{ "id": 308092, "title": "image resizing at different stages", "url": "/competitions/happy-whale-and-dolphin/discussion/308092", "author_name": "", "post_date": "2022-02-17T03:31:34.877363500Z", "votes": 13, "comment_count": 3, "views": 0, "content": "

I was recently reading @christofhenkel 's paper where he used a dolg-effnet model to gain stellar performance on the most recent landmark competition.
\n\"Screenshot-from-2022-02-17-08-52-36\"

in this paragraph ^ he mentions that the epochs were broken up into 3 different stages and at each stage different image sizes and \"dataset quality\" (see Google Landmarks Dataset literature for what clean dataset means) were used.

\n

Recently being active in CV competitions I've seen that these kinds of \"engineering tricks\" relating to ensembling/training procedure/other meta stuff are often used in implementations for high performing benchmarks while these are obviously not detailed in the official releases of the research papers/associated training procedures.

\n

For this specific resizing at different epochs technique, could someone please explain why exactly this might improve performance?

", "messages": [ { "id": "1693900", "postDate": "02/17/2022 03:31:34", "content": "

I was recently reading @christofhenkel 's paper where he used a dolg-effnet model to gain stellar performance on the most recent landmark competition.
\n\"Screenshot-from-2022-02-17-08-52-36\"

in this paragraph ^ he mentions that the epochs were broken up into 3 different stages and at each stage different image sizes and \"dataset quality\" (see Google Landmarks Dataset literature for what clean dataset means) were used.

\n

Recently being active in CV competitions I've seen that these kinds of \"engineering tricks\" relating to ensembling/training procedure/other meta stuff are often used in implementations for high performing benchmarks while these are obviously not detailed in the official releases of the research papers/associated training procedures.

\n

For this specific resizing at different epochs technique, could someone please explain why exactly this might improve performance?

", "rawMarkdown": "I was recently reading @christofhenkel 's [paper](https://arxiv.org/pdf/2110.03786.pdf) where he used a dolg-effnet model to gain stellar performance on the most recent landmark competition. \n\"Screenshot-from-2022-02-17-08-52-36\"

in this paragraph ^ he mentions that the epochs were broken up into 3 different stages and at each stage different image sizes and \"dataset quality\" (see Google Landmarks Dataset literature for what clean dataset means) were used. \n\nRecently being active in CV competitions I've seen that these kinds of \"engineering tricks\" relating to ensembling/training procedure/other meta stuff are often used in implementations for high performing benchmarks while these are obviously not detailed in the official releases of the research papers/associated training procedures. \n\nFor this specific resizing at different epochs technique, could someone please explain why exactly this might improve performance?", "votes": null }, { "id": "1693925", "postDate": "02/17/2022 03:57:51", "content": "

@ferlockx Kindly look into \"Progressive resizing\" and \"Progressive Learning\" techniques.

\n

Please see a similar discussion here

", "rawMarkdown": "ferlockx Kindly look into \"Progressive resizing\" and \"Progressive Learning\" techniques. \n\nPlease see a similar discussion [here](https://www.kaggle.com/c/happy-whale-and-dolphin/discussion/306777#1692796)", "votes": null }, { "id": "1695115", "postDate": "02/18/2022 00:52:19", "content": "

thanks for directing me there, effnetv2 paper was on my read-list :)

", "rawMarkdown": "thanks for directing me there, effnetv2 paper was on my read-list :)", "votes": null }, { "id": "1695208", "postDate": "02/18/2022 02:42:11", "content": "

Good Luck! :)

", "rawMarkdown": "Good Luck! :)", "votes": null } ], "comments": [ { "id": 1693925, "author_name": "init27", "author_url": "", "post_date": "02/17/2022 03:57:51", "content": "

@ferlockx Kindly look into \"Progressive resizing\" and \"Progressive Learning\" techniques.

\n

Please see a similar discussion here

", "votes": null, "replies": [ { "id": 1695115, "author_name": "ferlockx", "author_url": "", "post_date": "02/18/2022 00:52:19", "content": "

thanks for directing me there, effnetv2 paper was on my read-list :)

", "votes": null, "replies": [] }, { "id": 1695208, "author_name": "init27", "author_url": "", "post_date": "02/18/2022 02:42:11", "content": "

Good Luck! :)

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "1693900": "I was recently reading @christofhenkel 's [paper](https://arxiv.org/pdf/2110.03786.pdf) where he used a dolg-effnet model to gain stellar performance on the most recent landmark competition. \n\"Screenshot-from-2022-02-17-08-52-36\"

in this paragraph ^ he mentions that the epochs were broken up into 3 different stages and at each stage different image sizes and \"dataset quality\" (see Google Landmarks Dataset literature for what clean dataset means) were used. \n\nRecently being active in CV competitions I've seen that these kinds of \"engineering tricks\" relating to ensembling/training procedure/other meta stuff are often used in implementations for high performing benchmarks while these are obviously not detailed in the official releases of the research papers/associated training procedures. \n\nFor this specific resizing at different epochs technique, could someone please explain why exactly this might improve performance?", "1693925": "ferlockx Kindly look into \"Progressive resizing\" and \"Progressive Learning\" techniques. \n\nPlease see a similar discussion [here](https://www.kaggle.com/c/happy-whale-and-dolphin/discussion/306777#1692796)", "1695115": "thanks for directing me there, effnetv2 paper was on my read-list :)", "1695208": "Good Luck! :)" }, "source": "meta" }