I said before the results that I would facepalm when I'd read the top solutions.
\nWell, as I was going through all those great and creative ideas I suddently thought \"Wait, they are all using 768 or 1024 inputs?. All this time I thought I was forced to use the default input size of the backbone.\" 🤦♂️\nSure enough, when I train just 1 fold of Dino large with 768 input, it already outperforms all my careful ensembles with TTT (.62 PB).
\nAlso pointed out by @konitac in their writeup, there was a bug with StratifiedGroupKFold in some sklearn versions. Although in my experiments the impact of this is not as obvious.
\nDid anyone have similar realisations? I hope not as painful as mine. 🫠
", "messages": [ { "id": "3400337", "postDate": "02/01/2026 11:46:54", "content": "I said before the results that I would facepalm when I'd read the top solutions.
\nWell, as I was going through all those great and creative ideas I suddently thought \"Wait, they are all using 768 or 1024 inputs?. All this time I thought I was forced to use the default input size of the backbone.\" 🤦♂️\nSure enough, when I train just 1 fold of Dino large with 768 input, it already outperforms all my careful ensembles with TTT (.62 PB).
\nAlso pointed out by @konitac in their writeup, there was a bug with StratifiedGroupKFold in some sklearn versions. Although in my experiments the impact of this is not as obvious.
\nDid anyone have similar realisations? I hope not as painful as mine. 🫠
", "rawMarkdown": "I said before the results that I would facepalm when I'd read the top solutions.\n\nWell, as I was going through all those great and creative ideas I suddently thought \"Wait, they are all using 768 or 1024 inputs?. All this time I thought I was forced to use the default input size of the backbone.\" 🤦♂️\nSure enough, when I train just 1 fold of Dino large with 768 input, it already outperforms all my careful ensembles with TTT (.62 PB).\n\nAlso pointed out by @konitac in [their writeup](https://www.kaggle.com/competitions/csiro-biomass/writeups/54th-place-solution-all-i-need-is-fastai-timm-a), there was a bug with StratifiedGroupKFold in some sklearn versions. Although in my experiments the impact of this is not as obvious.\n\nDid anyone have similar realisations? I hope not as painful as mine. 🫠", "votes": null }, { "id": "3400353", "postDate": "02/01/2026 12:27:48", "content": "In my random crop strategy, I realized that increasing the resolution of the crops had more impact than increasing the number of crops. I was earlier using random crops of 448x448 which were then resized to 224x224. I later moved on to using 448x448 random crops directly, and hence removed the unnecessary resize operation as well. This was especially beneficial for Green and Clover as they require much more details compared to the other targets.
", "rawMarkdown": "In my random crop strategy, I realized that increasing the resolution of the crops had more impact than increasing the number of crops. I was earlier using random crops of 448x448 which were then resized to 224x224. I later moved on to using 448x448 random crops directly, and hence removed the unnecessary resize operation as well. This was especially beneficial for Green and Clover as they require much more details compared to the other targets.", "votes": null }, { "id": "3400390", "postDate": "02/01/2026 13:23:45", "content": "It makes sense that downsizing the input would decrease performance, unless we get some crazy overfitting. I had a similar system with bigger crops, so there was probably a significant bottleneck there.
", "rawMarkdown": "It makes sense that downsizing the input would decrease performance, unless we get some crazy overfitting. I had a similar system with bigger crops, so there was probably a significant bottleneck there.", "votes": null } ], "comments": [ { "id": 3400353, "author_name": "komilparmar", "author_url": "", "post_date": "02/01/2026 12:27:48", "content": "In my random crop strategy, I realized that increasing the resolution of the crops had more impact than increasing the number of crops. I was earlier using random crops of 448x448 which were then resized to 224x224. I later moved on to using 448x448 random crops directly, and hence removed the unnecessary resize operation as well. This was especially beneficial for Green and Clover as they require much more details compared to the other targets.
", "votes": null, "replies": [ { "id": 3400390, "author_name": "tennogh", "author_url": "", "post_date": "02/01/2026 13:23:45", "content": "It makes sense that downsizing the input would decrease performance, unless we get some crazy overfitting. I had a similar system with bigger crops, so there was probably a significant bottleneck there.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3400337": "I said before the results that I would facepalm when I'd read the top solutions.\n\nWell, as I was going through all those great and creative ideas I suddently thought \"Wait, they are all using 768 or 1024 inputs?. All this time I thought I was forced to use the default input size of the backbone.\" 🤦♂️\nSure enough, when I train just 1 fold of Dino large with 768 input, it already outperforms all my careful ensembles with TTT (.62 PB).\n\nAlso pointed out by @konitac in [their writeup](https://www.kaggle.com/competitions/csiro-biomass/writeups/54th-place-solution-all-i-need-is-fastai-timm-a), there was a bug with StratifiedGroupKFold in some sklearn versions. Although in my experiments the impact of this is not as obvious.\n\nDid anyone have similar realisations? I hope not as painful as mine. 🫠", "3400353": "In my random crop strategy, I realized that increasing the resolution of the crops had more impact than increasing the number of crops. I was earlier using random crops of 448x448 which were then resized to 224x224. I later moved on to using 448x448 random crops directly, and hence removed the unnecessary resize operation as well. This was especially beneficial for Green and Clover as they require much more details compared to the other targets.", "3400390": "It makes sense that downsizing the input would decrease performance, unless we get some crazy overfitting. I had a similar system with bigger crops, so there was probably a significant bottleneck there." }, "source": "meta" }