I tried many times using TTA in my models, but all of them get lower public score than original predictions. I think it just enhance time cost in Inference. I also tried sliding windows predictions, each of windows using augmentation, I think it maybe like TTA. And the result also like TTA, lower socre than original predictions and enhance time cost in Inference.
", "messages": [ { "id": "3182885", "postDate": "04/20/2025 03:03:51", "content": "I tried many times using TTA in my models, but all of them get lower public score than original predictions. I think it just enhance time cost in Inference. I also tried sliding windows predictions, each of windows using augmentation, I think it maybe like TTA. And the result also like TTA, lower socre than original predictions and enhance time cost in Inference.
", "rawMarkdown": "I tried many times using TTA in my models, but all of them get lower public score than original predictions. I think it just enhance time cost in Inference. I also tried sliding windows predictions, each of windows using augmentation, I think it maybe like TTA. And the result also like TTA, lower socre than original predictions and enhance time cost in Inference.", "votes": null }, { "id": "3190465", "postDate": "04/30/2025 17:31:30", "content": "TTA can be useful but it is key that your augmentations are something that your model understands. If your model performs better on the original image and loses performance on the augmented versions then the added noise will only harm the predictions. For example, if a model has never seen an upside down picture of a dog then it will likely be less confident in its dog prediction, but if it has seen on many occasions a dog upside down and can predict it just as well as the original then it will add confidence and different perspectives to the model.
\nMy recommendation is to test if TTA is valid for this competition by training your model using 1 type of augmentation at 50% and use TTA on that 1 augmentation to see if it gives a better CV than without TTA. Sometimes our augmentations make it more difficult to give a confident prediction (such as adding gaussian noise) which can make it a bad candidate for TTA if that makes sense. It is hard to give specific help without knowing your code or augmentations but this is my take because I had the same issue a few months ago where my TTA submission was worse whereas others reported having better performance. Try using one of your most dominant augmentations (highest probability and is seen often in the model) too. I hope this word vomit helps lol
", "rawMarkdown": "TTA can be useful but it is key that your augmentations are something that your model understands. If your model performs better on the original image and loses performance on the augmented versions then the added noise will only harm the predictions. For example, if a model has never seen an upside down picture of a dog then it will likely be less confident in its dog prediction, but if it has seen on many occasions a dog upside down and can predict it just as well as the original then it will add confidence and different perspectives to the model.\n\nMy recommendation is to test if TTA is valid for this competition by training your model using 1 type of augmentation at 50% and use TTA on that 1 augmentation to see if it gives a better CV than without TTA. Sometimes our augmentations make it more difficult to give a confident prediction (such as adding gaussian noise) which can make it a bad candidate for TTA if that makes sense. It is hard to give specific help without knowing your code or augmentations but this is my take because I had the same issue a few months ago where my TTA submission was worse whereas others reported having better performance. Try using one of your most dominant augmentations (highest probability and is seen often in the model) too. I hope this word vomit helps lol", "votes": null } ], "comments": [ { "id": 3190465, "author_name": "connorjd", "author_url": "", "post_date": "04/30/2025 17:31:30", "content": "TTA can be useful but it is key that your augmentations are something that your model understands. If your model performs better on the original image and loses performance on the augmented versions then the added noise will only harm the predictions. For example, if a model has never seen an upside down picture of a dog then it will likely be less confident in its dog prediction, but if it has seen on many occasions a dog upside down and can predict it just as well as the original then it will add confidence and different perspectives to the model.
\nMy recommendation is to test if TTA is valid for this competition by training your model using 1 type of augmentation at 50% and use TTA on that 1 augmentation to see if it gives a better CV than without TTA. Sometimes our augmentations make it more difficult to give a confident prediction (such as adding gaussian noise) which can make it a bad candidate for TTA if that makes sense. It is hard to give specific help without knowing your code or augmentations but this is my take because I had the same issue a few months ago where my TTA submission was worse whereas others reported having better performance. Try using one of your most dominant augmentations (highest probability and is seen often in the model) too. I hope this word vomit helps lol
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3182885": "I tried many times using TTA in my models, but all of them get lower public score than original predictions. I think it just enhance time cost in Inference. I also tried sliding windows predictions, each of windows using augmentation, I think it maybe like TTA. And the result also like TTA, lower socre than original predictions and enhance time cost in Inference.", "3190465": "TTA can be useful but it is key that your augmentations are something that your model understands. If your model performs better on the original image and loses performance on the augmented versions then the added noise will only harm the predictions. For example, if a model has never seen an upside down picture of a dog then it will likely be less confident in its dog prediction, but if it has seen on many occasions a dog upside down and can predict it just as well as the original then it will add confidence and different perspectives to the model.\n\nMy recommendation is to test if TTA is valid for this competition by training your model using 1 type of augmentation at 50% and use TTA on that 1 augmentation to see if it gives a better CV than without TTA. Sometimes our augmentations make it more difficult to give a confident prediction (such as adding gaussian noise) which can make it a bad candidate for TTA if that makes sense. It is hard to give specific help without knowing your code or augmentations but this is my take because I had the same issue a few months ago where my TTA submission was worse whereas others reported having better performance. Try using one of your most dominant augmentations (highest probability and is seen often in the model) too. I hope this word vomit helps lol" }, "source": "meta" }