Starting from this year, I had a bit more free time from work and other factors, so I put a lot of effort into this competition with the aim of winning a gold medal.
\n We tried various techniques like pretraining from dataset of 2021 to 2023, including distillation, stft, using an ensemble of 14 models, multithreading, and many other things. Seeing the public LB, our efforts kept improving our scores.
\nHowever, the result was that the more effort we put in, the our private score didn't improve actually. It's quite disappointing.
\nTo those who achieved good results, I sincerely congratulate you and wish you all the best!
Starting from this year, I had a bit more free time from work and other factors, so I put a lot of effort into this competition with the aim of winning a gold medal.
\n We tried various techniques like pretraining from dataset of 2021 to 2023, including distillation, stft, using an ensemble of 14 models, multithreading, and many other things. Seeing the public LB, our efforts kept improving our scores.
\nHowever, the result was that the more effort we put in, the our private score didn't improve actually. It's quite disappointing.
\nTo those who achieved good results, I sincerely congratulate you and wish you all the best!
I extend my condolences for the unfortunate outcome.
\nIt seems possible that we could have created an ensemble of up to 7 or 8 models. Based on the actual results, I believe our outcome would have been better if we had included more models in the ensemble. However, due to the time to the deadline and lack of subs, we were only able to use 4 models.
\nI am curious about the ensemble of 14 models. Were you able to perform inference on all 14 models within 2 hours? Additionally, I would like to know if the 14 models were selected based on their promising performance, rather than just changing folds and parameters.
", "rawMarkdown": "I extend my condolences for the unfortunate outcome.\n\nIt seems possible that we could have created an ensemble of up to 7 or 8 models. Based on the actual results, I believe our outcome would have been better if we had included more models in the ensemble. However, due to the time to the deadline and lack of subs, we were only able to use 4 models.\n\nI am curious about the ensemble of 14 models. Were you able to perform inference on all 14 models within 2 hours? Additionally, I would like to know if the 14 models were selected based on their promising performance, rather than just changing folds and parameters.", "votes": 3, "replies": [ { "id": 2866093, "postDate": "2024-06-11T06:17:15.840Z", "content": "mutlithread + openvino, by model size, it was possible
", "rawMarkdown": "mutlithread + openvino, by model size, it was possible\n" } ] }, { "id": 2866068, "postDate": "2024-06-11T06:06:57.427Z", "content": "Ensembling 14 models at once is pretty powerful, how did you do it?
", "rawMarkdown": "Ensembling 14 models at once is pretty powerful, how did you do it?", "votes": 1, "replies": [ { "id": 2866092, "postDate": "2024-06-11T06:16:43.260Z", "content": "mutlithread + openvino, but quite tiny models like effcientvit_b0
", "rawMarkdown": "mutlithread + openvino, but quite tiny models like effcientvit_b0", "votes": 2 } ] }, { "id": 2866003, "postDate": "2024-06-11T05:23:44.147Z", "content": "Sorry about the shake my friend. I too struggled a bit in this one, I had probably a dozen times I had improved my CV significantly and really never even got improvement on the LB. I was close on a few ideas that I will experiment with post comp, but ultimately it was just one of those comps this time around! I know you will bounce back in the next one!
", "rawMarkdown": "Sorry about the shake my friend. I too struggled a bit in this one, I had probably a dozen times I had improved my CV significantly and really never even got improvement on the LB. I was close on a few ideas that I will experiment with post comp, but ultimately it was just one of those comps this time around! I know you will bounce back in the next one!", "votes": 1, "replies": [ { "id": 2866102, "postDate": "2024-06-11T06:22:37.703Z", "content": "Thank you so much for your comforting and encouraging words.
\nI will strive to achieve better results in the next competition. I appreciate your consolation, and I look forward to better news in the next competition!
I extend my condolences for the unfortunate outcome.
\nIt seems possible that we could have created an ensemble of up to 7 or 8 models. Based on the actual results, I believe our outcome would have been better if we had included more models in the ensemble. However, due to the time to the deadline and lack of subs, we were only able to use 4 models.
\nI am curious about the ensemble of 14 models. Were you able to perform inference on all 14 models within 2 hours? Additionally, I would like to know if the 14 models were selected based on their promising performance, rather than just changing folds and parameters.
", "votes": 3, "replies": [ { "id": 2866093, "author_name": "kaggler", "author_url": "", "post_date": "2024-06-11T06:17:15.840000", "content": "mutlithread + openvino, by model size, it was possible
", "votes": 0, "replies": [] } ] }, { "id": 2866068, "author_name": "Donghui Zhang", "author_url": "", "post_date": "2024-06-11T06:06:57.427000", "content": "Ensembling 14 models at once is pretty powerful, how did you do it?
", "votes": 1, "replies": [ { "id": 2866092, "author_name": "kaggler", "author_url": "", "post_date": "2024-06-11T06:16:43.260000", "content": "mutlithread + openvino, but quite tiny models like effcientvit_b0
", "votes": 2, "replies": [] } ] }, { "id": 2866003, "author_name": "Cody_Null", "author_url": "", "post_date": "2024-06-11T05:23:44.147000", "content": "Sorry about the shake my friend. I too struggled a bit in this one, I had probably a dozen times I had improved my CV significantly and really never even got improvement on the LB. I was close on a few ideas that I will experiment with post comp, but ultimately it was just one of those comps this time around! I know you will bounce back in the next one!
", "votes": 1, "replies": [ { "id": 2866102, "author_name": "kaggler", "author_url": "", "post_date": "2024-06-11T06:22:37.703000", "content": "Thank you so much for your comforting and encouraging words.
\nI will strive to achieve better results in the next competition. I appreciate your consolation, and I look forward to better news in the next competition!