21:20", "content": "Yes, we used 3D unets from monai. I am not sure either but maybe it was due to our setup.
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3116441": "**References**\n\n[[lb0.748] experiment resuts on cyroET foundation model build with synthetic data](https://www.kaggle.com/competitions/czii-cryo-et-object-identification/discussion/545221)\n\n[if you are stuck below the benchmark.csv, read this!](https://www.kaggle.com/competitions/czii-cryo-et-object-identification/discussion/547350)\n\n[Has Anyone Been Successful With Anything But Denoised Data?](https://www.kaggle.com/competitions/czii-cryo-et-object-identification/discussion/555247)\n\n[Does anyone have results with the 10441 external dataset works?](https://www.kaggle.com/competitions/czii-cryo-et-object-identification/discussion/556103)\n\nThanks to autors for their insights.\n\n[**Training**](https://www.kaggle.com/code/sacuscreed/czii-2dto3d-exp)\n\n7 fold CV with experimental data only (0.808)\n\nNormalization between 1 and 99 percentiles\n\n64x128x128 voxels with hard masks at .5 particle radius\n\nAugmentations: yz Rot90, contrast, brigthness and yz flips\n\nLoss: Unweighted [Zhang Shuai](https://github.com/shuaizzZ/Dice-Loss-PyTorch/blob/master/dice_loss.py) Dice Loss. It has a small bug on weights implementation (the multiplication should be outside dtype check). \n\nThis initially led to 0.741/0.733, one more epoch of [fine tune](https://www.kaggle.com/code/sacuscreed/czii-2dto3d-ft) with weights [2,1,2,1,2,1] led to 0.740/0.735 (I'm not sure what to think about that)\n\n[**Inference**](https://www.kaggle.com/code/sacuscreed/czii-submission-2dto3d)\n\nA rotatory scan through the seven folds of 64x640x640 voxels each 16 slices so overlaps get 1,2,3,4,4,3,2,1 different rotation readings.\n\n**What didn't worked**\n\n[Synthetic data](https://www.kaggle.com/code/sacuscreed/czii-2d-to-3d-synthetic-data-pretraining)\n\n[Denoise synthetic data](https://www.kaggle.com/code/sacuscreed/czii-synth-denoiset) by training DenoisET that feeds freezed models already trained on experimental data.\n\n[Synthetic denoised data with real masks](https://www.kaggle.com/code/sacuscreed/czii-realmask-2dto3d-synth)\n\n[Home made](https://www.kaggle.com/code/sacuscreed/czii-2d-to-3d-homemade-data-training) data by folding different small voxels into a bigger one masking inter-voxel particles\n\nAugmentations: Fliping z or omelette rotation and free yz rotations\n\nPseudo relabel some false positives as positives trying to emulate beta at same time than correct possible wrong annotations\n\nPost processing: Filter by voxel_counts or discriminate false positves\n\n[Try to emulate F-Beta metric in loss](https://www.kaggle.com/code/sacuscreed/translating-metric-into-a-loss)\n\n**Final Curiosity**\n\nI knew the data was already highly standardized, but that surprised me.\n\n", "3116744": "Congratulations! We couldn't get any benefit from synthetic data too.", "3116869": "Did you used 3D encoders? I've seen many teams that got profit with synthetic with 3D UNets May was that the difference.", "3116896": "Yes, we used 3D unets from monai. I am not sure either but maybe it was due to our setup." }, "source": "meta" }